StableGen puts four models on a ComfyUI server you already run

The addon is a front end, and the models run elsewhere

StableGen does not contain a generative model. A single image or a text prompt goes to TRELLIS.2, Microsoft's 4B-parameter image to 3D system, and the resulting mesh is then textured by a diffusion model that lives somewhere else entirely. The TRELLIS.2 side runs through ComfyUI-TRELLIS2, a node pack kept in its own repository under PozzettiAndrea, and the page points at an installer.py for fetching it. Texturing, ControlNet, and IPAdapter all run through the user's existing ComfyUI installation, so the addon needs a reachable ComfyUI server before any of its features do anything.

The page's own hedges are informative. Blender is described as mostly responsive, because the heavy computation is offloaded rather than kept in the addon, and memory pressure is handled with disk offloading plus a configurable attention backend. FLUX.1-dev is the one backend called experimental, while SDXL and Qwen Image Edit are named without qualification. Since the addon leans on the server for the checkpoints and custom LoRAs it uses, the practical ceiling on output quality is whatever that server has installed, not whatever ships in this repository.

ControlNet units carry no architecture map, IPAdapter runs with no image at all

Local Edit Mode is the one feature with an explicit architecture claim. It points cameras at specific areas, blends the new texture over the original using angle-based and vignette-based feathering, and states that it works with SDXL, Flux, and Qwen Image Edit. The angle ramp and the silhouette edge feathering have separate controls, which is what makes the blend boundary tunable instead of automatic.

ControlNet gets finer controls and no architecture claim. Depth, Canny, and Normal units can run simultaneously so that generated textures respect the mesh geometry, and each unit carries its own strength plus start and end steps, with custom model mapping supported. Nowhere on the page are those units tied to a particular backend, so a reader has to work out which depth and normal models are usable with the server already in front of them.

IPAdapter is the odd one out. Its stated purpose is guiding style, mood, and content from external reference images, and it exposes strength, weight type, and active steps. The page also says it can be employed with no reference image at all, in which case it serves enhanced consistency across multi-view generation rather than style guidance. One control, two different jobs, and a single strength slider ends up meaning two different things depending on whether an image is attached.

Two of the four resolution modes are named Cascade and left undefined

Image to 3D runs at four listed resolutions: 512, 1024, 1024 Cascade, and 1536 Cascade, with 1024 Cascade marked recommended and 1536 Cascade presented as the option for maximum geometric detail. Nothing on the page says what a cascade is, so the recommended mode and the highest detail mode both carry a term that is never defined.

Two surrounding mechanisms are explained better. A preview gallery generates several candidate images from different seeds, so a choice can be made before committing to 3D generation, which puts the cheap step in front of the expensive one. The mesh side has its own repair path: automatic recovery from mesh corruption, configurable decimation and remeshing, import scaling, and a studio lighting setup. Those four are the difference between a failed run and a salvageable one, since generated geometry can come back broken or at the wrong scale.

TRELLIS.2's native PBR textures can be used as they arrive, or the generated mesh can be sent through SDXL, FLUX.1-dev, or Qwen Image Edit instead, which is the route the page prefers when diffusion textures matter more than the native ones.

The multi-color export names a format with nowhere to put the colors

The 3D printing suite exports textured meshes into a multi-color format, written as 3MF/STL, aimed at OrcaSlicer, Bambu Studio, and PrusaSlicer. 3MF carries a material palette; STL does not, so the second extension in that pair has nowhere to hold the color assignment the suite spends most of its work computing.

That computation is the interesting part. A physical color mixing solver evaluates filament mixtures using Kubelka-Munk theory, tracking absorption and scattering coefficients, so a requested color becomes a set of filament demands rather than one hex value. Two heuristics sit on top of it. A one-sided saturation-difference penalty stops pigments bleeding as dither speckles into highly saturated regions, and vectorized pruning of filament demands below an 8% threshold keeps FDM layers free of banding. The 8% figure is stated as a fixed property, with no preference exposed for changing it.

The preview path projects and dithers color layers onto a flat-shaded preview mesh named SG_SlicedPreview inside the Blender viewport. The entire tab sits behind an Addon Preferences switch, Enable 3D Print Exporter Tab, so a headline feature stays invisible until that preference is turned on.

