StableGen 0.4 for Blender: AI Texturing, Auto-Rigging and Mixamo Animation With ComfyUI (Install Guide)

Texturing a 3D model by hand takes hours, and rigging a character takes even longer. StableGen 0.4.0, released on October 28, 2026, does both from inside Blender. It is a free add-on that textures any mesh or whole scene with diffusion models, generates new meshes from a prompt or image with TRELLIS.2, and now auto-rigs characters with SkinTokens so you can drop Mixamo animations straight onto them. All the AI work runs on a ComfyUI server, on the same PC or another machine on your network. This guide covers what is new in 0.4.0, the exact install steps, which installer packages to pick, where every custom node and model goes, and how much VRAM you need.

What Is StableGen?

StableGen is a Blender add-on that connects Blender to ComfyUI. Blender handles geometry, cameras and materials. ComfyUI handles all the AI models. It started as a bachelor’s thesis by Ondřej Sakala at the Czech Technical University in Prague and has grown into one of the most complete AI texturing tools for Blender.

It works by projection. You place cameras around a model. For each camera, StableGen renders a depth guide, asks ComfyUI to generate a view that follows that geometry, and projects the result back onto every surface that camera sees. Each new view is generated with the earlier views as context, so the texture stays consistent all the way around. The views are then blended by how directly each camera sees each surface.

On top of texturing, StableGen can:

• Generate meshes from a prompt or an image with Microsoft’s TRELLIS.2.
• Estimate PBR maps (albedo, roughness, metallic, normal) with Marigold and StableDelight.
• Rig characters with SkinTokens and retarget animations onto them.
• Bake and export to glTF, FBX, 3MF and video.

What Is New in StableGen 0.4.0

Auto-rigging and animation

The headline feature. Auto-rig selected character adds a skeleton and skin weights using SkinTokens on the ComfyUI server. It can run as part of a TRELLIS.2 generation or on any mesh you select. Motion from FBX or BVH files, including Mixamo clips, can then be retargeted onto the new rig. A new Texturing Frame setting fixes which pose is used when texturing a rigged character, so arms and legs do not smear.

Masked Edit

Paint over part of a texture and describe the change. Masked Edit regenerates only that region, in every camera view that sees it, and leaves the rest alone. You get a preview to apply or discard, and it updates the PBR maps too, including emission.

Better PBR and projection

• Marigold V2 replaces v1 for albedo and normal maps, at up to 2048 px.
• Depth/Part Projection Guard (DPPG) reduces colour bleeding at depth and material edges, with Depth-Aware Margins, Coverage Rescue and Surface Blend feathering.
• The projection shader was rebuilt to stay within Cycles’ and EEVEE’s limits.

New models and a new panel

Qwen Image 2.1 joins with presets (Precise, Precise+, Safe, Context + Reference, Local Edit, Refine) at 1 MP by default and up to about 4 MP. Qwen Image 2.1 Turbo (8 steps), MOSS Image and FIBO 2.0 Edit Turbo (4 steps) arrive as experimental options. FLUX.1 and FLUX.2 Klein are now deprecated. The interface moves to a tabbed panel (Texture, Cameras, Blending, Tools, Mesh); the old sidebar is still available under Use Legacy Panel in preferences.

Faster

On a six-view SDXL scene, the non-AI work dropped from 27.1 s to 13.6 s, and the longest UI freeze from 4.4 s to 0.8 s.

Requirements and VRAM

Two practical notes. On Blender 4.2 to 4.5, projection uses Open Shading Language, so StableGen temporarily switches Cycles to CPU while it builds projection materials. Blender 5.1 and later use native Raycast nodes and skip that. Also, TRELLIS.2 needs compiled CUDA extensions, and the installer warns there are no pre-built wheels for Python 3.13, so run ComfyUI with Python 3.12 if you want mesh generation.

On a 16 GB card like an RTX 5080 you can use every architecture, though Qwen Image 2.1 at its higher megapixel settings will be slower per view. Our ComfyUI VRAM guide explains the trade-offs.

Step-by-Step Installation

1. Install ComfyUI

Install ComfyUI and make sure it starts and generates an image. If ComfyUI will run on another computer, do steps 1 and 2 on that computer.

2. Install the custom nodes and models with the installer

git clone https://github.com/sakalond/StableGen
cd StableGen
python installer.py /path/to/ComfyUI

Replace /path/to/ComfyUI with your ComfyUI folder (the one containing models and custom_nodes). For the Windows portable build, either the outer ComfyUI_windows_portable folder or the ComfyUI folder inside it works. If you leave the path out, the script asks for it.

The installer finds the Python that ComfyUI uses (including the portable build’s python_embeded and venv or .venv folders), shows a menu of packages, lists every component with its download size and your free disk space, and asks before downloading. Interrupted downloads resume, and files already present are skipped, so you can rerun it any time. It installs custom nodes into ComfyUI/custom_nodes/ and models into the matching ComfyUI/models/ subfolders. Restart ComfyUI when it finishes.

