StableGen

StableGen is an open-source Blender add-on for generating meshes, texturing existing objects and scenes, and preparing rigged 3D assets. It uses a ComfyUI server to run the selected models, then brings their results into Blender. The server can run on the same computer or another machine on the network. The add-on is developed by Ondřej Sakala and distributed under the GNU General Public License v3.0.^[1]

How it works

StableGen places cameras around scene geometry and renders guidance images from those viewpoints. Models generate or edit the corresponding images, which the add-on projects onto visible mesh surfaces. Overlapping projections receive weights based on visibility and the angle between the viewing direction and surface normal. Cameras contribute nothing to hidden surfaces or surfaces rejected by the angle threshold.^[2]

This makes camera placement part of the texturing process. A surface outside every camera's coverage receives the fallback color. Reprojecting recalculates the material from saved camera images after blending settings change; it does not run image generation again.^[2]

The SDXL architecture can use ControlNet for structural guidance and IP-Adapter for image references. Qwen-based image-editing workflows instead take reference images, such as depth maps, previous views and renders of the developing texture.^[3]

Texturing workflows

The available workflow depends on the selected architecture.^[4]

SDXL and MOSS Image support the text-to-image workflow family, including Grid and UV Inpaint. Qwen Image Edit, Qwen-Image-2.1, Qwen Image 2.1 Turbo and FIBO 2.0 Edit offer Generate, Refine, Local Edit and Masked Edit instead. Their Generate presets process views sequentially. Camera order matters because the first view establishes appearance for subsequent views.^[4]

Masked Edit lets a user paint a selection, describe a change and preview the result before applying or discarding it. It operates through the cameras that see the selected surface and supports multiple component meshes without permanently joining the originals.^[5]

Model architectures

StableGen separates texture generation from mesh generation. Its architecture selector identifies supported, experimental and deprecated options.^[3]

The 0.4.0 release, published on October 28, 2026, added the new panel, Masked Edit, automatic rigging and Qwen Image 2.1 support. It also marked FLUX.1 and FLUX.2 Klein architectures as deprecated and hid them by default, while retaining an option to expose them for existing files.^[6]

Mesh generation and rigging

The TRELLIS.2 integration uses a ComfyUI custom node to generate a mesh from an input image. For a text prompt, StableGen first creates a reference image. A preview gallery lets the user select an image before mesh generation continues. The resulting GLB is downloaded and imported into Blender. Texturing can use the native TRELLIS.2 texture mode or a separate diffusion architecture.^[7]

SkinTokens supplies automatic skeleton generation and skin weights through the ComfyUI server. It can run on an existing character mesh or between the mesh-generation and texturing stages of a TRELLIS.2 job. StableGen can also retarget animation from FBX or BVH files onto a rig.^[8]

Materials, baking and export

StableGen can estimate material maps from the generated camera images and project those maps onto the mesh. Available channels include albedo, roughness, metallic and normals; optional channels include height and experimental emission. Ambient occlusion uses Blender geometry rather than an image-generation model. The documentation explicitly distinguishes estimated PBR maps from measured material properties.^[9]

A projection material combines camera images at rendering time. Exporting it for another application requires baking into ordinary UV textures first. StableGen supports glTF Binary, glTF with separate texture files and FBX export, including rigs and optional animation actions. Export itself does not perform the bake.^[10]

These steps connect StableGen to a broader game-asset creation workflow. The source documentation also describes orbit renders and multicolor 3D-printing export.^[10]

Installation and hardware

StableGen requires both the Blender add-on and a working ComfyUI installation with the relevant models and custom nodes. The project's installer downloads dependencies for the selected features. The documentation lists these requirements:^[11]

On Blender 4.2-4.5, projection uses Open Shading Language and the add-on switches Cycles to CPU rendering while constructing projection materials. Blender 5.1 and later use native Raycast shader nodes.^[11]

Example scene workflow

In the linked street-scene demonstration, Sakala described using TRELLIS.2 meshes for individual props and a character, with simple blocks for buildings, road and sidewalk. Qwen Image 2.1 generated ten camera views for texturing. Per-camera prompts named the objects visible from each viewpoint, and the character received a separate camera setup. Separate texturing runs produced photorealistic, cyberpunk, anime and papercraft treatments. The author reported baking the results and rendering in Cycles.^[12]

