3D Generative AI Showdown: OpenAI Shap-E 2 vs Google’s AI Shopping View vs NVIDIA Picasso Edify
The world of 3D generative AI is heating up, with tech giants rolling out models and platforms that can conjure three-dimensional content from simple inputs. In this report, we compare three leading approaches: OpenAI’s Shap-E 2, Google’s new 3D Shopping View (powered by the Veo model), and NVIDIA’s Picasso Edify platform. We’ll break down what each one is, how they work, and where they shine:
- OpenAI Shap-E 2 – The anticipated successor to OpenAI’s Shap-E, a text-to-3D generative model. The original Shap-E (released 2023) could turn text prompts or reference images into 3D objects within seconds voicebot.ai. It introduced implicit shape representations (like NeRF neural fields) to produce more detailed, textured models than its point-cloud predecessor Point-E voicebot.ai. Shap-E 2 is expected to further improve detail and fidelity – essentially “a hypothetical sequel… to create 3D assets with a high degree of realism and detail using only a natural language description” madoverstories.com. OpenAI has not formally launched Shap-E 2 yet, but the tech community is buzzing with anticipation for an upgraded text-to-3D tool.
- Google 3D Shopping View (Veo) – Google’s latest AI-powered shopping feature that transforms a handful of product photos into an interactive 360° 3D view. Debuted in mid-2025, it uses Google DeepMind’s Veo generative video model to infer a product’s appearance from all angles research.google socialmediatoday.com. The goal is to replicate the in-store experience online – imagine scrolling through search results and being able to “pick up” a sneaker or gadget and spin it around. Google’s generative system can produce “high quality and shoppable 3D product visualizations from as few as three images”, eliminating the need for costly 3D scans research.google. It’s already live for shoes and expanding to more categories on Google Search and Shopping.
- NVIDIA Picasso Edify – NVIDIA’s enterprise-focused cloud AI foundry for visual content, with Edify as its multimodal foundation model. Picasso is a platform (launched 2023) that lets developers build generative AI tools for images, videos, and 3D on NVIDIA’s GPU cloud reuters.com. At its core is Edify, a diffusion-based AI architecture trained on licensed datasets (from partners like Getty Images and Shutterstock) to ensure high-quality, royalty-paid outputs reuters.com. Edify can power text-to-image, text-to-video, and crucially, text-to-3D generation. For example, it allows services where a user types “a cozy red sofa” and quickly gets a textured 3D model of a sofa. NVIDIA’s strategy is to provide this capability to creatives and brands via APIs and integrations (Adobe, Shutterstock, etc.), rather than a direct consumer app. In short, Picasso Edify is the behind-the-scenes workhorse enabling commercially safe generative media in many industries reuters.com.
In the sections below, we’ll delve deeper into each solution’s architecture, capabilities, use cases, user experience, performance, ecosystem integration, and how you can access them – complete with expert quotes and the latest news as of August 2025.
Core AI Architecture and Training Data
OpenAI Shap-E / Shap-E 2: Shap-E introduced a novel two-stage approach to 3D generation. First, an encoder network compresses 3D assets into a latent representation (the “parameters of an implicit function”). Second, a conditional diffusion model is trained on those encodings to generate new 3D shapes from prompts ngwaifoong92.medium.com ngwaifoong92.medium.com. Unlike explicit models that output point clouds or meshes directly, Shap-E produces an implicit 3D representation that can be rendered as a textured mesh or NeRF (neural radiance field) voicebot.ai. This implicit approach gives flexibility: the same internal representation can yield both realistic renderings and geometry. Shap-E’s diffusion model was trained on a large paired dataset of 3D models and text descriptions arxiv.org (the authors note it’s the same dataset used for Point-E, likely including open 3D asset collections with captions). The original Shap-E already outperformed Point-E’s point-cloud method in fidelity and speed voicebot.ai voicebot.ai, thanks to techniques like CLIP guidance and NeRF-based rendering. We can expect Shap-E 2, when released, to build on this: perhaps a larger model, more training data (including more complex shapes or photogrammetry data), and improved diffusion techniques for finer detail. OpenAI hasn’t published details of Shap-E 2’s training yet – it remains “hypothetical” for now madoverstories.com – but tech bloggers speculate it will maintain the implicit-function paradigm and push 3D fidelity closer to photorealism.
Google 3D Shopping (Veo): Google’s solution leverages advances in diffusion models for novel view synthesis. Initially, Google used Neural Radiance Fields (NeRFs) in 2022 to generate 360° spins from multiple photos research.google research.google. But NeRF alone struggled with sparse inputs (e.g. only 5 images of a sandal left gaps) research.google. The breakthrough came by incorporating diffusion: in 2023 Google researchers added a “view-conditioned diffusion prior” that could imagine missing views of an object, essentially asking the model “what would this product look like from another angle?” research.google. By 2025, they went further and fine-tuned Veo, a state-of-the-art video generator, for 3D product spins research.google research.google. Training Veo for this task involved an immense synthetic dataset: millions of 3D product models (from Google’s own collections or manufacturers) were rendered from various angles and lighting, producing paired data of static images and corresponding 360° video rotations research.google. This allowed Veo to learn how to go from a few static views to a smooth, consistent 3D spin. As Google’s team explains, “we curated millions of high-quality 3D assets… rendered from various angles and lighting… then supervised Veo to generate 360° spins conditioned on one or more images”, which taught the model to handle real product photos research.google research.google. The use of a diffusion-based video model is key – diffusion gives realism and coherence frame-to-frame. Google notes “a key strength of Veo is its ability to generate videos that capture complex interactions between light, material, texture, and geometry”, thanks to its powerful diffusion architecture research.google. In summary, Google’s core tech is a fine-tuned diffusion video model (Veo) that generates new views of a product, effectively “imagining” a 3D representation from limited input.
NVIDIA Picasso Edify: Edify is described as a “multimodal generative AI architecture” for visual domains blogs.nvidia.com. It’s essentially NVIDIA’s equivalent of a foundational model, but specialized for visuals (images, 3D, etc.) rather than text. Under the hood, Edify employs advanced diffusion models and possibly other generative techniques (the details are proprietary) to produce content across formats. One distinguishing factor is training data: Edify models are trained through NVIDIA’s AI Foundry program on licensed, high-quality datasets reuters.com reuters.com. For example, NVIDIA partnered with Getty Images and Shutterstock to train models on their vast libraries (over 650 million images and many thousands of 3D models, all properly licensed) marketscreener.com marketscreener.com. By training on curated data (and compensating contributors), Edify aims to be “commercially safe” – i.e. free of copyright worries – and also tailored to professional-quality outputs. In the 3D realm, NVIDIA and Shutterstock built a text-to-3D model using Edify that was trained on 500,000+ 3D models and corresponding images/metadata marketscreener.com. Technically, Edify’s 3D generator uses a mesh-based approach: it outputs quad-based meshes with textures and PBR (physically based rendering) materials for realism aibase.tech. An NVIDIA blog explains that Edify’s 3D pipeline integrates diffusion models that understand RGB color and surface contours, and outputs models with natural topology (quad meshes are easier to animate and edit than point clouds or triangles) aibase.tech. Essentially, Edify “imagines” a 3D object from a prompt, then constructs a mesh with realistic textures (metal, wood, fabric, etc.) mapped onto it – all consistent with real-world lighting physics thanks to PBR materials aibase.tech. Edify’s architecture is designed to be extensible: the same core model can be fine-tuned for different data or tasks. For instance, Getty Images uses a variant of Edify for image generation (with custom style controls), while Shutterstock uses an Edify 3D variant for model generation blogs.nvidia.com blogs.nvidia.com. All these models run on NVIDIA’s cloud (DGX Cloud infrastructure), and are accessible via NVIDIA NIM (NVIDIA Inference Microservice) APIs blogs.nvidia.com blogs.nvidia.com. In short, NVIDIA Picasso Edify is less a single model and more an AI factory: a collection of diffusion-based models trained on specific media domains, all built on a common architecture that emphasizes high fidelity and licensed training content.