3. Install the add-on in Blender

Download the add-on ZIP for your system (windows-x64, linux-x64 or macos-arm64, about 65 to 90 MB) from the StableGen releases page. In Blender, open Edit, Preferences, Get Extensions, choose Install from Disk from the menu in the top-right corner, and select the ZIP. Blender rejects a ZIP built for a different platform, so pick the right one.

4. Configure the preferences

In the add-on preferences, set an Output Directory and the Server Address (127.0.0.1:8188 by default, or your other machine’s IP and port). Then go to System, Network and enable Allow Online Access, otherwise Blender blocks the connection to ComfyUI.

Upgrading from an older version

Existing users must rerun the installer. Add --update-nodes to move already-installed custom nodes to the commits StableGen 0.4.0 was tested with. After any ComfyUI update, run python installer.py /path/to/ComfyUI --repair, because TRELLIS.2 runs in an isolated environment that can fall behind.

Which Installer Packages to Choose

For unattended installs, the installer takes command-line options:

python installer.py /path/to/ComfyUI --packages 13,14 --yes --verify

• --packages 3,8 installs those packages without the menu.
• --yes accepts every prompt, including licence notices.
• --list shows what each package installs; --dry-run changes nothing.
• --verify checks that the model files exist and ComfyUI registers every needed node; --verify-hashes also checks SHA-256.
• --python points to ComfyUI’s Python if the installer cannot find it.

For gated Hugging Face models, set the HF_TOKEN environment variable to your access token before running it.

Manual Install: Custom Nodes and Model Paths

The installer is the easiest route, but if you prefer to do it by hand, here is what it sets up. Every git command runs inside ComfyUI/custom_nodes/.

Custom nodes

TRELLIS.2 must be pinned to the tested commit and needs a specific comfy-env version:

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-TRELLIS2.git
cd ComfyUI-TRELLIS2
git checkout 9b878516f2dc2fd873f4f6cceadba403dd12d83e
pip install -r requirements.txt
pip install --force-reinstall --no-deps comfy-env==0.4.1

Afterwards, run installer package 8 anyway: it keeps your clone but applies compatibility patches. TRELLIS.2 models (about 15.4 GB) download on first use.

Marigold:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-Marigold.git
cd ComfyUI-Marigold
pip install -r requirements.txt
pip install "diffusers>=0.28"

On Windows, also run pip install triton-windows.

SkinTokens builds its own environment so ComfyUI’s Python is not touched:

cd ComfyUI/custom_nodes
git clone https://github.com/sakalond/ComfyUI-StableGen-SkinTokens.git
cd ComfyUI-StableGen-SkinTokens
python install_skintokens.py --target managed

Use ComfyUI’s own Python for that last command (on Windows portable, ..\..\..\python_embeded\python.exe). It creates vendor/skintokens_env with its own PyTorch and downloads about 1.6 GB of checkpoints. python doctor_skintokens.py reports its status. It needs an NVIDIA GPU.

Core SDXL models (exact paths)

The renames matter: StableGen’s presets look for those exact file names.

Qwen Image 2.1 models (exact paths)

• ComfyUI/models/diffusion_models/qwen_image_2.1_int8_convrot.safetensors (about 7.3 GB)
• ComfyUI/models/text_encoders/qwen3vl_8b_w4a8.safetensors (about 6.3 GB)
• ComfyUI/models/vae/qwen_image_2.1_vae_bf16.safetensors (about 0.7 GB)
• Optional Turbo: ComfyUI/models/diffusion_models/qwen_image_2.1_turbo_int8_convrot.safetensors (about 6.9 GB)

All are in the Comfy-Org/Qwen-Image-2.1 repo, and they use ComfyUI core nodes (ComfyUI from September 2026 or newer). See our Qwen Image 2.1 Turbo guide for more on this model.

Marigold V2 PBR models

From Comfy-Org/marigold-v2-0, keeping the folder layout under ComfyUI/models/: diffusion_models/qwen_image_edit_2509_int8_convrot.safetensors, loras/marigold_v2_albedo.safetensors, loras/marigold_v2_normals.safetensors, the two matching files in vae/ and the two conditioning files in embeddings/. About 23.3 GB in total.

How to Texture a Model

1. In Blender, press N in the 3D Viewport and open the StableGen tab.
2. Select your mesh and click Add cameras. Six views around the object is a good start.
3. On the Texture page, choose an Architecture (SDXL, Qwen Image Edit or Qwen Image 2.1), a Preset and a Checkpoint.
4. Describe the surface in Texture prompt, for example “weathered bronze statue, green patina, dark crevices”.
5. Click Generate textures. Switch the viewport to Rendered shading to see the result.
6. Use Bake to turn the projection into normal UV textures, then Export.

You can also do every step by hand: reorder cameras, give individual views their own prompts, change how views blend and reproject without regenerating, extract PBR maps, and fix any patch with Masked Edit.

