Empirical finding from 20 mesh-generation runs on RTX 5090 (34 GB):
1536_cascade DOES fit in standalone mode as long as low_vram=True is
on. Three back-to-back runs at the most-quality settings (16/16/16
steps, max_num_tokens=49152, texture_size=4096, decimation_target=300k)
completed in 64-71 seconds warm with no OOM. The README's "1536
currently OOMs on ≤34 GB cards" claim was true only under eager
device placement; with low_vram=True upstream Pixal3D offloads the
4× DinoV3 image_cond_model extractors to CPU between stages, which
frees enough headroom for the shape-1536 NAF attention to land.
Bundled workflow changes (both pixal3d_image_to_mesh.json and
_with_external_rembg variant):
- pipeline_type: 1024_cascade -> 1536_cascade
- max_num_tokens: 32768 -> 49152
- save_obj: False (default) -> True (writes .obj + .mtl + PNG)
New node feature - Pixal3DImageToMesh.save_obj (BOOLEAN, default
False, optional widget appended at the end so old workflows don't
shift their widgets_values list):
- When True alongside save_to_output_dir=True, after writing the GLB
call glb.export() again with .obj extension. trimesh writes the
.obj + .mtl sidecar + a base-color PNG next to it. OBJ is
base-color-only (no PBR slots in the format) - keep the GLB as the
canonical artifact and use the OBJ in DCC tools that prefer it
(Blender, ZBrush, Marmoset, older pipelines).
README rewrites:
- Memory + performance table promotes 1536_cascade as the top row
with verified-on-5090 numbers (~32 GB peak, ~65-70 s warm).
- New "1536_cascade unlock" subsection explains why low_vram=True is
the key, with the 3-runs-in-a-row evidence.
- Fragmentation footnote: 1024_cascade + low_vram=False + 49k tokens
succeeded 10 times then OOMed on the 11th in the same session.
Queue "Pixal3D: Free Pipeline" between batches or restart if you
hit this.
- Use section: brief note on save_obj producing the .obj + .mtl + PNG
trio alongside the GLB.
- Upstream roadmap: rephrased — visualbruno's use_tiled_decoder would
unlock 1536 on 24 GB cards (we already fit it on 34 GB via low_vram).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
