Merge PR #2 — Hunyuan3D 2.1 paint subsystem (includes the clean-room rewrite)

[jethac]

PR #2 (devin/1788065256-hunyuan21-paint → master) is the feature-only rebase of the Hunyuan3D 2.1 texture-paint subsystem. It is open, mergeable and green apart from cla-assistant, which fails only because the fork's PERSONAL_ACCESS_TOKEN repository secret is empty — not a real signature failure.

It already includes the clean-room rewrite (PR #4, merged into the paint branch), which replaced unet.py and attention.py wholesale and spliced in the five camera/mesh functions plus view_scale_mapping without reading the old bodies (net −386 lines).

Validation as it stands:

• Tier-2 parity against the reference UNet with real weights: final noise_pred max abs diff 8.7e-06, worst intermediate block 1.1e-03. Tolerances were not touched.
• Getting there required fixing three real bugs the first clean-room push had (commit 33d1f4d9): material self-attention view/material layout scrambling, reference-attention value packing, and a learned_text_clip_ref device-placement crash. Before those fixes parity was 9.1e-01, i.e. wrong.
• Fresh multiview albedo and textured-mesh renders were produced from the clean-room code on a GPU pod.
• Attention routes through ComfyUI's shared optimized_attention (comfy/ldm/hunyuan3d/paint/attention.py), per AGENTS.md.

Known gap for anyone re-running the tests locally: the paint test collection needs safetensors, and a local Torch/TorchVision version mismatch surfaces as RuntimeError: operator torchvision::nms does not exist. Use a matched pair.

Action: merge #2 into the fork's master (protected branch, so a human or authorized agent has to do it). Upstream submission is tracked separately.

[jethac]

Blocks #7.

[jethac]

Current independent review does not approve PR #2 for merge. Keep this issue and the pull request open.

Current GitHub facts: PR #2 is CONFLICTING with master; its merge state is DIRTY; the CLA check and the AI co-author check fail.

The remaining blockers are material:

• The Tencent-derived provenance and GPL compatibility question needs Tencent permission or qualified legal review.
• The model-patcher and checkpoint-loading contract needs an approved implementation.
• The optional DINO input has no usable first-class producer, and there is no complete supported workflow.
• The real end-to-end proof needs an owned or CC0 mesh and reference image, a matched workflow/parity comparison, and GPU bf16 proof.
• Upstream needs a decision about AI-assisted authorship before this work can be proposed there.

PR #10 merged the separate mesh export and GLB attribute fixes. It did not approve or replace this paint subsystem.

[jethac]

Correction/status audit: PR #2 remains conflicting and not approved for merge. Open technical blockers are the model-patcher/checkpoint contract, a supported DINO workflow, and complete matched parity/bf16 evidence with owned or CC0 assets. Provenance/legal and upstream maintainer decisions also remain open. Mesh-only PR #10 did not replace this paint work.

[jethac]

Status update: the clean-room rewrite is integrated and pushed to the existing upstream PR path at commit 2d1c6623. Current evidence is 75/75 focused paint/render/parity tests passing, plus a native direct-node/core run (6 views, 256px, 15 steps, CFG 3, 512px texture, DINO omitted) that produced a finite textured GLB in 52.019s with about 5.0 GB peak CUDA allocation. Full unit tests are 1660 passed, 17 skipped, with two unrelated baseline .m4a classification failures. Keep this issue open pending independent review of the implementation and workflow.

[jethac]

PR #2 was merged normally at 84ba857. All build, code, and platform checks passed; the two fork workflow-policy failures were documented. Upstream review remains tracked separately in PR Comfy-Org#15020 and issue #7.