Folder batching hides the scene and keeps one log file name

Pointing StableGen at a directory of reference images turns the addon into an unattended batch that generates and textures multiple assets in one run. Two conveniences come with it. Baking runs automatically after each model finishes, with resolution, ORM packing, and normal conventions applied as part of the pass. And a batch log records four durations, Shape, Texture, Bake, and Total, plus a success or error status per item, written to stablegen_batch_log.json.

Two gaps are visible in that description. The log has a fixed name with no per-run suffix and no directory given, so a second batch lands on top of the first batch's timings, which is unfortunate for a feature whose whole purpose is running unattended. And scene isolation hides every pre-existing asset during each generation stage, to keep texture bleed and projection contamination out of the results. The point of hiding assets is that they are visible again afterwards, and the page does not say what happens to visibility when a run stops partway through an error.

Sequential spends time on consistency, Grid spends it on speed

StableGen applies textures to every mesh object in a scene at once, from the camera viewpoints the user defines, or to only a selected set of objects. Two modes decide how that work is scheduled. Sequential mode generates viewpoint by viewpoint on each mesh, using inpainting and visibility masks to keep separate views consistent across complicated surfaces. Grid mode processes multiple viewpoints for all meshes simultaneously, which is the faster route and yields quicker previews, with an optional refinement pass available afterwards.

Weighted blending handles the joins between views in both cases. The trade is stated plainly: one mesh at a time is the consistent path, everything at once is the fast path, and a refinement pass can recover some of the difference after the fact.

Scene-wide operation changes what the choice means. A single asset job only weighs per-view consistency, while an environment or a set of related assets also weighs whether the meshes agree with each other. That second case is the one the page puts forward for concept art, look development, and batch texturing an asset library.

Seven camera strategies and an aspect ratio read off the silhouette

Camera placement is where the tuning happens, and the page names seven strategies: Orbit Ring, Fan Arc, Hemisphere, PCA-Axis, Normal-Weighted K-means, Greedy Occlusion Coverage, and Interactive Visibility-Weighted. The list is descriptive rather than prescriptive, with no guidance on which one suits a given shape.

Two details make this more than a canned set. Each camera gets its own aspect ratio, computed from that mesh's silhouette, so a view is framed to its subject instead of being letterboxed into a fixed shape. And the camera count is unlimited, with the page stating outright that the earlier eight-camera limit is gone. Order matters as well in Sequential mode, because results blend in sequence: a drag-and-drop reorder list with six preset strategies controls the processing order, and individual cameras can be cloned, mirrored, and given floating viewport prompt labels.

The consequence is that camera setup is a design step rather than a checkbox. More cameras means more consistency work in Sequential mode and more preview time in Grid mode, and the page offers no automatic stopping point.

The contents list promises ten sections and the text stops at Advanced Inpainting

The page opens with a linked contents list covering Key Features, How It Works, System Requirements, Installation, Quick Start Guide, Usage and Parameters Overview, Output Directory Structure, Troubleshooting, Contributing, License, Acknowledgements, planned features, and Contact. The text that follows is the feature list, and its final heading is Advanced Inpainting, which introduces nothing further. What is visible is an inventory of what the addon can do, with no accompanying statement of what it takes to run: no memory figure, no Blender version, no download step for the addon itself.

Version facts have to be recovered elsewhere. The v0.2.0 release title is the only place Blender 5.1 Support appears. The three tagged releases read like a changelog of the feature list itself: v0.2.0 for the camera overhaul and Local Edit Mode, v0.3.0 for 3D asset generation, PBR materials, and the scene queue, v0.3.1 for the TRELLIS.2 node upgrade, multi-color 3D printing, and folder batching. The newest tag is v0.3.1 dated 2026-06-12, while the repository's last push is 2026-09-29, so the tip of the default branch sits roughly three and a half months ahead of the newest packaged build.

Editorial conclusion

StableGen suits someone who already runs ComfyUI and wants image to 3D plus diffusion texturing inside Blender without wiring nodes by hand, and its camera placement work is the part most likely to repay the setup effort. It is not self-contained: TRELLIS.2 arrives through a separate node pack, the texture models live on your server, and the page states neither what the machine needs nor how to get the addon itself, so check the Blender version and your memory budget against your own setup first. If the asset pipeline has to be reproducible from a tag, pin to v0.3.1, because the default branch is months ahead of the newest packaged build.

Frequently asked questions

Official sources

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