StableGen: AI Texturing and Image-to-3D Inside Blender

The gap StableGen fills in a Blender texturing workflow

Texturing a model usually means leaving Blender, generating images somewhere else, and projecting them back by hand. StableGen moves both halves of that loop into the addon. The README describes it as an open-source Blender addon that brings generative AI into the 3D workflow: generate textured meshes from a single image or text prompt via TRELLIS.2, then texture and refine them, or any existing model, using SDXL, FLUX.1-dev or Qwen Image Edit through a ComfyUI backend. The intended user is a 3D artist or technical artist who already has a ComfyUI instance and wants the diffusion stack reachable from the viewport. It is not for someone who wants a hosted service. StableGen drives models you host yourself, and the README points at ComfyUI-TRELLIS2 as a separate component installed through installer.py.

How StableGen drives ComfyUI, cameras and ControlNet

The architecture is a client-server split. Blender holds the scene, the cameras and the mesh; ComfyUI holds the diffusion models. StableGen builds a camera set, renders or projects from those viewpoints, sends the work to ComfyUI, and blends the results back onto the mesh.

Camera placement is the part that decides output quality, and the README lists seven strategies: Orbit Ring, Fan Arc, Hemisphere, PCA-Axis, Normal-Weighted K-means, Greedy Occlusion Coverage, and Interactive Visibility-Weighted placement. Each camera gets its own aspect ratio computed from the mesh silhouette, and the README states there is no 8-camera limit. Generation order is reorderable by drag and drop, with six preset strategies controlling the processing order in Sequential mode.

Multi-view consistency has two modes. Sequential mode generates texture viewpoint by viewpoint on each mesh, using inpainting and visibility masks, and is aimed at complex surfaces. Grid mode processes multiple viewpoints for all meshes at once for faster previews, with an optional refinement pass. Both use weighted blending between views.

Geometric control comes from ControlNet units (Depth, Canny, Normal) that can run simultaneously, each with its own strength and start/end steps, plus custom model mapping. Style guidance comes from IPAdapter, which can also be used without a reference image to improve consistency in multi-view modes. Texturing is not limited to one object: the README states StableGen applies textures to all mesh objects in the scene simultaneously from the defined camera viewpoints, or to selected objects only.

Installing StableGen and texturing your first model

The repository's top level contains installer.py, a stablegen/ package, docs/, LICENSE and README.md. The README's installation section is the place to start, and it points at ComfyUI as a prerequisite backend. The TRELLIS.2 path additionally needs ComfyUI-TRELLIS2, which the README says is installable via installer.py.

System requirements are stated as Blender 4.2+ or 5.1+, per the version badge. Once the addon is enabled, the workflow is: place cameras, choose an architecture, generate. The README describes the backend as flexible rather than fixed, so check the addon preferences for the connection fields rather than assuming a default host.

For a first real run, texture an existing model rather than generating one. The Quick Start Guide separates "Texturing an Existing Model" from "Generating a 3D Model with TRELLIS.2". Texturing an existing model avoids the TRELLIS.2 download and the 4B-parameter model entirely, which makes it the cheaper way to find out whether your ComfyUI connection works. After a pass completes, the README documents an output directory structure; that is where the baked textures land.

TRELLIS.2 generation, resolution modes and the VRAM bill

The generation half uses Microsoft's TRELLIS.2, described as a 4B-parameter model. The README lists four resolution modes: 512, 1024, 1024 Cascade (recommended), and 1536 Cascade for maximum geometric detail. The recommended mode is not the cheapest one, which tells you where the quality threshold sits.

Two features soften the cost. StableGen supports disk offloading and a configurable attention backend, both described as VRAM-conscious measures. There is also a Preview Gallery: generate multiple candidate images with different seeds and pick the best before committing to 3D generation. That matters because 3D generation is the expensive step, and choosing the seed after the fact wastes it.

Mesh handling is defensive rather than clean. The README lists auto-recovery from mesh corruption, configurable decimation and remeshing, import scaling, and a studio lighting setup. Auto-recovery from corruption is an admission that generated meshes are not always well formed. Plan for decimation and remeshing as normal steps, not as rescue operations.

Texture sourcing is a real fork in the pipeline. You can keep TRELLIS.2's native PBR textures, or automatically re-texture the generated mesh with SDXL, FLUX.1-dev or Qwen Image Edit for what the README calls higher-quality diffusion textures. The second path costs another diffusion pass per asset.

Batch generation and the 3D printing exporter

Folder batching points StableGen at a directory of reference images and generates and textures multiple assets in one unattended run. Three details make it usable rather than a loop wrapper. Auto-baking runs texture baking (resolution, ORM packing, normal conventions) after each model is generated. Scene isolation hides pre-existing assets during each generation stage to prevent texture bleed and projection contamination. And a batch performance log writes execution durations for Shape, Texture, Bake and Total, plus success or error status, to stablegen_batch_log.json. That log is the honest part of the feature: unattended runs fail, and the file tells you which asset failed and where the time went.

