Viral GPT-6 3D Anatomy Demo Used Existing Assets; ifanr Builds New One With Hyper3D MCP

The original demo drew more than 32 million views. In the video, a complete human body could be separated layer by layer, with muscles, bones, blood vessels and organs becoming independent 3D parts. The post said GPT-6 generated 2,234 human body parts in one pass. Users initially concluded that GPT-6's 3D capability had reached a new level.

The finding changed that view. The 2,234 parts were not generated by GPT-6 from zero; they came from an existing professional 3D dataset made by experts based on MRI data. GPT-6's actual role was to call those ready-made models and organize them into an interactive webpage, according to ifanr.

As a general model, GPT-6 is not strong at generating complex 3D assets. It can use Three.js to build a 3D scene or use Computer Use to open Blender and model, but the results remain like the early video-generation example of 'Smith eating spaghetti': the actions are correct, but the outcome looks off. Direct GPT-6 generation works like stacking blocks, with each part placed, but proportions, details and connections appear stiff. It is better suited to low-complexity 3D scenes assembled from basic geometry, such as cubes, spheres, simple mechanical structures, low-poly buildings and abstract installations, where interaction and spatial relationships matter more.

The report said it found a GPT-6 + Hyper3D MCP combination that assigns 3D modeling to a professional 3D generation model. GPT-6 can use its general world knowledge, while the 3D generation model handles form and materials; GPT-6 can also use Computer/Browser Use tools to refine results. Hyper3D Rodin, described as a top AI 3D generation model, handles real objects with complex forms, rich surface details and needs for complete meshes and materials, including human bodies, animals, organs, complex props and architectural assets.

The workflow is largely a conversation with Codex. After a clear prompt, Hyper3D's MCP connects to Codex and handles 3D model generation in the background. For a prompt such as 'make a 3D car model,' Codex checks available tools, identifies Hyper3D and uses Hyper3D Rodin. The MCP reports progress; in the car example the task is split into five stages, and users can see when paint, glass and tire material textures are being generated.

To connect, users go to the Hyper3D website, click MCP in the upper right corner, copy the prompt that appears and send it to Codex. The prompt configures Hyper3D MCP in the current client, Codex or Claude Code: service name hyper3d, Streamable HTTP address https://api.hyper3d.com/api/mcp, and OAuth authorization, while preserving other settings. After configuration, Codex lists available Hyper3D tools, including 3D model generation with Hyper3D Rodin, recursive part splitting through BANG, progress queries and result retrieval.

ifanr then opened a new conversation and asked Codex to make an interactive 3D human dissection webpage, temporarily named 'Human Dissection Room,' using Hyper3D Rodin to generate the main 3D assets, then download, assemble and inspect them. The report said Codex took a long time and repeatedly checked images generated by Hyper3D Rodin, for example finding that the bottom of both lungs had been generated with too-deep cavities and an unsuitable shape. Hyper3D continued generating parts based on GPT-6's feedback, and GPT-6 automatically selected Rodin's Medium mode and the corresponding image-to-3D path. The final result was a complete head-to-toe muscular human body using 18 newly generated Hyper3D Rodin assets and offering 46 selectable regions.

In a separate project, ifanr changed the prompt to use GPT-6 to generate the main 3D assets instead of Hyper3D Rodin. The result was crude: GPT-6 produced transparent mannequins with organs inside rather than a complete body of muscles, bones and blood vessels. Because GPT-6 could only use the Three.js library, materials looked flat, and organs and bones were rendered in a single matte color, reducing surface detail and the impact of the 3D view. Hyper3D Rodin's assets, by contrast, included continuous body contours, muscle bundle relief and organ-specific shapes and surface textures. Those details were already present in the downloaded models, and Codex then handled position, scale and disassembly interaction. A generated heart model was so realistic that the screenshot was blurred.

Using the same combination, ifanr made a disassemblable steam beetle. It has a dark green enamel shell, brass edging and rivets and wear textures. A slider opens the two elytra, separates the head shell and body shell, and reveals gears, pistons and a boiler inside. The mechanism keeps running even after the shell is opened. A single elytron can be selected, rotated and changed to another enamel color, then reassembled. Hyper3D Rodin generated PBR material information, including surface color, fine bumps and different levels of metalness and roughness. Hyper3D's BANG recursive part-splitting function automatically separates the model into independent parts that can be selected, disassembled and interacted with in the Codex-generated webpage.

Another project put a black trench coat and sunglasses on an orange cat leaning back, with silver sardines flying past; one fish is nearly touching its face. The webpage freezes the moment, lets users orbit behind the cat to see the coat folds, and provides a time slider to move the fish forward or backward. The main assets were only a cat and a fish; the page reused the same fish eight times and calculated their positions. The cat remains leaning back while fish movement, camera orbit and greenish cinematic lighting create a 'bullet time' effect.

A Guangzhou street corner scene was made from five independent Hyper3D Rodin assets: two qilou buildings, a breakfast cart, a stone lion and a banyan bonsai. The buildings have green shutters upstairs and a continuous arcade below. Pulling the view back makes the street look like a tabletop miniature; street mode lowers the camera to pedestrian height so users can walk along the arcade. A drag assembly slider shows the road first, then buildings, then small street objects.

Other demonstrations included a bat observation page where the mouse acts as a flashlight: a fixed bat hangs in the dark, the inspection light follows the cursor, and clicking the head, wing membrane or claws moves the camera closer. Users can also rotate the model to see its back and the underside of the wings.

A game prototype used a teapot with antlers on a table, a lamp casting its shadow on a wall, and a passport showing the silhouette customs requires. Players rotate the teapot and adjust pitch and roll to align the shadow with the target. The direction of the antlers, the curve of the spout and the gap in the handle all change the shadow. When the outline matches, the passport receives an 'approved' stamp. The three levels use the same teapot, starting untimed and adding time pressure. The first teapot had antlers that were too small; after new reference images emphasized unequal large antlers, a curled handle and a curved spout, Hyper3D Rodin generated a new version. Codex connected projection, silhouette comparison, timing and stamp feedback. Hyper3D MCP was also used to generate ten more detailed antique 3D assets to increase difficulty.

Hyper3D Rodin comes from Hyper3D, a 3D generation model company under Yingmou Technology. The company was founded in 2020, and its core founding team comes from ShanghaiTech University; it has long worked on 3D modeling, graphics and generative AI. Hyper3D has appeared in a workflow demonstration in Jensen Huang's CES 2026 keynote and in the champion project of OpenAI's first Codex hackathon. Its scene generation research CAST won the Best Paper Award at SIGGRAPH 2025, studying how to generate a complete 3D scene containing objects, object relationships and physical constraints from a single image. In September this year, Hyper3D evolved CAST into the world generation model WorldGen, whose generation capability has entered real production scenarios including embodied intelligence simulation training, film and television production, game development and spatial computing.