StableGen 0.4.0 Textures and Rigs 3D Assets in Blender via ComfyUI

StableGen has been around since 2025, and it has always been a bridge rather than a model: Blender owns the geometry, the cameras and the materials, ComfyUI owns the inference, and the add-on moves images and depth between them. The 0.4.0 release is the largest change to that bridge so far. The sidebar becomes a full panel, texture edits can be masked down to a painted region, characters can be rigged during mesh generation, and the projection shader was rebuilt around a depth and part boundary guard.

A new panel

The sidebar is replaced by a panel with page tabs (Texture, Cameras, Blending, Tools, and Mesh for TRELLIS.2), a toolbar for the camera tools, a search field on Ctrl K that filters every setting, and a footer with progress plus Reproject, Bake and Export. Tags beside each architecture mark whether it textures a scene or generates an asset, and flag the experimental and deprecated ones. Preset tags show the workflow behind each preset, and the list only shows the presets that belong to the selected architecture. The old sidebar is still available through Use Legacy Panel in the preferences.

Qwen Image 2.1, plus three experimental architectures

Qwen Image 2.1 is the new default-grade architecture, with six presets: Precise, Precise+, Safe, Context + Reference, Local Edit and Refine. It runs at its native 2K resolution, about 4 MP per view, if you raise Target megapixels under Sampling & resolution; the default stays at 1 MP because higher budgets take longer per view. It generates RGBA views, so transparent pixels stay out of the projection, it carries thin detail such as trim lines on dark fabric from one view to the next more cleanly, and its driving prompts are editable templates. Requests can be exported as ComfyUI workflows.

Three more architectures are tagged Experimental, shown by default and hidden by Enable Experimental Architectures:

• Qwen Image 2.1 Turbo runs the same workflows with the few-step Turbo model at 8 steps, about three times faster than Qwen 2.1 Precise on the same scene.
• MOSS Image runs the SDXL workflows, including Generate Sequentially, Local Edit and UV Inpaint, but follows the geometry with its own depth control adapter instead of ControlNet.
• FIBO 2.0 Edit runs the Qwen workflows with FIBO 2.0 Edit Turbo at 4 steps and 1 MP per view.

FLUX.1 and FLUX.2 Klein are now deprecated: they and their presets are hidden unless Enable Deprecated Architectures is turned on. Files that use them still open.

Masked Edit

Painting a region on the model and describing a change makes StableGen regenerate only that region, through every camera that sees it, leaving the rest of the texture as it was. It works with every texturing architecture except the deprecated FLUX ones, holds the edit as a preview until you apply or discard it, and on PBR materials updates the maps of the edited region too, including new emission for glowing details.

The original texture	Masked Edit: "black neon tech jacket"

Rigging with SkinTokens

Auto-rig selected character adds a skeleton and skin weights with SkinTokens on the ComfyUI server, either as a step of TRELLIS.2 generation or on any mesh you select. Animation from FBX or BVH files, Mixamo clips included, can then be retargeted onto that rig, with a reference pose matched to the source, elbows and wrists that follow the source motion, and arms kept clear of the body.

Marigold V2 and PBR maps

Marigold V2 replaces v1 as the albedo and normal source, with sharper detail, better colour and support up to 2048 px. The albedo saturation correction now measures only the object instead of the whole frame, and it switches off under V2 because the V2 albedo already keeps the generated colours. There is a tile inference option for tiled extraction, and re-baking removes the maps of channels you switched off so they are not wired in again.

Projection: the Depth/Part Projection Guard

How generated views are projected and blended was rebuilt for cleaner seams, edges and occlusion. The new Depth/Part Projection Guard (DPPG) stops a view's foreground colours from spilling onto the surfaces behind them, such as an arm's skin onto the dress behind it, at depth and object or material boundaries. Depth-Aware Margins set separate foreground and background widths, Coverage Rescue relaxes the guard where only much weaker views are left, and the guard edge can be feathered with several methods including Surface Blend, rendered at a higher scale for smoother boundaries.

Guard off: every view leaves outlines on the wall	Guard on: the wall stays clean

The projection material is now split into labelled frames (surface geometry, image and UV, camera frustum, raycast visibility, facing angle, guard, per-view sections) with the main blending controls exposed as shader inputs, so what the material does can be followed and adjusted in the Shader Editor. It stays inside Cycles' shader stack and EEVEE's attribute limit, which stops scenes with many cameras from rendering black or magenta.

v0.3.1	v0.4.0

Faster, and the interface stays responsive

Guidance renders, visibility and projection builds now run without blocking Blender, and progress is reported for every ComfyUI stage. On the same six-view SDXL scene, everything StableGen does per run except model inference dropped from 27.1 s to 13.6 s, and the longest interface freeze from 4.4 s to 0.8 s. Visibility renders went from 11.0 s to 3.8 s, context renders from 9.8 s to 3.9 s, projection from 4.5 s to 2.9 s, while depth export rose from 0.6 s to 2.1 s because it now also renders the guard maps.

Two more additions help with rigged characters: a Texturing Frame setting fixes the frame used for posed geometry, visibility, projection and baking so animated characters texture in a consistent pose, and a Diagnostics page (with debug settings enabled) shows which view writes where, what each camera discards, and the context sent to the model, with replayable records.

Installer and documentation

The add-on now ships as one ZIP per platform, windows-x64, linux-x64 or macos-arm64, each about 65 to 90 MB instead of a single 230 MB file; Blender refuses a ZIP built for another platform. The installer has new packages for Marigold V2, Qwen Image 2.1, SkinTokens and the three experimental architectures, pins every custom node to a tested commit, and gains command-line options (--packages, --yes, --list, --dry-run, --verify, --verify-hashes, --update-nodes) for unattended runs. It checks disk space, Git and the ComfyUI version first, repairs the TRELLIS.2 environments after ComfyUI updates, and installs the Nunchaku wheel that matches ComfyUI's Python and torch. Existing installs need the installer run again; --update-nodes moves previously installed nodes to the pinned commits.

Batch texturing of DAE and GLB files can now be driven by a CSV manifest, FBX export puts the baked UV map first and drops unused material slots so engines such as Unreal map the baked textures correctly, and the 3D printing side adds a dithering method, a more reliable Make Solid, printable mesh checks and textured OBJ export.

The release also comes with the project's first full documentation: a 15-page wiki covering installation, architectures, cameras, blending, Masked Edit, PBR maps, rigging, export and troubleshooting.

Availability

StableGen is a Blender add-on, not a ComfyUI custom node: you install it in Blender, point it at a ComfyUI server, and use the installer to fetch the models and nodes it needs on that server. Code and releases are on GitHub (GPL-3.0), documentation on the project wiki, and questions go to the project's Discord.