The 3D printing suite is the most unusual thing in the repository. It exports textured meshes into a multi-color format (3MF/STL) compatible with OrcaSlicer, Bambu Studio and PrusaSlicer. The color conversion is not naive: the README describes a physical color mixing solver that optimizes filament mixtures using Kubelka-Munk theory, evaluating absorption K and scattering S coefficients. A one-sided saturation-difference penalty prevents pigments from bleeding as dither speckles into highly saturated regions. A vectorized 8% threshold pruning step filters out minor filament demands to keep layers clean. A viewport preview projects and dithers color layers onto a flat-shaded preview mesh named SG_SlicedPreview inside Blender. This tab is off by default: it requires enabling "Enable 3D Print Exporter Tab" in Addon Preferences.

Limitations and where StableGen is the wrong tool

StableGen is not self-contained. It is a Blender client for a ComfyUI server, and the README treats ComfyUI as an external dependency you supply. If you do not already run ComfyUI, the addon is the smaller half of the setup.

The README does not document rollback. There is no described mechanism for reverting a texture pass that went wrong, so the practical safety net is Blender's own undo and file versioning, not the addon. Treat a long sequential pass on a complex scene as a commitment.

Hardware is the second boundary. TRELLIS.2 is a 4B-parameter model, and the recommended generation mode is 1024 Cascade, not 512. Disk offloading and attention backend selection exist precisely because the default path is heavy. A machine that runs Blender comfortably may still not run this pipeline.

Finally, scope. The README's own framing is concept art, look development and batch-texturing asset libraries. If you need deterministic, hand-authored UV work with repeatable results, a diffusion pipeline that samples from seeds is the wrong instrument, no matter how good the blending is. The output directory structure and the batch log are the only audit trail described; there is no versioned, reproducible texture build.

StableGen compared with StableProjectorz and PixelModeller AI

The closest alternatives people search for alongside StableGen are StableProjectorz and PixelModeller AI. The difference is where the intelligence lives.

StableProjectorz is a projection-oriented texturing tool: you paint or project images onto a model and the value is in the projection and blending controls. StableGen's Local Edit Mode is the comparable feature, and the README describes it as pointing cameras at specific areas to modify, with the new texture blended over the original using angle-based and vignette-based feathering, separate angle ramp and silhouette edge feathering controls, and support for all architectures (SDXL, Flux, Qwen Image Edit). So StableGen covers that ground, but it is one mode among several rather than the whole product.

What StableGen adds on top is generation and scene-scale work: TRELLIS.2 image-to-3D, seven camera placement strategies, scene-wide multi-mesh texturing, folder batching with a JSON performance log, and the 3D printing exporter with the Kubelka-Munk filament solver. PixelModeller AI is a generative texturing tool for Blender; the README does not compare against it, and nothing in the repository describes its internals, so the honest statement is that both sit in the generative-texturing slot and StableGen is the one whose backend is explicitly ComfyUI and whose scope extends to mesh generation and print export.

The practical distinction is dependency shape. StableGen assumes you run ComfyUI and are willing to manage model files, attention backends and VRAM. A tool that hides the backend trades that control for a shorter setup.

Licence, maintenance and upgrade cost

StableGen is GPL-3.0. The addon is distributed under that licence, and the repository carries a LICENSE file at the top level. If you plan to redistribute a modified version or bundle it into a product, read the GPL-3.0 terms yourself; nothing here is legal advice. Note also that the addon orchestrates third-party components with their own terms: TRELLIS.2, ComfyUI, ComfyUI-TRELLIS2, SDXL, FLUX.1-dev and Qwen Image Edit are separate projects, and the licence of the addon says nothing about the licence of the models you load into ComfyUI.

On maintenance, the last push to the default branch was on 2026-07-07. The repository is not archived. The release cadence visible in the repository is steady rather than fast: v0.2.0 on 2026-02-15, v0.3.0 on 2026-03-05, and v0.3.1 on 2026-06-12. The release titles show what upgrades cost you. v0.2.0 was a camera overhaul that added local edit mode and Blender 5.1 support. v0.3.0 added 3D asset generation, PBR materials and a scene queue. v0.3.1 upgraded the TRELLIS.2 node, added multi-color 3D printing, folder batching and modularization. Each of those touches the parts of the addon users build habits around: camera placement, the node graph, the export path.

The upgrade cost is therefore not the addon download, which is small. It is the backend. A TRELLIS.2 node upgrade in v0.3.1 implies a matching ComfyUI-TRELLIS2 version, and the README already routes that install through installer.py. Pin your ComfyUI and ComfyUI-TRELLIS2 versions before upgrading StableGen, and re-run a small batch after each upgrade so stablegen_batch_log.json can confirm the pipeline still completes.

Editorial conclusion

Adopt StableGen if you already run ComfyUI and want multi-mesh texturing or image-to-3D generation inside the viewport, and if your Blender is 4.2+ or 5.1+. Skip it if you want a one-click installer or have no GPU budget for TRELLIS.2 and a diffusion model. Before committing, confirm three things: that your ComfyUI install satisfies the addon's requirements, that your GPU memory survives the 1024 Cascade mode, and that the folder-batching log stablegen_batch_log.json reports success on a small test directory before you point it at an asset library.

Frequently asked questions

Official sources

1. Issues
2. License: GPL-3.0
3. README
4. Releases
5. sakalond/StableGen on GitHub