3.56.0

Rigged character generator (#1121, #1124, #1125)

A human character generator built from two open data sets: the MHR body (Apache-2.0) with its own head, plus parts from ICT-FaceKit (MIT).

• A standard 52-joint skeleton with body and head in one mesh, eyeballs / teeth / mouth interior as separate parts, and 51 ARKit expression shapes as morph targets, fitted onto MHR's own face-expression vectors and exactly left/right symmetric.
• Open eyes by default (the thin layer between the lids is removed and the eyeballs and mouth interior are seated relative to eyelid and lip landmarks, so they follow every head shape). --closed-eyes keeps the sculpted lids.
• Surfaces: qtmesh character (CLI), MCP generate_character, and an Object-mode Inspector "Character Generator" with named body and head shape sliders, a live preview that edits the character in place (one undo step), and skin and eye colour. A new dark eye colour is available everywhere.
• Textured eyes (opt-in, local): QTMESH_CHARACTER_EYE_OBJ + QTMESH_CHARACTER_EYE_TEX swap the flat-colour eyeballs for a textured eye mesh; the eye colour picks the closest of six iris textures. No eye asset is bundled, because its licence is not cleared.
• Walk and run keep the shoulders up. Applying a generated walk or run used to drop the character's shoulders (MHR's spine is an S-curve and the clip's clavicle slope was copied onto the target). The spine pivots are now lined up and the collars retarget from the character's own bind pose; Spine2 rotation on a walk went from 40° to 8°.
• Fix: textured parts imported from MetaHuman DNA files rendered a single flat colour, because Ogre's software skinning replaces the position and normal buffers and the UVs were packed with the normal. They now get their own buffers.

Not usable yet: the two data bundles (mhr_character.qmcb, ict_character.qmcb) are not hosted yet, so the generator cannot download its data on first use and will report that it is unavailable until they are published. The feature and its tests are in this release; the data is not.

FLUX.3 Image Klein, with files you provide (#1123, #1127)

The text-to-image step of prompt-to-3D can now run FLUX.3 Image Klein (5.6B) as an alternative to FLUX.2-klein, which stays the default and is unchanged.

• It runs in a separate helper process (qtmesh_sdcli, stable-diffusion.cpp's own sd-cli, shipped next to the editor). The editor's built-in stable-diffusion.cpp shares llama.cpp's ggml and a version new enough for FLUX.3 aborts on it, so FLUX.3 gets its own ggml in its own process. Shipped in the macOS and Linux builds (verified end to end on macOS/Metal; the Linux build compiles and packages it but generation there has not been run yet); Windows is not wired up.
• There is no download for it in AI Model Settings, deliberately. FLUX.3 is under the FLUX Kommunity License, which is not permissive (free commercial use of outputs only under US$5M revenue; personal and non-sublicensable). A download entry would skip the licence acceptance Black Forest Labs' own repository enforces, so the editor does not host, bundle, mirror or offer to fetch these weights. If you have accepted the licence and obtained the files yourself, put them in <AppData>/ai_models/flux3_klein/ (a FLUX.3 diffusion GGUF whose name contains flux, the FLUX.2 VAE with vae in its name, and a Qwen3-VL-4B text-encoder GGUF with qwen in its name) and select FLUX.3-Image-Klein (qtmesh generate3d --image-model FLUX.3-Image-Klein).
• FLUX.3 supports reference-image editing and has its own step count ("FLUX.3 Steps", default 16) in Generation Settings. It has no depth ControlNet, so a mesh-aware texture request on FLUX.3 is refused with a clear message instead of silently dropping the depth map; use an SD 1.5 model for that.
• If you downloaded FLUX.3 through a development build, you can free the space by deleting <AppData>/ai_models/flux3_klein/.

Smaller improvements

• AI Model Settings now says why a model cannot be downloaded: an "Unavailable" label and a tooltip with the reason, instead of a button that silently does nothing.
• The image-model list refreshes itself when a download finishes, so a freshly downloaded model (FLUX.2-klein included) appears without pressing Refresh.
• New docs/AI_MODELS.md: a one-page index of every AI model the editor uses, what it does and under which licence.

Removed

• MetaHuman DNA body fitting (#1122): qtmesh dna fit, the MCP fit_dna_body tool and the Inspector "Fit DNA Body" section are gone; the results were poor and generating a character with the native MHR rig gives a better one. DNA import, body correctives and the face morph bake (qtmesh dna info|pose|face|psd) are unchanged.

Under the hood

• CI: a unit-test suite that stalls or crashes on the software-GL runner is now retried once; assertion failures are never retried (#1126).
• Website dev dependency: vite 8.0.16 to 8.3.5 (#1120).