Capabilities and Use Cases
Each of these AI offerings has different strengths and target applications:
- OpenAI Shap-E 2 Capabilities: Shap-E (and by extension Shap-E 2) is all about text-to-3D and image-to-3D generation for creative content. Give it a prompt like “a green armchair shaped like an avocado” and it will generate a 3D model of exactly that – complete with textures designboom.com. The original Shap-E could output either a textured mesh or a NeRF (neural radiance field) for the object voicebot.ai. This means you can both visualize the model in 2D from any angle or import the mesh into 3D software (e.g. Blender or Unity) for further use. The models Shap-E creates tend to capture “fine-grained textures and complex, detailed shapes”, far better than earlier methods that gave only sparse point clouds voicebot.ai. For example, Shap-E can handle not just the shape of a chair but also its surface pattern (leather vs. fabric) and color variations, which Point-E struggled with voicebot.ai. However, Shap-E’s outputs are still somewhat artistic or stylized in many cases – they are plausible and recognizable, but not necessarily photorealistic copies of real objects. This is acceptable for its primary use cases: gaming, art, concept design, and prototyping. Shap-E is great for ideation – an artist or game designer can generate lots of whimsical 3D concepts (e.g. “a flying car with dragon wings”) in seconds, then pick one to refine. Because it’s open-source and fast, developers have integrated Shap-E into tools like Blender (via plugins) to streamline creative workflows reddit.com reddit.com. Target use cases include entertainment (quickly generating 3D assets or characters), education (students visualizing objects from text), and possibly architecture or product design in early stages. Shap-E 2, when it arrives, will likely push these capabilities further – aiming for more realism and accuracy so that the generated models could even be used in AR/VR or e-commerce. There’s speculation that OpenAI might integrate Shap-E’s tech into future ChatGPT plugins or a 3D version of DALL·E, allowing everyday users to say “make me a 3D model of a new toy” and get an asset ready to print or animate madoverstories.com. For now, Shap-E remains a playground for creators and researchers rather than a turnkey business solution.
- Google 3D Shopping View Capabilities: Google’s generative shopping feature is laser-focused on product visualization. It takes existing products (say, a sneaker that a retailer has listed online) and creates a 360-degree interactive view of that item. Importantly, it’s not generating wild new 3D objects from scratch – it’s reconstructing real products in 3D from a few photographs. The capability here can be described as image-to-3D conversion, or more specifically multi-view synthesis. With as few as 3 input photos, Google’s AI can produce a coherent spin that shows the product from angles that were never actually photographed research.google research.google. This includes filling in hidden sides (e.g. the back of a shoe from only front and side photos) by hallucinating likely appearance based on learned knowledge. Notably, the system handles complex lighting and materials – Google touts that “Veo was able to capture complex lighting and material interactions (shiny surfaces), which was challenging for earlier approaches” research.google. The result is a photorealistic 3D asset that consumers can rotate on their screen, as if the product were in their hand. The key use case is e-commerce: helping shoppers inspect products online just like they would in a store. Shoes were the pilot category (launched with 360° spins on Google Search in 2022 research.google), and by 2025 this expanded to categories like sandals, heels, boots, and even furniture and electronics research.google research.google. Google explicitly mentions that “as few as three images capturing most object surfaces are sufficient” for a high-quality 3D result research.google – a game-changer for small businesses who might not have 3D scanners. Another capability is speed and scale: Google’s diffusion approach allowed scaling up to generate thousands of models for products in their catalog daily research.google. The generative 3D models are also fully shoppable – meaning they are linked to product listings, prices, and can be embedded in ads or search results. We’re also seeing Google integrate this with AR and virtual try-on features. For instance, Google’s new AI shopping suite lets users virtually try on apparel by generative AI (using your photo to see how clothes fit), and 3D spins for product images complement that by showing items in 3D business.google.com. Future use cases could include viewing these 3D products in your space (AR home placement via your phone) or even powering VR shopping malls. Google’s own blog hints at this, saying such 3D models could feed into virtual worlds where users can engage with real products socialmediatoday.com. In summary, Google’s 3D Shopping View is a tool for merchants and shoppers: merchants get an easy way to showcase their goods in 3D (driving more engagement), and shoppers get confidence through a richer, more interactive browsing experience.
- NVIDIA Picasso Edify Capabilities: NVIDIA’s Edify, being a broad platform, wears many hats. Its capabilities span generating and editing images, creating 360° HDR environment maps, producing 3D models, and even potentially video medium.com medium.com. Focusing on 3D: Edify (through partners) enables text-to-3D asset generation at a quality level suitable for professional use. A user (typically via a partner interface like Shutterstock’s) can input a text prompt or reference image, and Edify will output a fully textured 3D model in a chosen format (for example, OBJ, GLB, or USDZ file) marketscreener.com marketscreener.com. These models come with clean geometry (often quad meshes) and PBR textures applied, so they’re ready to import into game engines, 3D editors, or even for 3D printing with minimal cleanup marketscreener.com. An impressive feature is Edify’s speed: it can generate a preview image of the 3D object in ~10 seconds, allowing the user to approve the design before committing to the full 3D generation blogs.nvidia.com. Once approved, the service refines it into a high-quality 3D model with realistic materials like concrete, wood, leather, etc., matching the prompt blogs.nvidia.com. For instance, a creator could type “Victorian-style wooden chair with red velvet cushions” and quickly get a preview image; if it looks good, they’ll receive the 3D model of that chair, correctly textured, to use in their scene. Another capability is 360° HDRi background generation – using Edify to create panoramic environment maps from text (e.g. “sunset beach”) for lighting 3D scenes marketscreener.com blogs.nvidia.com. This is hugely useful for rendering and VFX, as artists can generate custom lighting environments on the fly instead of scouting or shooting HDR panoramas. Target use cases for Picasso Edify are enterprise and professional media creation:
- Gaming & Metaverse: studios can rapidly generate props, scenery, or even characters to populate virtual worlds aibase.tech. Edify’s outputs are compatible with popular tools and game engines (and NVIDIA is deeply involved in the OpenUSD format for interoperability).
- Product Design & Prototyping: companies like Mattel use Edify to ideate toy designs – a designer can describe a concept toy and visualize it in 3D, accelerating the iteration process blogs.nvidia.com aibase.tech. HP showcased using Edify to create prototype models that can then be 3D-printed for real-world evaluation blogs.nvidia.com.
- Advertising & Marketing: Edify is powering tools like Generative AI by Getty Images and Shutterstock’s 3D, which agencies (e.g. WPP, Omnicom) use to create custom visuals for clients blogs.nvidia.com blogs.nvidia.com. Need a unique 3D scene for a commercial? Edify can generate the assets and even composite them into shots. One example: creative teams generated a 3D model of a Land Rover Defender car and placed it in different AI-generated environments for promotional materials blogs.nvidia.com aibase.tech.
- Architecture & AR/VR: Designers could quickly generate interior decor objects or entire environments using text prompts, then use them in AR/VR walkthroughs. Because Edify’s models aim to be high-fidelity, they can serve as drop-in assets for visualization.
- Content Marketplaces: Shutterstock’s integration suggests a future where customers can simply generate the 3D stock content they need (“on-demand stock 3D”) rather than searching a library blogs.nvidia.com. This flips the stock media industry into a generative mode.
In essence, NVIDIA’s Edify is positioned for enterprise creativity – its killer capability is letting professionals generate custom, high-quality visuals (2D and 3D) with minimal effort, all while respecting IP/licensing. It’s less for casual users and more for powering the next generation of creative software and services (Adobe Firefly’s upcoming 3D features, for example, are being co-developed with NVIDIA blogs.nvidia.com).
User Experience and Interface
The user experience differs greatly among these three, reflecting their different audiences:
- Shap-E 2 (OpenAI) UX: Currently, Shap-E is a research project released as open-source code, so using it requires some tech savvy. Early adopters run Shap-E via Python notebooks or command-line, inputting a text prompt and receiving a 3D file or visualization. There are community demos (e.g. on Hugging Face Spaces) where you can type a prompt and see a rendered result. Additionally, independent developers have built plugins – for example, a Blender addon allows artists to invoke Shap-E from within Blender and import the generated model directly reddit.com reddit.com. This makes it somewhat easier for 3D artists to experiment without leaving their primary tool. That said, Shap-E doesn’t have a slick GUI or app of its own yet. If Shap-E 2 becomes part of OpenAI’s product lineup, we might see it integrated into an interface like ChatGPT or a web app, where a user could chat “Create a 3D model of X” and get a downloadable model. For now, the UX is geared toward developers and researchers – you need to write code or use a community-provided web demo. Once you have a Shap-E model output, interacting with it is up to you (opening it in a 3D viewer, game engine, etc.). The speed is a plus for UX: with ~10–20 seconds generation time on good hardware, it feels nearly real-time to get a result voicebot.ai. Users can tweak the text prompt or supply a reference image to guide the output if the first result isn’t right. There is no dedicated “interface” for Shap-E beyond these techie workflows, so usability for the general public is limited at the moment. We anticipate that could change – if OpenAI sees demand, they might launch a user-facing 3D model generator (similar to how DALL·E 3 is now integrated into Bing Chat, perhaps a Shap-E model could integrate into some platform). In summary, current UX: experimental and code-driven; potential future UX: maybe a ChatGPT plugin or a simple web form for text-to-3D.
