Spaces:
Running on Zero
Running on Zero
Commit History
-
switch DINOv3 to HF Hub via AutoModel.from_pretrained
torch.hub.load from facebookresearch/dinov3 fails because hubconf.py imports segmentors which transitively require torchmetrics. Using HF transformers AutoModel avoids the dependency chain entirely. Pre-norm features extracted via output_hidden_states[-1] + F.layer_norm. Co-authored-by: Cursor <cursoragent@cursor.com>0775b31 -
switch DINOv3 from HuggingFace transformers to facebookresearch/dinov3 GitHub repo
Loads model via torch.hub.load with auto-downloaded weights from dl.fbaipublicfiles.com. Removes transformers dependency for image conditioning. Co-authored-by: Cursor <cursoragent@cursor.com>85f32da -
Migrate from TRELLIS v1 to TRELLIS.2 pipeline
- Replace trellis/ with trellis2/ (3-stage: sparse structure -> shape -> texture) - Model: microsoft/TRELLIS.2-4B (upgraded from TRELLIS-image-large) - PyTorch 2.6.0 + CUDA 12.4, flash_attn_3, o_voxel, cumesh, flex_gemm - GLB export via o_voxel.postprocess.to_glb with PBR materials - Add HDRI envmaps for shaded rendering (forest, sunset, courtyard) - Keep existing Gradio UI: LitModel3D, video preview, STL export - Remove old extensions/, wheels/, Dockerfile (Gradio SDK space) Co-authored-by: Cursor <cursoragent@cursor.com>e6de801