The demonstration's reference-image logo editing was identified as development-branch functionality awaiting the next release. It should be distinguished from the released Masked Edit feature.^[12]

Licensing

StableGen's software license does not replace the licenses of its model weights or dependencies.^[1]

The Qwen Image 2.1 license grants rights for non-commercial use and defines that use as research or evaluation. StableGen's documentation identifies Qwen Image 2.1 Turbo as subject to the same license. Its architecture documentation also identifies FIBO 2.0 Edit as non-commercial under CC BY-NC 4.0.^[3][13]

The TRELLIS.2 integration documentation separately identifies restrictions on the NVIDIA libraries used by its Native texture mode. Those libraries are not used by the shape-only or projection-based texture paths. Model-specific terms still apply when another texture architecture is selected.^[7]

Research context

StableGen's documented components draw on research in conditioned image generation and multiview texturing. ControlNet adds spatial conditioning, including depth and edge inputs, to pretrained diffusion models. IP-Adapter introduces image-prompt conditioning alongside text. TEXTure studies iterative painting of a 3D shape from different viewpoints and the consistency problems caused by independently generated views.^[14][15][16]

These papers describe the underlying methods and related research. Their evaluation results are not performance measurements of StableGen itself.

Continue reading

• ComfyUIComfyUI is a free, open-source node-based graphical user interface and backend for generative AI workflows, primarily focused on image generation using diffusion models such as Stable Diffusion and FLUX.Mentioned 5 times above
• CControlNetControlNet is a neural network architecture that adds spatial and structural control to large pretrained text-to-image diffusion models.Mentioned twice above
• IIP-AdapterIP-Adapter (short for Image Prompt Adapter) is a lightweight neural network module that adds image-prompt conditioning to a pretrained text-to-image diffusion model, allowing a reference image to guide…Mentioned twice above
• Qwen-Image-2.1Qwen-Image-2.1 is an open-weight image generation and image editing model released by Alibaba's Qwen team on September 20, 2026.Also in Image Generation
• Creating Game Assets with AICreating game assets with AI is a production workflow in which generated images, meshes, sound or text are reviewed, edited and imported into a game.Also in Generative AI
• Adobe FireflyAdobe Firefly is a family of creative generative AI models developed by Adobe and integrated across its product ecosystem.Also in Image Generation

More to read

• Recraft AI
• Higgsfield AI Influencer
• Best AI Image Generators
• PixAI
• Stable Diffusion
• DALL-E

Topics:AI Tools & ProductsGenerative AIImage Generation

References

1. ^1 ^2Sakala, Ondřej. StableGen repository and README. Project description, authorship and software license.
2. ^1 ^2StableGen documentation. Blending and projection. Updated October 28, 2026.
3. ^1 ^2 ^3StableGen documentation. Architectures and presets. Updated October 28, 2026.
4. ^1 ^2StableGen documentation. Generation modes. Updated October 28, 2026.
5. ^StableGen documentation. Masked Edit. Updated October 28, 2026.
6. ^StableGen. Version 0.4.0: New Architectures, Masked Edit, Automatic Rigging & New Panel. October 28, 2026.
7. ^1 ^2StableGen documentation. TRELLIS.2 mesh generation. Updated October 28, 2026.
8. ^StableGen documentation. Rigging and animation. Updated October 28, 2026.
9. ^StableGen documentation. PBR maps. Updated October 28, 2026.
10. ^1 ^2StableGen documentation. Export. Updated October 28, 2026.
11. ^1 ^2StableGen documentation. Installation. Updated October 28, 2026.
12. ^1 ^2Sakala, Ondřej, posting as sakalond. I textured an entire street in Blender with Qwen Image 2.1, then made cyberpunk, anime and papercraft versions of it. Project-author demonstration and workflow description.
13. ^Qwen. Qwen Image 2.1 Research License Agreement. September 20, 2026.
14. ^Zhang, Lvmin; Rao, Anyi; Agrawala, Maneesh. Adding Conditional Control to Text-to-Image Diffusion Models. 2023.
15. ^Ye, Hu; Zhang, Jun; Liu, Sibo; Han, Xiao; Yang, Wei. IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. 2023.
16. ^Richardson, Elad; Metzer, Gal; Alaluf, Yuval; Giryes, Raja; Cohen-Or, Daniel. TEXTure: Text-Guided Texturing of 3D Shapes. 2023.

Show all 16 references