- Google 3D Shopping View UX: Google has made the experience seamless for end-users. If you’re shopping on Google (Search or the Shopping tab) and the item has a generative 3D model, you’ll see an interactive viewer – usually a rotatable image or a 3D icon you can tap. With a swipe or drag of your finger, you can spin the product 360 degrees thespacelab.tv thespacelab.tv. For example, searching for running shoes might show certain results labeled as “Interactive 3D”. Clicking one brings up a viewer where the shoe rotates as you drag, letting you inspect the sole, sides, and back. The look and feel is much like a high-quality product model that was manually created, except it was AI-generated behind the scenes. Users do not need to know anything about AI – it’s just a richer media experience within Google’s interface. On mobile, these 3D views can often be expanded to fullscreen and sometimes integrated with AR (“View in your space” using your phone’s camera), though AR is typically available for products where companies have provided 3D models. With Google’s generative approach, AR could potentially be enabled for many more products automatically. From the merchant side, the UX is also straightforward: they just upload product photos (the typical shots from a few angles). Google’s system does the heavy lifting to create the 3D spin overnight. No special action or cost is required from the seller – which is a huge UX win compared to having to commission 3D models or photography. Google Merchant Center likely has guidelines encouraging merchants to provide at least 3 images (front, side, back) to trigger the generative 3D creation research.google research.google. Once generated, the 3D spins are integrated into search results, shopping listings, and even Google Ads. In fact, Google has started using these “3D spins” in advertising creatives on its platform business.google.com – meaning if you’re running a product ad, Google can automatically show an interactive 3D version to users, which is more engaging than a flat photo. The quality of the experience is notably high: reviewers have noted it “makes it feel less like browsing a flat webpage and more like holding the product in your hand.” thespacelab.tv. It’s silent and smooth (a looping turntable animation if you don’t interact). In short, Google’s UX is consumer-friendly and integrated – users might not even realize AI is involved, they just get a better way to shop. There’s no standalone app; it’s part of Google Search. The only limitation is that it’s currently available for select product types and in certain regions (rolling out first in the U.S. for shoes, etc. thespacelab.tv). As coverage grows, this could become a standard expectation: e.g. “Why can’t I spin this product? Oh, maybe it doesn’t have enough images or isn’t supported yet.”
- NVIDIA Picasso Edify UX: Since Edify is a B2B platform, most end-users will interact with it indirectly through partner applications. For instance, Shutterstock’s Generative 3D service provides a web interface where a user can enter a prompt, see a preview image of the generated 3D model, and then download the model files blogs.nvidia.com blogs.nvidia.com. That interface is designed by Shutterstock, but powered by Edify behind the scenes. Similarly, Getty Images’ generative AI (for pictures) has a web UI where you type prompts and get four image results, thanks to Edify’s model blogs.nvidia.com blogs.nvidia.com. In the context of 3D, NVIDIA also demonstrated a Blender plugin workflow at SIGGRAPH 2024: artists could use a plugin to generate 3D objects right inside Blender’s viewport blogs.nvidia.com. Imagine sculpting a scene and simply typing “/generate a vase” to place a new AI-created vase model into your scene – that’s the kind of integration being explored. For developers, NVIDIA offers the NIM API and the Build interface (at build.nvidia.com) where they can test Edify models via cloud API calls blogs.nvidia.com blogs.nvidia.com. This is more technical (requires coding or using REST API calls). In short, Edify’s UX varies depending on the implementation:
- For a creative professional using a partner platform (Shutterstock, Getty, eventually Adobe Creative Cloud), it’s a feature within tools they already use. It might feel like an AI “assistant” inside their content creation software – e.g. a “Generate 3D” button in a stock media site or a Firefly panel in Photoshop/Blender that taps NVIDIA’s models.
- For enterprise developers, the UX is through APIs and SDKs – they integrate Edify’s capabilities into their own products or pipelines. NVIDIA provides documentation and presumably an online console to manage model usage.
Performance: Speed, Realism, and Interactivity
When it comes to performance, we look at how fast these systems generate output, how realistic or detailed the results are, and how interactive the output is for the end-user.
- Shap-E / Shap-E 2 Performance: One of Shap-E’s bragging points was speed. The researchers reported that “each Shap-E sample… took about 13 seconds to generate on a single NVIDIA V100 GPU, whereas Point-E took ~2 minutes for the same.” voicebot.ai This is a huge leap, making Shap-E feel almost instant in practical terms. That speed likely comes from the efficient diffusion-decoder design and the implicit representation (which is compact to generate). In terms of realism, Shap-E’s outputs are a mixed bag. They are coherent and textured, but often have a stylized or fuzzy appearance compared to photorealistic CGI. For example, a Shap-E generated “shark” might clearly have a shark shape and even a nicely colored skin, but up close it won’t have the intricate scales or perfectly sharp features a real shark (or a high-poly model) would twitter.com. The use of NeRF for rendering gives a smooth, somewhat soft look (good for avoiding jagged geometry, less good for fine surface detail). Shap-E was a research prototype, so it wasn’t pushing the limits of resolution. Shap-E 2, if it leverages advances in diffusion and larger models, could improve detail – perhaps generating higher resolution textures or more complex geometry. Another aspect is consistency: Shap-E can sometimes produce odd artifacts or incomplete parts if the prompt is very complex (the paper noted it “still struggles with some complex objects” and large-scale scenes) voicebot.ai. So it’s best at single objects or small object sets, not entire landscapes or scenes (for those you’d generate pieces and assemble them). Interactivity of Shap-E outputs: once you have the 3D model, you can fully interact with it – orbit, zoom, even edit it if you import into modeling software. The models can be animated or used in games (given they are meshes/NeRFs). However, Shap-E itself doesn’t create animated models or enable interaction in the generation process. It’s one-shot generation of a static asset. So, interactivity depends on what you do with the output. For example, a user could generate an object and then use an AR app to place it in their room – that’s feasible because Shap-E provides standard 3D formats. Performance-wise, Shap-E’s strength is quick generation; its weakness is realism when compared to the likes of Google or NVIDIA’s outputs. It’s more akin to an AI prototype that’s “pretty good for seconds of work” rather than a polished, photoreal asset pipeline. That said, for many creative purposes, that level of detail is sufficient (and you can always touch up the model manually). We’ll have to see if Shap-E 2 narrows the realism gap – possibly through training on larger 3D datasets or employing techniques like Score Distillation (from DreamFusion) which produce more lifelike results at the cost of compute.
- Google Veo Shopping View Performance: Google’s generative 3D models are built for high realism and fluid interaction, since they face consumer eyeballs directly. From reports and examples, the generated spins are often indistinguishable from a real 360° product photo shoot. They capture material properties – e.g. you can see the sheen of patent leather or the knit texture on a sneaker as it turns. Google emphasizes how Veo handles reflections and highlights: “complex interactions between light, material, texture, and geometry” are preserved research.google. This suggests a very photoreal rendering quality. One limitation: since the output is effectively a rendered 360° video or image set, it’s not a true 3D mesh that you can, say, zoom into infinitely or place under different lighting easily. The lighting and background are typically neutral or white, similar to product photography. But within that scope, it’s extremely realistic – likely sufficient for any shopper to examine details like tread patterns or fabric weaves. In terms of speed, for the end-user the rotation is real-time (the images are precomputed, so it’s just loading them). For generation, Google hasn’t disclosed exact times, but the fact that they are “gradually rolling it out” socialmediatoday.com and mention scaling improvements implies it might take on the order of minutes per product with heavy compute, done offline. (They generate these spins in advance, not on the fly when you search – so latency for users is just like loading normal images.) The scale is impressive: Google is generating many of these. They stated that the majority of daily-viewed shoes on Google now use this generative tech research.google, meaning it’s robust enough to handle thousands of products and variations. Interactivity is high from a user perspective: the spins are fully interactive (users can drag to rotate 360°, possibly tilt if multi-axis, though most are single-axis spins). It’s a smooth experience with no noticeable AI glitches – any hallucinated parts are blended in so well that users likely don’t notice that, say, the back of a shoe was never actually photographed. As Steve Seitz of Google Labs wrote, “each technology has played a key part in making online shopping feel more tangible and interactive” research.google – the interactivity is core to that. Another measure of performance is how the model deals with sparse input: evidently, with just 3-5 images, it manages to produce a full 360. Earlier methods would produce blurriness or errors when views were missing (like a sandal’s thin straps confounding a NeRF) research.google. Veo’s diffusion prior hallucinates plausible fills for those gaps. Of course, if the input images are very scarce or all from the same angle, there’s only so much it can do – extreme cases might still yield some blur on the fully unseen side. But Google claims as few as 3 well-chosen images are enough to “reduce hallucinations” significantly research.google. The trade-off: some fidelity increases with more images (if you feed 5-10 images, you’ll get even crisper results, as the AI has less guesswork) research.google. In summary, Google’s 3D Shopping View offers photorealism and smooth interactivity, generated in a reasonably efficient pipeline that can operate at Google scale. It’s tailored for visual accuracy because trust is key – shoppers won’t use it if the product looks wrong. And based on its reception, it appears to hit that mark, delivering a “better-looking 3D product depiction with fewer inputs” than prior methods socialmediatoday.com.