Prompt to Rigged Character

With packages 8, 13 and 14 installed, StableGen can go from a text prompt to a textured, rigged character in one run:

1. Set Architecture to TRELLIS.2 and choose a preset tagged Mesh + texture.
2. On the Mesh page, set Generate mesh from to Prompt, describe the character, and enable Preview gallery.
3. Under Rigging & Animation, enable Auto-rig.
4. Click Generate asset and pick your favourite preview image.

TRELLIS.2 builds the mesh, SkinTokens adds the skeleton and skin weights, and generated views are projected into the texture. You can also start from an image instead of a prompt, skip the rig, or keep TRELLIS.2’s own texture. Once rigged, import an FBX or BVH animation, such as a Mixamo walk cycle, and retarget it onto the character.

If you only need a quick 3D preview of an object rather than a game-ready asset, a splat model is faster. See our Idol image-to-3D splat guide.

Which Texture Model to Use

The VRAM column for the non-SDXL rows follows StableGen’s own “16 GB or more for Qwen Image” guidance and the file sizes; it is not a measured figure for each model. If you sell assets, stick to SDXL, Qwen Image Edit 2509 or MOSS Image. Experimental architectures are hidden until you turn on Enable Experimental Architectures. Learn more about two of them in our MOSS-Image and FIBO 2 guides.

Exporting to Unreal, Unity and Godot

Bake the result, PBR maps included, into UV textures and export it with its rig and animations:

• glTF/GLB for Godot and the web.
• FBX for Unreal Engine and Unity.
• 3MF with multiple colours for 3D printing.
• Orbit GIFs and MP4s for previews.

The developer’s demo is a whole village in Unreal Engine 5.7 built only from StableGen exports, with five rigged, walking characters. For big jobs, the Scene Queue runs batches unattended, and batch texturing from CSV manifests works for DAE and GLB files.

Troubleshooting

Blender cannot connect to ComfyUI

Check that ComfyUI is running, that the Server Address matches (default 127.0.0.1:8188), and that Allow Online Access is enabled under Preferences, System, Network. For a remote server, start ComfyUI with --listen so it accepts network connections.

The add-on ZIP will not install

You downloaded the ZIP for a different platform, or you are on Blender 5.0, which is not supported. Use 4.2 to 4.5 or 5.1 and later.

TRELLIS.2 breaks after a ComfyUI update

Run python installer.py /path/to/ComfyUI --repair and restart ComfyUI. Avoid updating the TRELLIS.2 node through ComfyUI Manager, because the installer pins it to a tested commit.

Missing preset models

Check the renamed files in the table above. A depth ControlNet left as diffusion_pytorch_model.safetensors will not be found. Run the installer with --verify to list what is missing.

Masked Edit is greyed out

It does not work with the deprecated FLUX architectures. Switch to SDXL or Qwen.

FAQ

Is StableGen free?

Yes. The add-on is free and open source under GPL-3.0. Some models it can use have non-commercial licences.

Does StableGen work without an NVIDIA GPU?

The add-on runs on Windows, Linux and Apple Silicon Macs, but an NVIDIA GPU is recommended for the ComfyUI server, and SkinTokens rigging needs NVIDIA. You can also run ComfyUI on a separate PC with an NVIDIA card.

How much VRAM does StableGen need?

About 8 GB for SDXL texturing and 16 GB or more for Qwen Image models.

Can I use StableGen assets commercially?

The add-on itself allows it. The models decide: SDXL, Qwen Image Edit 2509 and MOSS Image are fine, while Qwen Image 2.1, its Turbo version and FIBO 2.0 are non-commercial.

Can StableGen rig any character?

Auto-rig works on TRELLIS.2 characters and on any selected humanoid mesh. Results are best on clean, single-piece character meshes in a neutral pose.

Which Blender version should I use?

Blender 5.1 or later is best, because it uses native Raycast nodes and does not need to switch Cycles to CPU during projection.

Final Thoughts

StableGen 0.4.0 turns Blender into a near-complete AI asset pipeline: mesh, texture, PBR maps, rig and animation, all driven by local models through ComfyUI. The installer makes setup far easier than wiring up six custom node packs by hand, and the new auto-rigging closes the last big gap between a generated mesh and something you can drop into a game engine. Start with package 3 for SDXL texturing on an 8 GB card, or packages 8, 13 and 14 on a 16 GB card if you want the full prompt-to-character workflow. Just keep an eye on the licences if you plan to sell what you make.

More guides:

• Idol in ComfyUI: Turn One Image Into a 3D Gaussian Splat
• ComfyUI Pixal3D Workflow: Image to 3D with TRELLIS.2
• Best Qwen Image 2.1 Turbo Workflow for 16 GB

Sources

• GitHub: sakalond/StableGen (README, read October 31, 2026)
• StableGen wiki: Installation and Manual installation
• ComfyUI Wiki: StableGen 0.4.0 Textures and Rigs 3D Assets in Blender via ComfyUI
• GitHub: VAST-AI-Research/SkinTokens