Skip to content

3.39.0

Choose a tag to compare

View all tags
@fernandotonon fernandotonon released this 15 Sep 20:10
3.39.0
dc4ba6b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Three new local-AI features, a text-to-motion upgrade, and two fixes for silent failures.

AI: Image → 3D

BiRefNet high-quality matting (#1016) — a new Best tier for background removal, alongside U²-Net as Fast. BiRefNet (MIT) runs at 1024² versus U²-Net's 320² and is measurably cleaner: on a reference asset, 0.83% vs 2.38% ambiguous edge pixels (2.9× crisper).

That matters for reconstruction quality, not just the cutout. On a test goblin the two mattes disagree on 1.96% of pixels — but 45:1 in favour of subject BiRefNet keeps and U²-Net erases (32,394 px vs 719 px). TRELLIS.2 can only reconstruct what survives the matte, so the Best mesh resolves ear tips, harness straps and individual fingers the Fast one flattens.

Opt-in (--matting best, or the GUI checkbox) — it is a ~930 MB download and several CPU-seconds. Default behaviour is unchanged.

Photo depth estimation (#1018) — Depth-Anything-V2-Small (Apache-2.0) estimates relative depth from a single photograph, emitting an 8-bit map with near = bright, the same convention the mesh-depth renderer uses. A photo and a rendered mesh are therefore interchangeable as ControlNet conditioning images.

qtmesh material --photo-depth photo.png, the Material Editor's "Depth from photo…" button, and MCP photo_depth.

Text to motion

MotionComposer (#1010) — multi-step prompts now compose. "walk forward, sit down, wave twice, then stand up" previously played a single walk clip; it now produces one continuous clip, stitching library takes at their best-matching pose pair with a short crossfade. Works offline with no LLM installed (deterministic connective parsing), and agents can supply a timeline directly via the MCP script argument.

Loop metadata (#1009) — per-clip loop points, so "wave twice" cycles the wave itself instead of replaying its lead-in and settle. Derived at parse time, so existing libraries gain them with no redownload.

Paint v2

Bake-up workflow (#552) and CLI/MCP parity (#553) — closes epic #543. Painted layers bake to engine-ready PBR sets (Unity, Unreal, Godot, glTF, generic), and 14 paint_* MCP tools bring the paint surface to parity.

Mesh tools

Weld duplicate vertices (#991) — fixes animation tearing on generated meshes, where co-located but unconnected vertices pull apart under deformation.

Fixes

  • Decal/projection painted an all-white layer on non-square textures (#995).
  • UniRig reported a misleading error (#1025). --algo unirig has never run the ML model: the hosted encoder emits 100% NaN latents, so every call silently fell back to the template rig. The failure surfaced 1000+ decode steps later as "constrained decode reached a dead state", pointing at the tokenizer. It now names the real cause and fails in 2.4s instead of 2m24s. The re-export is tracked in #1025.
  • TRELLIS.2 license-notices concat wrote 0 bytes on Windows.

Notes

  • New models download on first use from the HF models repo; both are also mirrored with standalone model cards.
  • Every new export asserts torch-vs-ONNX parity before upload and refuses to write a graph containing non-finite output — added after #1025 showed a bad export can go unnoticed for months.
  • MCP SERVER_VERSION → 1.10.0 (new photo_depth tool, 14 paint_* tools, matting argument).
Assets 15
Loading