- NVIDIA Edify Performance: NVIDIA’s Edify, especially as demonstrated in Shutterstock’s 3D generator, shows strong performance on multiple fronts. Generation speed: We saw that a preview of a 3D model can be obtained in ~10 seconds blogs.nvidia.com, and a full detailed model presumably takes a bit longer (perhaps a few tens of seconds to a minute, depending on complexity). For comparison, older research like DreamFusion often took many minutes (or even hours) to optimize a single 3D model via neural networks. Edify’s ability to deliver useful previews almost instantly is a big productivity win – it leverages heavy pretraining so that runtime is fast. Realism and detail: Edify’s outputs are high-quality, given that partners are positioning them as commercial-grade assets. The generated 3D models have “clean geometry and layout”, meaning no messy mesh artifacts blogs.nvidia.com. And because they include PBR materials, the objects look realistic under proper lighting – metals shine, glass is transparent, etc. The resolution of textures can go up to what professionals expect (Shutterstock mentions 16K HDRi for backgrounds blogs.nvidia.com, and for models, having 4K texture maps is likely feasible). In terms of accuracy to the prompt, Edify benefits from training on specific domains: e.g. Shutterstock’s model was trained on real furniture, vehicles, etc., so if you ask for “a cozy sofa,” it knows the nuances of sofas. The outputs should reflect the style and proportions of real-world objects, rather than the somewhat cartoonish look a generic model might produce. And because each company can fine-tune it, a brand can enforce a certain aesthetic (e.g. Getty can fine-tune for photographic style in images blogs.nvidia.com; similarly a furniture retailer could fine-tune for their catalog’s style in 3D). This yields very accurate and on-brand outputs – a big plus for enterprise use. On the interactivity front: Edify’s generated 3D models are standard files, so they are fully interactive in any 3D environment. A user can take the model, view it from any angle, edit it, animate it, etc. In contrast to Google’s approach (which yields a baked spin), Edify yields the actual asset. This means, for instance, a game developer could generate a bunch of props and then directly drop them into a game, collision and all. Or an AR app could call the API to generate a new piece of furniture and then immediately let the user place it in their room in real-time. The only consideration is polygon count and optimization – but since Shutterstock’s service offers multiple formats and presumably LOD options blogs.nvidia.com, developers can choose a game-ready format. Edify’s performance in handling complex scenes: currently, it’s often one object at a time (plus separate background generator for scenes). If you need a full scene, you might generate several objects and arrange them, or use an image generator to compose them into a 2D scene. But we see hints of more interactive scene generation: e.g. NVIDIA mentions Accenture Song using Edify to create entire “cinematic, interactive 3D environments via conversational prompts,” combining generated environments with digital twins (like a car model) blogs.nvidia.com blogs.nvidia.com. This suggests the roadmap might include generating not just single assets but whole 3D scenes on the fly. Performance there would be heavier, but NVIDIA’s focus on GPU acceleration and DGX Cloud means they aim to make even complex generation reasonably fast for enterprise workflows. If a creative studio can generate a whole city block in a few minutes instead of modeling for weeks, that’s a performance revolution (even if some manual tweaking is needed afterward). In summary, NVIDIA Edify offers fast turnaround and high-fidelity outputs intended for production use. Its outputs are as interactive as any manually created 3D model, giving it an edge in flexibility. The trade-off is you need powerful cloud GPUs (provided by NVIDIA) to achieve this performance – but that’s abstracted away in the services that use it. Also, because each Edify instance may be tailored, performance can vary (the Getty image model can make 4 images in 6 seconds at 4K blogs.nvidia.com, which is very fast; the 3D model might take a bit more per asset, but still in the order of seconds to a couple minutes, not hours). Overall, Edify’s performance is enterprise-grade – speed suitable for near real-time use, and quality good enough for commercial projects.
Integration with Platforms and Ecosystems
Each solution plugs into (and enhances) a broader ecosystem:
- OpenAI Shap-E Integration: As of 2025, Shap-E is not integrated into OpenAI’s commercial API or platforms (ChatGPT, etc.) yet. It exists as a standalone project on GitHub ngwaifoong92.medium.com. However, its open-source nature means the community has integrated it into various tools. For example, there’s a Blender addon (community-made) that lets artists invoke Shap-E inside Blender reddit.com. This is a natural integration, as Blender is a popular open-source 3D suite – having AI assist in model creation there can speed up workflows. We’ve also seen Shap-E models available on Hugging Face and demo sites, making it easier to experiment without setup huggingface.co. Looking at OpenAI’s trajectory, it’s plausible Shap-E or its successor could be woven into their API offerings. OpenAI has a “multimodal” vision (e.g. GPT-4 can handle text and images; DALL-E handles image generation). A logical step is extending to 3D. If OpenAI does this, we might see Shap-E 2 integrated into development platforms or design software via official plugins. Consider an Autodesk plugin or Adobe dimension integration where OpenAI’s API is used to generate models – that could be in the cards if OpenAI pursues enterprise partnerships. Also, OpenAI could integrate 3D generation in ChatGPT’s interface via a plugin: imagine asking ChatGPT to “generate a 3D model of a tree” and it returns an interactive widget or a download link. That would significantly broaden the reach of Shap-E beyond the developer community to everyday users and creators. In the broader ecosystem, Shap-E benefits from open research – it can be extended or fine-tuned by anyone. Researchers might integrate Shap-E with robotics (to generate object models on the fly for simulation) or with gaming engines (some have done experiments to generate Minecraft models from text, for instance). While OpenAI’s own ecosystem integration is minimal right now, Shap-E is basically a building block that others can integrate wherever 3D content is needed quickly. If Shap-E 2 comes out with improved capabilities, we expect it to appear in more AI toolchains – possibly integrated with 3D printing services (type what you want, get an STL file printed), or in educational platforms (students create virtual scenes with text prompts). OpenAI’s strength is its developer community, so integration often happens bottom-up: lots of smaller projects incorporating Shap-E into apps, mods, and creative pipelines.
- Google 3D Shopping Integration: Google has tightly integrated the 3D Shopping View into its Search and Ads ecosystems. On Google Search, products that have AI-generated 3D spins are highlighted and interactive business.google.com. This not only improves user experience but potentially SEO – retailers might get an engagement boost if their listings have 3D (which could encourage them to supply the needed images). Google also integrated these spins into Google Shopping ads (Performance Max ads). Advertisers can let Google automatically generate 3D models of their products to use in interactive display ads business.google.com. This integration is a win-win: ads become more engaging (which Google hopes leads to better click-through or conversion), and advertisers don’t have to do extra work. Moreover, Google’s 3D models are part of its Shopping Graph – the big database of products and attributes. We could foresee a scenario where Google offers an API or a feature in Google Cloud for retailers to embed the 3D viewer on their own sites (though currently the feature is primarily on Google’s surfaces). There is also synergy with ARCore/Google Lens: Google has AR search results (e.g. you can view animals in AR). With generative 3D, they could dramatically expand AR search content – any product could potentially be viewed in your space via AR, even if the retailer never made a 3D model, because Google’s AI made one. While not explicitly announced, it seems plausible that Google will integrate generative 3D into its AR services and possibly allow consumers to use it. Another integration point is Google Cloud Vertex AI – Google could offer the underlying tech (Veo model for view synthesis) as a service for developers (for example, an API where you feed a few images and get a 3D spin). As of mid-2025, there’s no public Vertex API for image-to-3D yet, but Google did mention Veo 3 is available in the Gemini API for video generation developers.googleblog.com. It’s possible that the product spin model is considered a specialized internal tool rather than a general API product (since it’s tailored to consumer goods). In any case, the integration strategy is clear: enhance Google’s own products (Search, Shopping, Ads, AR) using this generative AI. This deep vertical integration means Google controls the entire pipeline from generation to user interface. It also cements Google’s value proposition to retailers: list your products with us, and we’ll automatically make them look awesome in 3D. This could be a competitive moat against platforms like Amazon (which has some 360 photo viewer capabilities but not at this AI scale). It’s worth noting privacy/ethics integration: Google likely has safeguards in place – e.g. ensuring the AI model doesn’t alter logos or misrepresent the product. They would integrate those checks because trust in shopping content is crucial. All said, Google’s generative 3D is now a native part of the Google Shopping ecosystem and likely to expand across Google’s consumer and developer offerings, but it’s not something end-users directly call via an API or external app at this time.
- NVIDIA Picasso Edify Integration: NVIDIA’s approach to integration is partnership-driven and infrastructure-level. Edify models are integrated into the offerings of major creative and media companies:
- Shutterstock: Integrated Edify into its platform as Shutterstock Generative 3D and generative 360 HDRi services blogs.nvidia.com blogs.nvidia.com. This means Shutterstock’s users (photographers, designers, businesses) can generate 3D models or environment maps directly on Shutterstock’s site, and these appear alongside traditional stock models for download. It’s a deep integration where Shutterstock even handles the ethical licensing side (paying contributors whose data was used) – a model NVIDIA encourages blogs.nvidia.com.
