Use sm90a gsplat wheel

Use rebuilt gsplat wheel

Force xformers CUTLASS attention on ZeroGPU

Use rebuilt nvdiffrast wheel

Restore custom nvdiffrast wheel

Load default image from assets in callbacks

Avoid stale Gradio default image path

Switch nvdiffrast wheel source

Pin supported ZeroGPU PyTorch stack

Prefer CUDA PyTorch wheels for ZeroGPU

Update ZeroGPU runtime dependencies

ui: move vid_roll to top-right row, restore vid_hdri below vid_cam

ui: remove PBR GLB export from online demo

perf(glb): further reduce bake params to avoid 120s timeout

perf(glb): speed up bake by reducing hardcoded defaults

revert: disable NeAR main-process preload; restore torchvision weights=...

ui: update title and description

chore: bump default video frames back to 48, step 2

fix(app): remove @torch .no_grad from generate_glb

fix: enable NeAR CPU preload by bypassing torchvision weights CUDA init

fix(app): split renderings into two @GPU callbacks; lower defaults

fix(app): do NOT preload NeAR in main process; warm disk caches only

fix(app): lazy-import trellis, preload NeAR on CPU, fix ZeroGPU duration

fix: replace DINOv2 torch.hub.load with download_url_to_file in preload

fix: use @GPU (no duration) for geometry and slat, matching app_hyshape.py

fix: increase GPU lease duration and reduce Hunyuan3D inference steps

fix: pre-cache RMBG-2.0 and DINOv2 in preload worker to prevent GPU lease timeout

fix: enable gradients for texture baking in GLB export

fix: restore clear_cache function definition

refactor: merge ③④ GPU callbacks for ZeroGPU stability

fix(zerogpu): lazy-import nvdiffrast in postprocessing_utils

fix(preload): match app_hyshape.py — load Hunyuan3D to CPU RAM, cache NeAR to disk

fix(zerogpu): remove _MODEL_LOCK from GPU callbacks; preload = disk cache only

fix(preload): split into disk-cache phase + CPU-load phase to unblock GPU callbacks

fix(zerogpu): skip NeAR CPU preload — BiRefNet/DINOv2 init triggers CUDA

refactor(app): ZeroGPU-safe 4-step pipeline with Hunyuan PBR GLB export

fix(zerogpu): add explicit duration, disable geometry offload by default

fix(zerogpu): remove stale CUDA flags, enable CPU preload by default

fix(near): defer renderer/tone-mapper init to first GPU ensure

refactor(app): ZeroGPU-oriented NeAR UI; path-only SLaT; dual videos

fix(gsplat): 2DGS image fit — expand colors to [1,N,3] for packed=False

docs: clarify Space app_file vs dashboard; fix README structure bullets

feat(gsplat): Gradio image-fitting app for ZeroGPU; default Space app_file

Space: move pipelines to CUDA only when spaces SDK is present

Gradio: apply NEAR_GRADIO_THEME (blue) to Blocks

HF Space: load pipelines before module-level demo, fix import entry

Space app: SLaT session cache, tone-mapper safety, CUDA/inference fixes

revert: remove sparse attention _linear inference-tensor clone hack

fix(sparse-attn): clone inference feats before Linear when grad is enabled

fix(sparse): spconv replace() must wire new feats into data.features