Spaces:
Running on Zero
Running on Zero
Commit History
14 days ago
-
fix: lower default temp to 0.1, raise MaskGIT steps to 20 for sharper frames Running
Higher temperature (0.5) was causing stochastic blending across tokens, producing blurry/mixed frames. Lower temperature makes each MaskGIT step pick the most confident token, reducing visual noise. More steps (20 vs 10) gives the diffusion process more iterations to refine uncertain positions. Also opens Quality settings accordion by default so users can see/adjust. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>7ba6bdd -
fix: use CONTEXT_WINDOW=4 to match training context_length=4
Dynamics model was trained on 4-frame sequences (context_length=4 in training.yaml). Running inference with only 2 context frames creates an out-of-distribution scenario for every temporal attention layer, causing incoherent predictions. Changes: - CONTEXT_WINDOW 2→4: init, step_game, and run_inference all use 4 frames - step_game: action_latent is now [1, CONTEXT_WINDOW, A] = [1, 4, A], built as lam.encode(ctx)[1,T-1,A] + cur_q[1,1,A], matching training - Free mode: repeats last inferred action instead of expanding past action - run_inference (auto-generate): fixes unconditioned path to infer past actions from context instead of passing conditioning=None (OOD) - Added debug print of tensor shapes in step_game Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>eac35c2 -
fix: use raw LAM encoder output (not quantized) for past-action conditioning
Mirrors the use_gt_actions path in inference_utils.py which skips quantization for the historical action slot — reduces quantization error in context and matches the higher-quality training mode more closely. Continue: [act_raw, act_raw] Action buttons: [act_raw, quantized_user_token] Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>bbf1324 -
fix: revert play defaults to temp=0.5/steps=10; fix expand->repeat for action latent
- Temperature 0.1 and steps=20 produced worse output than the original auto-generate defaults (0.5/10). Revert to match tested working config. - act_q.expand() created a non-contiguous view; replace with .repeat() to give the dynamics model an independent copy for each time step. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>8939f70 -
fix: infer action from context for Continue btn; add VT reconstruction frame for diagnosis
- Continue button now runs lam.encode(ctx) to get the natural next action instead of conditioning=None, matching training-time behavior more closely - init_game shows a 'VT recon' frame (gold border) = video tokenizer roundtrip of context; if this looks bad the issue is in the tokenizer - Debug print of next_frame value range to HF Space logs Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f766fb4 -
feat: add unconditioned 'Continue' button for coherent frame continuations
The dynamics model runs without action conditioning (action_latent=None), predicting the most likely next frame from context alone. This avoids the visual discontinuity caused by arbitrary action token injection. D-pad buttons are kept as experimental/exploratory controls. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>66d27dd -
fix: lower default temperature to 0.1, raise MaskGIT steps to 20 for better play quality
Expose temperature + MaskGIT steps sliders in 'Quality settings' accordion. Higher step count gives the masked diffusion more iterations to resolve tokens; lower temperature picks more confident (sharper) predictions. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f4083f5 -
fix: skip title screens + add color-coded borders for context vs generated frames
- Sample from middle 60% of dataset to avoid title/loading screens - Grey-blue border on context frames, green border on generated frames - Same sampling fix applied to auto-generate tab Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>9432abc -
feat: replace single play frame with accumulating frame grid (gr.Gallery)
- init_game seeds gallery with 2 context frames labeled 'Start 1/2' - step_game appends each new frame with caption 'Step N <action>' - Gallery shows 4 columns so the temporal sequence is easy to read - State now carries gallery list and step counter Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>d94fee2 -
feat: add interactive Play tab with D-pad controls and 4x pixel-art upscaling
- New 'Play' tab with D-pad (←↑↓→) and face buttons (A/B) for turn-based play - Each button maps to a learned action token; user discovers effects by experimenting - init_game() loads random context frames to start a session - step_game() runs one dynamics inference step per button press - Auto-generate tab: upscale video 4x (INTER_NEAREST) for crisp pixel-art look - Added 'How it works' tab with architecture table and latency explanation - CSS for pixel-perfect image rendering and styled D-pad/face buttons Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>4e0f9aa -
style: polish Gradio UI with Soft theme, header, accordion, and How it works section
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>3378d20 -
fix: use _find_ckpt_pth + _pth_to_ckpt_dir in get_models to load single .pth HF checkpoints
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>28e2620 -
debug: print all HF model repo files to diagnose checkpoint structure
0f4be9d -
feat: add multi-game selector (all 5 games) with suite fallback
df1fde2 -
fix: add @spaces.GPU decorator for ZeroGPU compatibility
765cb06 -
fix: pin starlette<0.39 and jinja2<3.2 for gradio 4.44.0 compat
169f3e4 -
fix: add api_name=False to skip gradio_client schema bug on gr.Video
20bf309 -
fix: remove gradio from requirements, let sdk_version manage it
9d4c5ac -
fix: pin gradio_client==0.9.1, remove gr.Progress(), add server_name
c848e1e -
fix: pin huggingface_hub<0.26 to keep HfFolder, revert gradio to 4.44.x
4a6ac8c -
fix: upgrade gradio to 5.x to resolve HfFolder import error
974dc7d
15 days ago
-
Initial HF Space deployment
dcc4594