- Getty Images/iStock: Integrated Edify to launch Generative AI by Getty Images, focusing on image creation initially, with fine-tuning options for enterprise clients blogs.nvidia.com blogs.nvidia.com. This is delivered via Getty’s website and API, giving Getty’s customers (advertisers, editors) a way to instantly create on-brand visuals. It’s a notable integration because Getty was historically cautious with generative AI (due to copyright) but chose NVIDIA as a safe partner reuters.com reuters.com.
- Adobe: While not a consumer integration yet, Adobe and NVIDIA announced a partnership to bring Edify’s 3D generative tech into Adobe’s products (Firefly and Creative Cloud) blogs.nvidia.com. This could manifest as Adobe’s 3D tools (like Substance 3D, Adobe Dimension, or even Photoshop’s 3D features) getting a “generate 3D” function using Edify under the hood. Adobe’s huge user base means this integration is poised to make generative 3D mainstream for millions of creatives, once rolled out.
- Omniverse and Industrial Partners: NVIDIA integrates Edify with its Omniverse platform (which is like a hub for 3D design and simulation, using Pixar’s USD format). They demonstrated Edify-generated assets being used in Omniverse workflows (e.g. WPP and Accenture using Edify 3D models to populate virtual environments for marketing) blogs.nvidia.com aibase.tech. Omniverse is also used in automotive, architecture, digital twins, etc., so having generative AI there means professionals can fill in world details or prototype designs rapidly.
- APIs for Developers: NVIDIA provides access to Edify models through NVIDIA AI Foundations – Picasso is one of these, alongside models like NeMo (for language). Developers can sign up to use these via cloud APIs. The integration here is that any software company can incorporate NVIDIA’s generative models by calling their API (with appropriate credentials and likely costs). For example, a fashion e-commerce startup could integrate the Edify API to let users generate custom jewelry designs on their website, without ever touching an NVIDIA GPU themselves – it’s all through Picasso’s cloud service.
- Hardware Integration: Since NVIDIA makes GPUs, they ensure Edify is optimized for their hardware. Big customers might integrate Edify into their own data centers using NVIDIA’s DGX Cloud or on-prem installations, effectively integrating at the infrastructure level. This matters for enterprises that want a self-hosted solution (for confidentiality or latency reasons).
Access, API Availability, and Pricing
The ability to access these technologies and their cost structure vary widely:
- OpenAI Shap-E Access & Pricing: Shap-E (the original) was released as an open-source GitHub repo with model weights ngwaifoong92.medium.com. This means anyone can download it and run it locally (provided they have a GPU with sufficient memory) or on cloud platforms. There is no licensing fee – it’s free under the terms OpenAI provided (likely MIT or similar license for the code and a lenient license for the weights). For casual users, there’s no official OpenAI web service for Shap-E yet. You can find community-run demos (which are usually free or have limited capacity) on sites like Hugging Face or small startups integrating it. As of mid-2025, OpenAI does not offer Shap-E through its paid API. All of OpenAI’s API offerings have so far been around text (GPT-3/4), images (DALL-E 2 via Bing or limited API for enterprise), and audio (Whisper). If Shap-E 2 becomes production-ready and in-demand, OpenAI could introduce it as a paid API (similar to DALL-E, which started as invite-only API access for businesses). Pricing then would likely be per generation or per second of compute, etc. But currently, to use Shap-E, one must self-host or find a free demo. The cost then is mainly compute – running a generation for 10 seconds on a GPU might cost a few cents of cloud GPU time. It’s quite accessible for hobbyists. In terms of commercial usage, since it’s open-source, companies can use Shap-E models internally without paying royalties, but they should consider the license. OpenAI usually releases research models under terms that allow reuse, but one should double-check if any restrictions apply (for instance, some OpenAI models have clauses against certain use cases). Assuming it’s standard open-source, a startup could incorporate Shap-E into their product at no cost (beyond computing). That is a big advantage for open innovation but also means support and updates are community-driven. Shap-E 2, if open-sourced similarly, would follow that model. If not, and it’s kept internal, then access might come via some OpenAI service. In summary, Shap-E access: open and free (DIY style); pricing: none officially, just compute costs on user’s side. For a user, that could be zero if using a free demo, or say $0.1–0.2 per model if using cloud GPU instances.
- Google 3D Shopping View Access & Pricing: For end-users (shoppers), access is simply via Google – it’s free to use as part of the search experience. If you have the Google app or go to google.com and search for a supported product, you can use the 3D viewer without any additional app or cost. It’s a value-add feature Google provides to keep users engaged. For merchants and advertisers, there is also no direct fee to have your products shown in 3D. If you’re listing products on Google Shopping (free listings) or running ads, Google will automatically create 3D spins given the image inputs (assuming you meet the criteria like providing multiple angles). Google’s Andrew Hutchinson writes that “Google is gradually rolling out its advanced object generation models for Shopping displays.” socialmediatoday.com There’s no indication of a charge for this – it’s part of Google’s platform. In fact, it may incentivize merchants to list on Google more. Advertisers who use Google’s services like Performance Max might indirectly pay for ads as usual, but the 3D feature is included to improve performance. From an API perspective, Google has not announced any public API where a developer can submit images and get a 3D model or spin. It seems to be an internal capability only for Google’s use. So if a retailer wants to get the 3D models Google generates, there isn’t a straightforward official way. (Perhaps one could scrape the frames from the viewer, but that’s not provided as a service.) Google’s priority is to keep users on its platform, not necessarily to provide generative 3D as a service to others. However, Google Cloud may integrate some of this tech in the future – for instance, the Vertex AI platform could offer a product imagery enhancement API that might include 3D view synthesis. If it did, it would likely be a paid API (Google Cloud services usually charge per usage). But again, as of August 2025, no such product is openly advertised. So, for businesses, the cost is essentially zero, aside from providing the imagery. This is a big deal because previously, making 3D models for your products could cost tens to hundreds of dollars per product (if hiring studios or using special rigs). Google removed that barrier. It’s “free” in monetary terms, though one might say the price is your data – Google gets to use your images to train and generate these models (the research blog implies they trained on synthetic data rather than merchant-provided photos for the model, so maybe that’s less of a concern). Bottom line: Access is through Google’s own interfaces, there’s no standalone app or API for others. Pricing is n/a for consumers and integrated into normal ad costs (if at all) for businesses. It’s essentially a free feature aimed at boosting the overall value of Google’s shopping ecosystem.
- NVIDIA Picasso Edify Access & Pricing: Edify is offered through NVIDIA’s cloud services and through partner platforms. There are a few avenues:
- NVIDIA AI Foundations (Picasso): Enterprises or developers can apply to use Picasso services which include Edify models. Initially, this was in early preview. In March 2024, NVIDIA announced Edify’s 3D and image models in preview, and by mid-2025 they shifted it out of preview (perhaps to a more available state) blogs.nvidia.com blogs.nvidia.com. Typically, NVIDIA would set up a contract or subscription for access. It might not be self-serve with a credit card (at least in early stages) – often you go through an NVIDIA rep. For instance, Getty and Shutterstock obviously have deep partnerships. However, NVIDIA did put some models on their Build (NIM) platform where developers can test them. If one wants to integrate, they likely pay for DGX Cloud hours or API credits. Pricing isn’t public, but we can guess it might be similar to other generative model APIs which charge per image or per model generated. An interesting clue: Shutterstock’s generative 3D API is offering “generative credit packs starting as low as $25” marketscreener.com. This implies a credit-based pricing (e.g. you buy credits, each 3D generation might cost a certain number of credits depending on complexity). $25 likely buys a small bundle of generations, suggesting maybe on the order of a few dollars or less per 3D model generation, which is remarkably low compared to manual modeling costs. This price is from Shutterstock’s end, but since Shutterstock has to cover NVIDIA’s costs, it reflects the economics. We can surmise that NVIDIA’s pricing to partners is wholesale (possibly cloud compute time plus licensing). Partners then mark it up or include it in subscriptions. For example, Getty might include a certain number of AI generations in a premium plan for enterprise clients.
- Partner Platforms: If you’re a user of Getty, Shutterstock, etc., you access Edify by using those platforms. Shutterstock’s early access generative 3D was in beta, likely at no extra cost for beta testers, but eventually, it will be a paid feature (with the credit packs as mentioned). Getty’s generative images are priced per image or as part of a plan (Getty’s model is interesting because they also provide indemnification and licensing for the outputs). We can expect output-based pricing: e.g. $X per 3D model generated, or a subscription that allows Y generations per month.
- Enterprise Fine-tuning: If a company wants to fine-tune their own Edify model (say a car manufacturer training a model on their vehicle designs), NVIDIA likely offers that as a service (NVIDIA AI Foundry). This would be a custom project, probably quite expensive (think cloud computing hours, plus possibly professional services). It’s geared to large enterprises with specific needs.
