--- title: StableGen emoji: 🎨 colorFrom: green colorTo: indigo sdk: gradio sdk_version: 6.31.0 app_file: app.py short_description: Texture 3D meshes from a prompt with SDXL projection python_version: "3.12" startup_duration_timeout: 30m license: gpl-3.0 --- # StableGen texturing demo A ZeroGPU port of the core texturing workflow from [StableGen](https://github.com/sakalond/StableGen), a Blender add-on by sakalond. It follows the **SDXL Default** preset with the **Generate Sequentially** workflow described in the [wiki](https://github.com/sakalond/StableGen/wiki/Generation-modes): 1. Cameras are placed around the mesh, and a depth guide is rendered from each one (normalized and inverted, near = white). 2. The views are generated one at a time with RealVisXL V5.0, the SDXL-Lightning 8-step LoRA and a depth ControlNet (`xinsir/controlnet-depth-sdxl-1.0`, strength 0.5), using 8 steps at CFG 1.5. 3. Each later view inpaints only the regions earlier views have not covered. A render of the texture blended so far is the context, and differential diffusion uses a soft coverage mask (ramp 0.15 to 1.0, grown by 3 px and blurred). 4. Every view is projected onto the surface. Views are blended per texel with weight cos(angle)^3, taking occlusion into account, and baked into a UV texture (xatlas unwrap). The result is exported as a GLB. This demo approximates the Blender add-on rather than running it, so the differences are listed here: - Rasterization uses nvdiffrast instead of Blender/Cycles. - Cameras are placed on an orbit instead of the add-on's normal-weighted auto placement. - The add-on's `dpmpp_2s_ancestral` sampler is replaced by Euler ancestral with trailing spacing. - No IPAdapter is used. ## Example meshes The meshes are from the [Khronos glTF Sample Assets](https://github.com/KhronosGroup/glTF-Sample-Assets), all CC0-1.0, with the textures stripped: Suzanne (Norbert Nopper), Lantern (sbtron / Frank Galligan), BarramundiFish and Corset (Microsoft).