Image: GitHub
Why it mattersA marketing team that needs one 3D model can hand a photo to Claude Code or Cursor. The agent returns the mesh without the user learning a 3D app.
Gokay, who publishes as gokayfem on GitHub, released image-to-sculpt on 8 November. The repository has gained 289 stars and 21 forks in six days. The project calls itself "a skill for coding agents (Claude Code, Codex, Cursor, Copilot, Gemini)". A user hands the agent a photo. The agent runs the skill and produces a 3D mesh that matches that photo, inside Blender.
How it runs
The README names eight phases. Scaffold builds a case folder from the input photo. Prep pulls the depth, the focal length and the normals from the image using the MoGe and Marigold models. Views asks the fal API for nine extra views of the subject, plus their masks and normals. Hull registers every view's camera, carves a visual hull and finds the mirror plane for symmetric subjects. Fit then runs 38 recorded passes that refine the shape in stages: forms, surface, detail, polish. Cut and Smooth open gaps and remove loose faces. Check validates every output map against the first photo. Blender renders the replay as a timelapse and exports the mesh as GLB, OBJ or STL.
The author states a cost of around $2 per case through the fal API. The repository uses Python 3.10 or later, Blender 5.2 or later, and FFmpeg. Everything the author lists is MIT licensed.
What makes it different from a one-shot
The README lists three claims, each about a built-in check. A view the agent asks for at 30 degrees may be returned at 72 degrees by fal. The skill detects the real angle at 72.0 degrees. So a wrong view does not pollute the fit. One of the nine generated views is held back before the fit starts. It is used only to grade the result at the end. Normal angles are reported to 0.3 degrees. The signed distance field is reported to a median of 0.07 voxel.
Where it does not work
The README lists four cases the skill refuses. Full scenes and rooms. Hair-thin structures. Clear or mirrored surfaces. Multi-subject photos with no SAM-3 selection to split them apart. The author writes these as limits a user should know, not as faults to patch around.
Most image-to-3D tools are a black box. A photo goes in. A mesh comes out. The user has no record of how the tool got from one to the other. This skill works differently. Every carving step is logged. Every generated view is checked against the first photo. Blender can replay the whole build as a video. A product team doing one 3D asset for a marketing image can watch this trail. It is the difference between a mesh they keep and a mesh they throw away. For a coding agent whose output fails one check out of ten, the trail is where it looks first.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.