As for API availability, NVIDIA’s Build interface provides REST APIs for these generative services blogs.nvidia.com. So a developer can get an API key, send prompts, and get back outputs (images encoded or model files downloadable from a URL). It’s likely running on NVIDIA’s cloud (so you pay NVIDIA for usage, or the partner you’re going through). Because Edify is no longer in free preview as of June 2025 blogs.nvidia.com, using it beyond demos will require a paid plan or partnership. That said, NVIDIA did show some willingness to let developers play: they allowed test-driving Getty and Shutterstock models via NIM during GTC announcements blogs.nvidia.com. We might see a more public rollout where any developer can sign up and use a credit card to access generative services (similar to how OpenAI or Stability API operate). For now, it’s somewhat gated – likely to ensure quality and responsible use. On pricing specifics: we don’t have official numbers from NVIDIA’s side, but given that Shutterstock’s API is $25 for some package, using that as a benchmark, the cost per 3D generation might be on the order of single-digit dollars or even less if at volume. An enterprise contract might be a monthly fee plus compute fees. In short, Edify access is through closed beta/enterprise programs or via third-party platforms (Getty/Shutterstock) for end users. Pricing is currently enterprise-negotiated or credit-based via those platforms. It’s not cheap if you compare to a free open-source model, but it is cheap compared to hiring artists or photographers for similar output. And NVIDIA’s positioning is that the outputs are licensed for commercial use, which is part of what you’re paying for (peace of mind that you won’t get sued using them). We’ll likely see more pricing info as these services exit beta – possibly tiered plans for small businesses vs. big companies.
Quotes from Experts and Companies
To better understand these tools, let’s hear from the people behind them and industry observers:
- OpenAI on Shap-E’s innovation: The creators of Shap-E highlighted how implicit representations give it an edge. “We find that Shap·E matches or outperforms a similar explicit generative model given the same dataset… These results highlight the potential of generating implicit representations, especially in domains like 3D where they can offer more flexibility than explicit representations.” voicebot.ai. This insight from OpenAI’s researchers (Heewoo Jun and Alex Nichol) underlines why Shap-E’s approach was notable – it validated that focusing on implicit NeRF-like outputs can be just as good as direct point clouds, but with more versatility in rendering. It’s a clue that Shap-E 2 will likely continue down this path.
- Google on generative shopping models: Google’s blog announcement by Distinguished Scientist Steve Seitz conveys their excitement: “We’ve developed new generative AI techniques to create high quality and shoppable 3D product visualizations from as few as three product images… This technology is already enabling the generation of interactive 3D views for a wide range of product categories on Google Shopping.” socialmediatoday.com. Google is essentially saying: we’ve cracked the code on scaling 3D for commerce with minimal input. Another quote, emphasizing Veo’s strengths, from Google’s research blog: “A key strength of Veo is its ability to generate videos that capture complex interactions between light, material, texture, and geometry. Its powerful diffusion-based architecture… enable it to excel at novel view synthesis.” socialmediatoday.com. This was highlighted in Google’s messaging and picked up by tech media – it reassures that the spinning products will look realistic (shiny things stay shiny, textures look right under light). And on the results, Google said: “We discovered that this approach generalized effectively across a diverse set of product categories… [Veo] was not only able to generate novel views that adhered to the available product images, but it was also able to capture complex lighting and material interactions… challenging for the first- and second-generation approaches.” socialmediatoday.com. In other words, Google is proud that their generative 3D works not just for a sneaker but for glossy heels, metallic objects, fabrics, etc., overcoming previous tech limitations.
- NVIDIA and partners on Edify: NVIDIA’s own blog by exec Gerardo Delgado announced Edify’s expansion, noting partnerships: “Shutterstock 3D generation enters early access; Getty Images introduces custom fine-tuning for enterprises… Adobe to bring 3D generative AI to Firefly… Be.Live, Bria and Cuebric choose NVIDIA Picasso.” blogs.nvidia.com. This quote from the blog header showcases the momentum – multiple industry players adopting Edify in different ways. From Reuters, we have Getty Images CEO Craig Peters praising the NVIDIA collaboration: “This collaboration (with Nvidia) is testament to the feasibility of a path of responsible AI development… It is in line with our belief that generative AI is an exciting tool that should be based on permissioned data.” reuters.com. This underscores the ethical licensing angle of Edify – a subtle jab at open models trained on scraped data – and it positions Edify as the responsible choice. Another perspective from Greg Estes, NVIDIA’s VP of developer programs, about why companies want Picasso: “Other software providers or enterprises… don’t want to be involved [with generative AI] not knowing what the provenance is of the underlying training images.” reuters.com. Estes basically highlights that businesses fear “unstable diffusion” (unknown copyright status content), and NVIDIA’s curated approach assuages that. On performance, NVIDIA’s blog on Shutterstock’s 3D beta marveled: “The AI model first delivers a preview of a single asset in as little as 10 seconds… If users like it, the preview can be turned into a higher-quality 3D asset, complete with PBR materials… The 3D assets… are ready to edit…and available in popular formats. Their clean geometry gives artists an advanced starting point.” blogs.nvidia.com blogs.nvidia.com. This quote highlights exactly what 3D professionals care about – speed and clean output – coming straight from NVIDIA’s showcase.
- Industry watchers: Tech writers have also chimed in. For instance, Andrew Hutchinson (Social Media Today) commented on Google’s move: “Better product depictions, enabling more responsive, engaging shopping, while the process will also facilitate the expansion of Google’s 3D object corpus, which could also be fed into VR models… likely the next level of online shopping.” socialmediatoday.com. This quote frames Google’s 3D shopping as not just a neat gimmick but a strategic build-up of a 3D library that could power future VR shopping experiences. It shows the broader vision outsiders see: a race to own virtual commerce. Another observer, the tech blog Spacelab, wrote about Google’s tool: “Google just launched a new AI tool that transforms regular product photos into 360-degree 3D models… Next time you’re deep in a sneaker hunt… don’t be surprised if your shopping results start spinning back at you.” thespacelab.tv. This playful take highlights how mainstream this could become – users will begin to expect interactive spins as a normal part of shopping.
These voices collectively echo that:
- OpenAI’s Shap-E brought a fresh technical approach and more is anticipated from a sequel.
- Google’s advancement is seen as a game-changer for online retail, making shopping more immersive and potentially bridging into AR/VR.
- NVIDIA’s Edify is being lauded for its ethical stance and integration into professional workflows, with speed and quality that attract big-name partners.
Latest News and Announcements (Mid-August 2025)
The landscape of generative 3D is evolving rapidly. Here are the latest developments for each as of August 2025:
OpenAI Shap-E / Shap-E 2: While OpenAI has been quiet about any official “Shap-E 2” release, the community and leaks suggest something may be brewing. In late 2024, some AI enthusiasts speculated Shap-E 2 would arrive in 2024 or 2025 with improved detail madoverstories.com. Indeed, a Medium blog’s “top AI art generators to watch in 2024” listed Shap-E 2 as a “hypothetical sequel… expected to generate 3D images with more detail… from text descriptions.” madoverstories.com. This indicates that people expect OpenAI to double down on 3D. There have been no official OpenAI blog posts or papers on Shap-E beyond the initial release, but OpenAI’s focus on multimodality (e.g. releasing the GPT-4 Vision update for images) might extend to 3D later. It’s possible that Shap-E 2 could be folded into a more general model – for example, OpenAI’s rumored “GPT-5” or future multi-modal systems might handle text, images, and 3D jointly. There’s also interest in whether OpenAI will integrate 3D generation into its developer platform; an item in OpenAI’s leaked roadmap (hypothetical) could include a “Point-E/Shap-E API”. As of mid-2025, however, no new official Shap-E versions have been announced. The latest news in the open-source world is that others are building on Shap-E: for instance, a project called Shap-Explorer (2023) created a GUI to manipulate Shap-E outputs in real-time figshare.com. Academic research also continues to cite Shap-E and propose improvements. In summary, the “news” is mainly anticipation – Shap-E 2 is one to watch, especially as competitors forge ahead. If OpenAI wants to stay at the forefront, we might hear something in late 2025 or 2026 about their next 3D model initiative.
