| title: DVSM | |
| emoji: 🦀 | |
| colorFrom: yellow | |
| colorTo: indigo | |
| sdk: gradio | |
| sdk_version: 6.16.0 | |
| python_version: '3.10' | |
| app_file: app.py | |
| pinned: false | |
| license: other | |
| # TTB-DVSM — interactive view synthesis on ZeroGPU | |
| Upload images / a short orbit video → π³ pose estimation + TTB-DVSM | |
| reconstruction (one GPU call) → a time-boxed interactive viser viewer. | |
| Models (`nvidia/DVSM`): uploads are reconstructed with `DVSM-TTB-p8-general-960p`; the gallery examples were baked with `DVSM-TTB-p8-DL3DV-960p`. | |
| ## ZeroGPU notes | |
| Select **ZeroGPU** hardware in the Space Settings (it is not set via this | |
| frontmatter). ZeroGPU forks a fresh worker per `@spaces.GPU` call and cannot hold a GPU | |
| permanently. The interactive viewer therefore runs inside a `@spaces.GPU` | |
| **generator** that holds the GPU for one bounded session (`SESSION_SECONDS`, | |
| default 180); the main process reverse-proxies `/viser-proxy/*` to the | |
| worker-local viser port. When the session expires the GPU is released — click | |
| "Start / Continue Live Session" to resume from the already-computed artifact. | |
| `flash-attn` is not installed; `dvsm/models/modules.py` falls back to torch SDPA | |
| (numerically equivalent). A second-phase optimization is to add a FlashAttn-3 | |
| branch in `_attn_func`. | |
| Env knobs: `SESSION_SECONDS`, `RECON_DURATION`, `BENCH_WIDTHS`. | |
| Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference | |