Issues

• Pixal3D multiview on Vulkan (Strix Halo / RADV): SS stage produces 0 active voxels; same input on CPU gives 4385

  raven38/pixal3d.cpp#89 · Filter by author juanjocop opened on Sep 29, 2026

• studio (macOS): dropped files do not load; the file picker works

  raven38/pixal3d.cpp#87 · Filter by author Zwilnik opened on Sep 26, 2026

• num(flow): real-weight forward of tex/shape flow DiTs diverges from the PyTorch reference from step 0 (1.3% tex, 4.1% shape; not the B3 cause)

  raven38/pixal3d.cpp#85 · Filter by author raven38 opened on Sep 22, 2026

• test(pixal3d): ss_res 32/64 E2E quality, memory and backend release gate

  enhancement
  raven38/pixal3d.cpp#84 · Filter by author raven38 opened on Sep 22, 2026

• web(wasm32): bound Pixal3D decode/postprocess within the 4 GiB host heap

  enhancement
  raven38/pixal3d.cpp#83 · Filter by author raven38 opened on Sep 22, 2026

• perf(webgpu): keep ss_res=64 conditioning sparse and below per-buffer limits

  enhancement
  raven38/pixal3d.cpp#82 · Filter by author raven38 opened on Sep 22, 2026

• fix(pixal3d): make cascade coordinate normalization ss_res-aware

  enhancement
  raven38/pixal3d.cpp#81 · Filter by author raven38 opened on Sep 22, 2026

• feat(pixal3d): parameterize sparse-structure resolution and remove hardcoded ss_res=32 assumptions

  enhancement
  raven38/pixal3d.cpp#80 · Filter by author raven38 opened on Sep 22, 2026

• test(pixal3d): pin ss_res=32/64 reference oracle and cascade fixtures

  enhancement
  raven38/pixal3d.cpp#79 · Filter by author raven38 opened on Sep 22, 2026

• epic(pixal3d): sparse-structure res 32/64 correctness + 4 GiB memory safety

  enhancement
  raven38/pixal3d.cpp#78 · Filter by author raven38 opened on Sep 22, 2026

• research(trellis2-mv): confidence-weighted fusion to reduce multidiffusion shape drift

  enhancement
  raven38/pixal3d.cpp#69 · Filter by author raven38 opened on Sep 20, 2026

• test(trellis2-mv): stage parity, E2E quality, performance and memory matrix

  enhancement
  raven38/pixal3d.cpp#68 · Filter by author raven38 opened on Sep 20, 2026

• test(trellis2-mv): pin Python reference and generate golden multi-image fixtures

  enhancement
  raven38/pixal3d.cpp#59 · Filter by author raven38 opened on Sep 20, 2026
  1

• feat(trellis2-mv): native multi-image TRELLIS.2 pipeline (stochastic + multidiffusion)

  enhancement
  raven38/pixal3d.cpp#58 · Filter by author raven38 opened on Sep 20, 2026
  4

• num(metal): decide whether f16 K/V should be the Metal FlashAttention default (accuracy follow-up of #30)

  raven38/pixal3d.cpp#47 · Filter by author raven38 opened on Sep 20, 2026

• web(vram): tile the C2S hn [64,M+1] buffer for maxBufferSize-bound WebGPU adapters

  enhancement
  raven38/pixal3d.cpp#42 · Filter by author raven38 opened on Sep 20, 2026
  1

• vram: verify the MV 1024 pipeline on a real 8 GB card (CUDA) and on Vulkan (Metal: PASS)

  raven38/pixal3d.cpp#41 · Filter by author raven38 opened on Sep 20, 2026
  3

• vram: fit the MV 1024 pipeline in 8 GB CUDA cards with headroom (desktop VRAM budget, tracking)

  enhancement
  raven38/pixal3d.cpp#38 · Filter by author raven38 opened on Sep 20, 2026
  3

• perf(metal): remove avoidable CONT/CPY around QKV and FlashAttention layouts

  enhancement
  raven38/pixal3d.cpp#32 · Filter by author raven38 opened on Sep 19, 2026

• perf(metal): benchmark F16 flow weights vs Q8_0 for large-N DiT GEMMs

  enhancement
  raven38/pixal3d.cpp#31 · Filter by author raven38 opened on Sep 19, 2026

• perf(metal): evaluate F16 FlashAttention K/V fast path with numerical gates

  enhancement
  raven38/pixal3d.cpp#30 · Filter by author raven38 opened on Sep 19, 2026
  2

• web: estimate single-image camera FOV in-browser (replace fixed 20° default)

  raven38/pixal3d.cpp#12 · Filter by author raven38 opened on Sep 16, 2026