Fetching metadata from the HF Docker repository...
Sep 30 20 days ago 04:09 Update project page and GitHub links to antigravity-tech Running Sep 13 14:22 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 14:09 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 13:53 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 12:53 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 11:05 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 10:55 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 10:43 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 08:43 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 08:04 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 07:39 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 07:36 Auto point cloud after Run; clarify Indoor/Outdoor is PC-only Single on_run handler; backup at backup/app_pre_dap_debug.py 06:48 Restore point cloud UI; Large only; keep ZeroGPU compute_depth Single on_run handler; backup at backup/app_pre_dap_debug.py 05:24 Fix Gradio 6: remove allow_flagging from Interface Single on_run handler; backup at backup/app_pre_dap_debug.py 05:18 Fix startup: remove open3d hard import for debug Space Single on_run handler; backup at backup/app_pre_dap_debug.py 05:09 Fix startup: remove open3d hard import for debug Space Single on_run handler; backup at backup/app_pre_dap_debug.py 04:25 gr.Interface + compute_depth @spaces.GPU; .to('cuda') fix Single on_run handler; backup at backup/app_pre_dap_debug.py 02:18 Gradio 6.15 + module-level demo (PaGeR ZeroGPU pattern) Single on_run handler; backup at backup/app_pre_dap_debug.py Sep 12 15:42 ZeroGPU: compute_depth directly on .click (HF doc pattern) Single on_run handler; backup at backup/app_pre_dap_debug.py 15:34 PaGeR .then() ZeroGPU chain + duration=120 for Large Single on_run handler; backup at backup/app_pre_dap_debug.py 15:28 Large only, DAP-shaped predict_depth + DA360 KITTI colormap Single on_run handler; backup at backup/app_pre_dap_debug.py 15:25 Fix ZeroGPU: bind @spaces.GPU directly to Run click Single on_run handler; backup at backup/app_pre_dap_debug.py 14:58 Fix Run hang: no .to(cuda) on click, preprocess inside GPU Single on_run handler; backup at backup/app_pre_dap_debug.py 14:16 Debug: DAP-style depth-only, DA360 scale-invariant colormap Single on_run handler; backup at backup/app_pre_dap_debug.py 12:15 Fix Space Run error: gpu_run_depth single GPU .then() handler 09:34 Sync Space deploy: local fixes, show_progress, ZeroGPU chain Aug 24 14:37 Fix ZeroGPU hang: gpu_forward via .then() chain, duration 25s 14:07 Fix example image flicker: gr.skip when no resize needed 14:00 Fix hang: CPU resize before ZeroGPU, split predict_depth/gpu_infer Aug 23 13:40 Restore checkpoint dropdown and Indoor/Outdoor example metadata 13:34 Match DAP/PaGeR: numpy in, depth-only GPU click, optional point cloud, no filepath on worker Aug 22 03:46 Request only 25s ZeroGPU so leftover free quota can still be scheduled 03:19 Root fix: @spaces.GPU(duration=60) on the Gradio click handler (was requesting 180s vs 154s left) Aug 19 15:01 Fix ZeroGPU abort: numpy IPC, no generator nest, .then() chain, duration 120s 14:37 Fix RUNTIME_ERROR: css belongs on Blocks, not launch under spaces+Gradio 5.50 Aug 17 14:43 Fix GPU task aborted: use global model in @spaces.GPU, raise Small duration to 60s Aug 16 15:20 Fix Markdown rendering: dedent so Gradio 5 does not treat title as code 15:18 Use native gr.ImageSlider; ssr_mode=False for ZeroGPU auth 15:18 Space card: sdk_version 5.50.0 15:18 deps: Gradio 5.50 for ZeroGPU login quota 15:14 Request only 10s GPU for packed Small so low leftover quota can run 15:08 Stream depth first; speed up point cloud; skip Large download at startup 15:07 Docs: Small/Base prefetch, two-step UI 14:59 Tighten ZeroGPU duration so low remaining quota can still run Small 14:55 ZeroGPU: pack model on cuda at startup 14:55 Fix Space README YAML metadata 14:53 Upload README.md with huggingface_hub 14:49 Upload folder using huggingface_hub Aug 6 05:32 Use dynamic ZeroGPU duration: 25s once model is already on GPU, 90s only for a cold model switch Since the default model is now moved to GPU at startup, most calls only need
to cover the forward pass. Requesting a smaller duration reduces quota cost
per call (quota is charged against the declared duration, not actual usage). 05:27 Add stage timing logs (load_model, preprocess, GPU forward, postprocess, PLY export) DAP uses the same vitl backbone at nearly identical resolution (512x1024 vs
our 518x1036), so the model forward pass itself shouldn't explain the gap.
Logging per-stage wall time to find where it's actually going (cold-start,
transfer, forward, or post-processing) instead of guessing further. 