Google 3D Shopping (Veo) Updates: Google’s 3D shopping feature officially launched around May 2025 research.google, and since then Google has been expanding it. By August 2025, Google confirmed that 360° spins are available for a wider set of products – not just shoes, but also categories like apparel (e.g., handbags) and home goods. A Google marketing article noted: “Over the last year, we launched… 3D spins for ad images. We also expanded virtual try-on to dresses, and integrated this experience into ads as well.” business.google.com. This was from Vidhya Srinivasan (Google’s VP of Ads and Commerce), indicating that by early 2025 Google had rolled out 3D spins in advertising and was linking it with their virtual try-on for clothing. In the news cycle, Google I/O 2025 (held in May) featured shopping announcements – according to AI Business, Google introduced “a new range of AI-powered shopping features” at I/O 2025, including virtual try-on and generative 3D views aibusiness.com. This suggests that Google is heavily promoting these features as part of their AI-driven future of search. By mid-August, some users in the US are seeing 3D spins on more search results, and feedback seems positive. Google’s Next steps likely involve launching this globally. Also, competition is responding: Amazon, for instance, might accelerate its own 3D/AR shopping initiatives. But Google’s advantage is that it can do this for any retailer’s products via search. Latest announcements also tease deeper integration of these generative models into Google’s Cloud offerings under the Gemini AI brand (Google’s next-gen foundational model suite). A Google AI tweet from July highlighted using generative AI (including Veo) to transform 2D images to 3D for shopping x.com. This public communication shows Google’s confidence in the tech. So the latest news: It’s out in the wild for footwear and more, merchants in the US can now use it (anecdotally, Google has likely contacted merchant partners to ensure they have enough images to utilize it). No word yet on an external developer API, but if Google sees demand (say large retailers want to use the tech on their own app), they could offer a private capability through Google Cloud. For now, mid-2025, it’s a Google exclusive feature fueling their e-commerce experience.
NVIDIA Picasso Edify News: Lots has happened here by mid-2025. In June 2025, as noted earlier, NVIDIA closed the “preview” of Edify microservices – implying that the service might be moving to general availability or a new phase blogs.nvidia.com. Around July/August 2024, NVIDIA and partners made a splash at SIGGRAPH with generative 3D. Shutterstock’s Generative 3D API was announced in July 2024 and went into beta, and by mid-2025 it’s in commercial beta/early access for enterprise clients blogs.nvidia.com blogs.nvidia.com. In fact, on July 30, 2024, Shutterstock’s press release proudly called it the “first ethical generative 3D API” marketscreener.com, highlighting the exclusive training on Shutterstock content. This service was showcased at SIGGRAPH 2024 with live demos (Blender plugin, 3D printing on site) blogs.nvidia.com. By 2025, we have real enterprise use cases: WPP (a big ad firm) integrating it for virtual production, Mattel using it for toy design, etc., which we already touched on blogs.nvidia.com aibase.tech. Getty Images in the meantime upgraded their Edify-powered service in mid-2024 to be faster and more feature-rich blogs.nvidia.com blogs.nvidia.com, and that service is fully live on their site. In September 2024, Getty’s CEO was publicly discussing their Gen AI tool, which suggests by 2025 it’s part of their regular offerings. Another tidbit: Adobe MAX 2024 (late 2024) and Adobe Summit 2025 might have revealed more about Adobe’s integration – possibly announcing Firefly for 3D beta, though specifics are scarce, everyone’s expecting Adobe to launch generative 3D features soon with NVIDIA’s help. On NVIDIA’s end, the next big event is likely GTC 2025 (March 2025). It wouldn’t be surprising if Jensen Huang (NVIDIA’s CEO) announced that Picasso and Edify are generally available for all businesses, with success stories from early adopters. They might also introduce Edify Model 2 or improvements (for instance, a version that can generate higher-resolution or animate 3D objects). NVIDIA Research is also active in related areas, like 4D (dynamic 3D) generation and simulation, which could feed into Edify’s evolution. Another piece of news: OpenUSD consortium – NVIDIA, Pixar, Apple, and others are pushing Universal Scene Description (USD) as a standard. In 2025, with more 3D content being AI-generated, NVIDIA is likely advocating that Edify outputs adhere to USD for seamless pipeline integration (which they mentioned with Dassault Systèmes and others adopting USD workflows with generative HDRi in Omniverse blogs.nvidia.com).
In summary, as of August 2025: Shutterstock’s Gen-3D and Getty’s Gen-AI are in market, Edify is out of preview and presumably now a product developers can sign up for (with NVIDIA likely charging usage fees). The industry is abuzz with partnerships – you can almost sense a race between cloud providers: NVIDIA’s Picasso vs. emerging offerings from others (perhaps OpenAI or stability.ai might attempt something similar for 3D). But NVIDIA’s head-start with enterprise ties is clear in the news. One more news angle: Regulatory/ethical – Because Edify focuses on licensed data, Getty and Shutterstock often mention the “ethical” label marketscreener.com. This has attracted positive press as companies worry about AI training lawsuits. So far, Edify’s approach has avoided the legal troubles that hit others like Stability AI. That’s a newsworthy achievement in itself by mid-2025, given how many generative models were facing court cases – Edify’s partners are instead launching products comfortably.
Upcoming Features, Roadmap, and Rumors
What does the future hold for these three? Let’s indulge in some well-founded predictions and rumors:
Shap-E 2 and OpenAI’s 3D Roadmap: With all the buzz about Shap-E, it’s natural to expect OpenAI will push the envelope further. Insiders speculate that OpenAI’s next multimodal model might integrate 3D understanding or generation. For example, a future GPT could output code for 3D scenes or directly invoke a model like Shap-E under the hood. There’s also potential that OpenAI might combine its image and 3D models – imagine generating an image with DALL-E 3 and simultaneously getting a rough 3D model of the scene. This could be powerful for consistency across media (think video game concept art and instant assets). In terms of leaks: nothing concrete has surfaced about Shap-E 2’s specs, but given the trend, we’d expect:
- Higher resolution outputs (maybe 4K texture support, more detailed geometry).
- Possibly the ability to generate not just single objects, but simple scenes or multi-object arrangements via textual scene descriptions.
- Integration with ChatGPT plugins – maybe a 3D viewer plugin so ChatGPT can show you a generated model right in the chat.
- More modal control – e.g. the ability to input a sketch or a rough 3D model and have Shap-E refine or texture it (going from coarse to fine).
OpenAI might also have a eye on competition: Google and NVIDIA are proving practical use-cases; OpenAI might aim for a more consumer-friendly killer app for 3D generation. Perhaps a partnership with a platform like Sketchfab or an AR app to bring Shap-E to a broader audience. Since Shap-E started as open-source, one question is whether Shap-E 2 will also be open or if OpenAI will keep it proprietary if it’s much improved (similar to how they handled GPT-2 vs GPT-3). The community hopes it remains open so the 3D research momentum continues. On timeline, if nothing was announced by mid-2025, maybe late 2025 or 2026 could see something (OpenAI’s dev cycles for major models are often 1-2 years). It’s also possible OpenAI is focusing elsewhere (like enhancing DALL-E or their coding models) and is happy to let others handle 3D for now. But given how central 3D is to the future of AR/VR and even web (think e-commerce), it’s likely on their radar. So, keep an eye out for OpenAI DevDay 2025 or research papers from OpenAI’s team that might drop hints of a “Shap-E v2” in the works.
Google’s Future Features: Google will likely expand 3D Shopping View to more product categories. By late 2025, we might see generative 3D for clothing (though clothing is much harder due to deformability – but they are tackling it via virtual try-on for models). Perhaps accessories like hats, toys, beauty products (360° of perfume bottles, etc.). One can imagine a timeline where any product image search results in a 3D view option if images are available. Google will also probably improve the fidelity further: maybe using higher resolution generation or enabling zoom (currently, 360 spins might be at a fixed resolution; future versions might let you zoom in to see fine texture, which would require more detail generation on the fly). Also, Google could integrate lighting control for users – e.g., a user might adjust lighting in the 3D viewer to see how a product looks in different conditions. Because the model inherently knows about lighting (via Veo’s training), this isn’t far-fetched. Another likely feature: Background replacement. Google already has AI that can generate backgrounds for product images. They could allow users to toggle backgrounds in the 3D view (see the product in a living room, or on a model, etc.). On the merchant side, Google might open up some tools: maybe a feature in Merchant Center to preview the AI-generated 3D model of your product and allow merchants to flag if it’s incorrect. That would help them get feedback to improve the model (for instance, if the AI hallucinated something wrong, the merchant could correct it). In terms of rumor, there’s the broader Google project Gemini (a next-gen multimodal model expected late 2023/2024) – some speculate parts of Gemini could handle vision and 3D jointly given DeepMind’s expertise in 3D (they did AlphaFold for 3D protein structures, for example). Google might unify these technologies so that one AI system can parse images, generate 3D, generate video, etc. If that happens, we may see developers getting access to these capabilities via Google’s API. For example, a Google Cloud API that given images yields a 3D model (which could compete with NVIDIA’s offering). So far, Google hasn’t commercialized it externally, but pressure from industry might cause them to, especially if retailers say “we want that tech on our app.” Also, Google’s moves might push Apple or Meta to do something similar – Meta has showcased related research (like NeRF in smartphones, and generative tech for avatars). If others join the race, Google will accelerate improvements. Summing up: expect Google to make 3D spins a standard part of online shopping, possibly connect it with AR search results, and maybe by 2026 offer it as a service for retailers and developers. They’ll be refining Veo and maybe training a specialized model for each product domain (one for shoes, one for furniture, etc., or a single multimodal powerhouse that handles all).
