Spaces:
Paused
Paused
Commit History
-
cleanup, repo name
Silvio Galesso11ac3b5
-
canvas changed appearance
Silvio Galesso380643c -
shipping compilation artifacts
Silvio Galesso465f8e6 -
more compile fixes
Silvio Galesso75aed47 -
more compile fixes
Silvio Galesso6744baf -
compile fix dit-v2
Silvio Galesso75da25d -
adapted timestep emb to older torch compile versions
Silvio Galesso54abf75 -
minor refactor and sampler change
Silvio Galesso4a4e752 -
updated app, engine
Silvio Galesso3bd48a2
-
Expose separate L1 and L2 sampler step counts
Sudhanshu Mittal +1: Claude Opus 4.8L1 and L2 are sampled independently and want very different step counts, but only L1's was reachable: --num_steps is passed to roll_out as the L1 NFE, while L2's came from the config's `l2_pred_NFE` with no CLI override. - rollout_demo_v2.py: add --l2_num_steps, applied via maybe_apply_l2_num_steps() alongside the existing scale overrides. l2_pred_NFE is read at sample time, so setting it post-init is sufficient. Raises if the loaded model has no separately sampled L2 predictor. - app2.py: replace the single "Diffusion steps" slider with two — L1 (5-20, default 10) and L2 (4-8, default 4, matching the config). L2 is consistency-distilled and designed for very few steps, so its useful range is much narrower; the demo text now says so. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>56da3b2 -
Add Orbis 2 hierarchical world model demo (app2.py)
Sudhanshu Mittal +1: Claude Opus 4.8Vendors the Orbis 2 inference code into orbis2/ and adds app2.py, a Gradio demo for the hierarchical L1-L2 model. Orbis 1 (app.py, orbis/) is untouched; the Space still boots app.py until app_file is switched. Unlike Orbis 1, the Orbis 2 rollout takes an mp4 directly and samples its own L1 (high-rate) and L2 (low-rate, further back) context windows, so app2.py re-encodes the upload to a constant 10 fps instead of extracting frames and writing a val config. - orbis2/evaluate/rollout_demo_v2.py: add --num_videos to roll out N futures from one context as a single minibatch (tiling context/steering before building condition_kwargs), and save each to fake_images/sequence_XXXX/, matching the layout app2.py reads. - orbis2/models/second_stage/fm_model_v2.py: label the rollout tqdm bars so the app can scrape progress from the child process. - app2.py: downloads the three models (L1 detail predictor, L2 abstract predictor, tokenizer) from sud0301/orbis2_test into the HF cache and po…a6cc5f0