Spaces:
Running on Zero
Running on Zero
| """Start frames (Picture 1) for streaming generation. | |
| A stream starts from a start frame: Picture 1, its static caption and the proxy of that frame. Training | |
| draws Picture 1 from many points of a clip -- the first frame of every ANCHOR_EVERY-th chunk -- either | |
| right at the start of the sampled span (generation from that image, no history) or earlier in the clip. | |
| Start frames inside a caption window flagged by QC are skipped. Per start chunk m: | |
| CLIP/prompts/static_cMM.txt the caption (rtcwm.data.captions --static) | |
| CACHE/CLIP/anchor_cMM.png the frame | |
| CACHE/CLIP/anchor[n]_cMM.safetensors its Picture 1 latent (n: native 1344x768, else 2048 short edge) | |
| CACHE/CLIP/static[n]_cMM.text.safetensors its Qwen states (rtcwm.train.preencode --stage static_text) | |
| """ | |
| from __future__ import annotations | |
| from pathlib import Path | |
| from cwm_h3_inference.constants import STRIDE_FRAMES, WINDOW_FRAMES | |
| from .. import paths | |
| from .clip import read_meta | |
| ANCHOR_EVERY = 3 # chunks between start frames | |
| CLIP_FRAMES = 17 # frames per chunk (one VAE clip) | |
| STREAM_CHUNKS = 54 # chunks of a 934-frame render's stream | |
| LAST_SPAN_START = STREAM_CHUNKS - 6 | |
| def anchor_frame(m: int) -> int: | |
| return CLIP_FRAMES * m | |
| def _flagged(clip: str): | |
| from ..data.qc import lists | |
| _, bad = lists.load() | |
| return [(STRIDE_FRAMES * w, STRIDE_FRAMES * w + WINDOW_FRAMES) for w in bad.get(clip, ())] | |
| def clip_anchors(clip: Path | str) -> list[int]: | |
| """Start chunks of a clip: every ANCHOR_EVERY-th chunk up to the last span start, outside flagged windows.""" | |
| name = clip if isinstance(clip, str) else clip.name | |
| flagged = _flagged(name) | |
| return [m for m in range(0, LAST_SPAN_START + 1, ANCHOR_EVERY) | |
| if not any(a <= anchor_frame(m) < b for a, b in flagged)] | |
| def caption_path(clip: Path, m: int) -> Path: | |
| return clip / "prompts" / f"static_c{m:02d}.txt" | |
| # Picture 1 resolution. "native": the renders' 1344x768 frame as is (1008 H3 rows, ~2.6k Qwen rows; equal or | |
| # better than the released 2048 short edge, which only upsamples our frames -- ops/dbg/res_test.py). | |
| # "2048": the released CWM anchor size (7168 H3 rows, ~8.8k Qwen rows), used by stream_df_v1 / v2. | |
| PICTURE1_RES = ("native", "2048") | |
| def cache_paths(clip: str, m: int, res: str = "native") -> dict: | |
| assert res in PICTURE1_RES, res | |
| d = paths.CACHE / clip | |
| tag = "n" if res == "native" else "" | |
| return dict(png=d / f"anchor_c{m:02d}.png", latent=d / f"anchor{tag}_c{m:02d}.safetensors", | |
| text=d / f"static{tag}_c{m:02d}.text.safetensors") | |