v0.2.0

0.2.0 (2026-07-07)

Features

• bench: default case matrix (tri-budget sweep + object variety) (6755c28)
• bench: Modal harness — real runtime + postprocess, per-case metrics/cost, visible failures (b91daec)
• gpu-backend: multi-arch ext build (8.0;8.6;8.9) — runs on A10G/A100/L40S; default GPU A10G (584860a)
• gpu-backend: real TRELLIS-2 image — from-source CUDA-extension build (nvdiffrast/diffoctreerast/diff-gaussian-rast/vox2seq, xformers/spconv/kaolin) (19b5abb)
• gpu-backend: thread target_tris through the contract to the runtime (9389939)
• preview: clay preview — self-contained interactive model-viewer HTML (2f38238)

Bug Fixes

• bench: use the real shared TRELLIS image (pip install git+ was invalid) (95447d4)
• gpu-backend: --no-build-isolation for CUDA extension builds (nvdiffrast et al.) (6fb66db)
• gpu-backend: add fast-simplification/pygltflib for postprocess (late layer, keeps ext cache) (e122e0f)
• gpu-backend: drive TRELLIS to_glb simplify from tri budget (texture-aware decimation) (b27149a)
• gpu-backend: drop vox2seq (not part of TRELLIS setup); ext build complete (4e5adbd)
• gpu-backend: force gcc/g++ for CUDA ext builds (inline, keeps torch layer cached) (1226f78)
• gpu-backend: install setuptools/wheel before --no-build-isolation ext builds (685c8ca)
• gpu-backend: set ATTN_BACKEND/SPCONV_ALGO before import, run TRELLIS in-place, robust face count (c4742a3)
• postprocess: preserve provider-baked textures — skip destructive re-unwrap/decimate when textured (8e238aa)

Documentation

• bench: first real TRELLIS-2 results on A10G (5/5 textured, ~1.5-11c/asset) (6d43b40)
• bench: how to run the Clay benchmarks (5907e46)
• bench: input image prompts (single-object, Clay-branded, 3D-friendly) (d0c39ad)