NVIDIA Edify’s Roadmap: NVIDIA is known to iterate quickly in AI. One thing to watch is whether they announce a new version of Edify (say Edify 2.0) with the latest research improvements. NVIDIA Research has been working on things like latent NeRFs, better 3D segmentation, and even generative models that produce dynamic scenes (with motion). We might see Edify incorporate capabilities to generate animated 3D models or sequences – for example, generating a short 3D animation from a prompt (“a bird flapping its wings”). Currently Edify covers static 3D and video separately, but a fusion could happen. There’s also the possibility of real-time 3D generation with upcoming more powerful GPUs – maybe not fully real-time in 2025, but eventually being able to tweak a prompt and see the 3D object update live would be huge for creators. NVIDIA will also likely deepen integration with Omniverse and CAD software. They might provide plugins for tools like Maya, 3ds Max, or SolidWorks to directly generate design prototypes via Edify (imagine a car designer using Edify to generate concept models in their CAD suite). In terms of leaks, we know from their blogs that Adobe integration is coming, which implies by late 2025 Firefly will have some 3D features (like generating a 3D object or material). Also, Unity or Unreal Engine might partner with NVIDIA to integrate generative AI for game devs (NVIDIA already works closely with Unreal on RTX and DLSS tech, so adding Picasso services for asset generation could happen). Pricing might become more transparent as it goes GA – possibly a usage-based model similar to cloud computing charges (maybe a few cents per second of GPU usage, etc.). Another interesting frontier: edge deployment. Right now Edify runs on cloud GPUs. But NVIDIA might optimize smaller versions of generative models to run on local GPUs for interactive applications (especially with their new powerful GPUs in workstations). For example, a future Nvidia RTX GPU for consumers could come bundled with generative AI software for creators, running a trimmed Edify locally for instant results (this is speculative, but aligns with NVIDIA’s push to sell more hardware by giving it unique AI abilities). On the content side, Edify’s reach may grow – more stock media companies might join (perhaps Adobe Stock, or Sketchfab for 3D). We saw Ultralytics writing an article about Edify 3D ultralytics.com, which shows even AI communities are eyeing it. Perhaps Getty and Shutterstock’s success might spur competitors like Adobe Stock or Freepik to also license Edify or similar models. So NVIDIA’s likely future moves involve dominating the generative content backbone – if someone is generating any visual media, NVIDIA wants their tech (and thus their GPUs) in the loop. Roadmap likely includes:
- Fine-tuning ease – letting more clients bring their own data to train custom versions (NVIDIA Edify being an “AI foundry” means they will streamline that process).
- Possibly incorporating multi-modal prompts for 3D (like using text + reference image + sketch all together to refine outputs).
- Enhanced outputs: maybe generating materials and animations along with static models.
- Competition: NVIDIA will face competition from startups and open models in this space. There are open projects (though none as advanced with licensing). Stability AI, for instance, might attempt a 3D model (they had DreamStudio for images; maybe “DreamFusion Studio” someday). If that happens, NVIDIA will emphasize the proprietary advantages (quality, data, integration).
In rumor-land, there were whispers of NVIDIA working on an AI model called “Edify for Video” (since Veo is Google’s, NVIDIA might have their own video generation separate from images). If true, Picasso will cover all visual media – an all-in-one. But focus wise, for 3D, the rumor mill has been mostly positive: no negative press yet, just anticipation of when it opens up fully for public use.
Conclusion
In this rapidly advancing arena of AI-generated 3D, each player brings unique strengths:
OpenAI’s Shap-E 2 (and its predecessor) shines as a creative tool – ideal for inventing new objects on the fly from a simple prompt. It democratizes 3D model generation by being open and (relatively) lightweight. Shap-E can empower individual artists, game modders, and small studios to prototype without 3D modeling skills. Its future likely holds greater realism and integration (perhaps one day you’ll ask ChatGPT to “make a 3D dragon” and it’ll send you the file). However, as of now, it’s somewhat limited to enthusiasts and the output quality, while impressive, isn’t fully photoreal. It’s where imagination runs wild, and it could evolve to challenge the others if OpenAI pours more resources into it. Keep an eye out for OpenAI’s next moves here – a Shap-E 2 could bridge the gap and even integrate with VR/AR content creation.
Google’s 3D Shopping View is a targeted masterstroke – it takes a real-world problem (product visualization) and solves it with AI at scale. Where it shines is user experience and trust: shoppers get interactive, realistic product models that instill confidence in what they’re buying. Google leveraged its AI research to create a solution that slots naturally into its products, potentially giving it an edge in e-commerce battles. In this “competition”, Google’s tool isn’t directly competing with Shap-E or Edify on features; instead, it’s competing for adoption in retail. And it’s doing well – retailers don’t have to lift a finger or pay extra, and suddenly their listings become 3D experiences. Google’s likely to extend this lead, making online shopping ever more immersive. In the long run, Google’s approach could expand beyond shopping – imagine Google Earth integrating generative 3D to reconstruct interior spaces, or YouTube allowing 3D product showcases generated from 2D videos. For now, Google’s tool shines in practicality: it has one job (show the product) and does it excellently thespacelab.tv. Its proprietary nature means others (like Bing or Amazon) will need their own solutions to keep up – so expect competition in commerce AI. But Google’s head-start and integration across search, ads, and AR make it a formidable pioneer in mainstream 3D content delivery.
NVIDIA’s Picasso Edify emerges as the Swiss Army knife and powerhouse for generative visuals in the enterprise. Its strength lies in quality, customizability, and ecosystem. By focusing on licensed data and partnering with industry leaders, NVIDIA ensured Edify is ready for commercial use, not just tech demos reuters.com. It is the platform that companies can build on – whether it’s a stock agency offering on-demand 3D assets, an ad agency generating virtual scenes, or a game studio populating a world in minutes. Edify’s outputs are already impressively realistic and editable, and it’s connected to a whole pipeline (Omniverse, Adobe, etc.), meaning it can slot into existing production workflows rather than replace them. NVIDIA has effectively positioned Picasso Edify as the backend for the Metaverse content factory. Where it might lag is direct consumer reach – you won’t see kids using Edify directly the way they might play with a DALL-E app. But you will see the fruits of Edify in media, games, ads, and more, without knowing it (like how “Powered by NVIDIA” appears subtly in game intros). Edify’s future looks bright and is somewhat roadmapped – more features, faster generation, deeper integrations – all driving the demand for NVIDIA’s hardware and cloud. In competitive terms, NVIDIA is ahead in B2B generative visuals, but they’ll have to continue justifying their not-inexpensive services against open-source alternatives that might arise. Still, the combination of data quality, speed, and enterprise support is NVIDIA’s moat.
In the bigger picture, these three aren’t so much direct competitors as they are parallel efforts addressing different needs:
- Shap-E (OpenAI) – agile creativity and research.
- Google/Veo – tailored consumer experience in shopping.
- NVIDIA/Edify – heavy-duty content generation infrastructure for industry.
It’s entirely conceivable that in a few years, all three could coexist and even complement each other. For instance, a game developer might use OpenAI’s tool to brainstorm asset ideas, then use NVIDIA’s platform to generate them at production quality, and finally use Google’s techniques to create 360° marketing spins of those assets for promotion. The future of 3D AI is one of convergence – as hardware, algorithms, and data improve, generating a 3D world might become as easy as writing a paragraph. We’re seeing the first steps of that with these three.
No matter who “wins” the race, the real winners are creators and consumers. Creators are gaining superpowers – a solo indie developer can now do in hours what once took a team weeks (modeling, texturing, rendering). Consumers gain richer experiences – shopping becomes interactive, learning becomes more visual (imagine textbooks with AI-generated 3D models of historical artifacts you can inspect).
Each tool we discussed excels in its domain: OpenAI in open creativity, Google in user-centric design, NVIDIA in enterprise-grade solutions. As competition heats up, we’ll likely see each borrow ideas from the others (OpenAI might incorporate more training data and focus on quality, Google might expose APIs, NVIDIA might streamline to attract smaller creators).
In conclusion, OpenAI Shap-E 2, Google’s 3D Shopping View, and NVIDIA’s Picasso Edify are collectively pushing 3D content generation from a futuristic idea to an everyday reality. Whether you’re a shopper spinning a sneaker, a developer conjuring a 3D model from thin air, or a brand rendering an entire ad campaign on the fly – this trio of AI innovations is shaping the 3D future, one prompt at a time.
Sources: OpenAI/Voicebot coverage of Shap-E voicebot.ai voicebot.ai, Google AI Blog & press on generative shopping research.google socialmediatoday.com, NVIDIA and partner announcements on Edify reuters.com blogs.nvidia.com, and industry analysis socialmediatoday.com, as detailed throughout this report.
