Load DINOv2 from Hugging Face without GitHub

Update optimized gallery assets and move samples to row three

Replace gallery assets with optimized GLBs

Hide sample names without changing gallery layout

Expand sample gallery to twelve objects

Remove OAuth gate and enforce anonymous daily quota

Use OAuth-fixed Gradio compatible with model stack

Upgrade Gradio OAuth flow to 6.20

Fix OAuth login redirect target

Enable FCGF with torch 2.11 CUDA 13 MinkowskiEngine

Inject the minimal signed-out CSS at Blocks creation

Keep the signed-out view to the official login button

Show only the official HF button to signed-out users

Register the OAuth gate inside the Blocks context

Gate the entire demo behind official HF sign-in

Add robust official HF OAuth and account quota

Remove custom OAuth quota UI

Require HF login and limit accounts to 5 runs per day

Restore production mode with dedicated GPU preload

Use vox2seq CUDA sm89 wheel for Nvidia L4

Use PyTorch3D CUDA sm89 wheel for Nvidia L4

Diagnose native CUDA kernels on L4

Optimize runtime for dedicated GPU

Pin Space SDK to Gradio 5.50.0

Upgrade Gradio runtime to 5.50.0

Restore authenticated ZeroGPU request path

Disable experimental Gradio SSR

Use public-tier ZeroGPU duration for lazy production

Use tier-compatible duration for cold lazy loading

Fit lazy production callback within ZeroGPU maximum duration

Use ZeroGPU-compatible duration for lazy production inference

Run CANIS production with lazy model loading inside ZeroGPU worker

Apply native wheel compatibility inside ZeroGPU worker

Diagnose ZeroGPU callback without model preload

Avoid ZeroGPU packed-model restore; move preloaded public runtime inside worker

Restore fast public TRELLIS hybrid model loading

Deploy CANIS FPFH production runtime on Torch 2.11 CUDA 13

Use Torch 2.11 CUDA 13 SM120 xFormers wheel

Test Torch 2.11 CUDA 13 native wheels without model preload

Restore verified Torch 2.11 ZeroGPU diagnostic baseline

Probe matched Torch 2.8 and Torchvision 0.23 ZeroGPU stack

Isolate Torch 2.8 in pure ZeroGPU worker probe

Isolate ZeroGPU worker from all user dependencies

Probe ZeroGPU worker with pure Python callback

Defer native CUDA imports in ZeroGPU worker probe

Diagnose ZeroGPU worker with zero model packing

Call a private module-level ZeroGPU job entry

Use FPFH registration without MinkowskiEngine

Stabilize ZeroGPU callback and cached model loading

Use low-overhead preload diagnostics