| |
| """genbench 结果拼成每个源一个对比视频（2 列网格，每格 832×480 带标签）。 |
| usage: genbench_compose.py <tag> ... tag ∈ train_v10 train_v9 val_v10 val_v9 L3""" |
| import os |
| import sys |
|
|
| import cv2 |
| import imageio.v3 as iio |
| import numpy as np |
|
|
| G = "/opt/dlami/nvme/zhiyangdeng/maliang/outputs/genbench" |
| L1 = "/opt/dlami/nvme/zhiyangdeng/maliang/outputs/sft_v10/eval_step-8000/L1_cond" |
| W, H = 832, 480 |
| STY = {"orig": "原首帧", "inkwash": "水墨（没见过）", "clay": "黏土（没见过）", "papercraft": "纸艺低多边形（没见过）"} |
|
|
|
|
| def panel(path, k): |
| """k-th 832-wide panel of an mp4 (title bar 48 px cropped) → [81,480,832,3].""" |
| v = iio.imread(path) |
| return v[:, -H:, k * W:(k + 1) * W] |
|
|
|
|
| def label(fr, text): |
| out = fr.copy() |
| cv2.rectangle(out, (0, 0), (W, 34), (0, 0, 0), -1) |
| |
| cv2.putText(out, text, (10, 24), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (255, 230, 120), 2, cv2.LINE_AA) |
| return out |
|
|
|
|
| def grid(tiles, ncol=2): |
| T = min(len(v) for _, v in tiles) |
| while len(tiles) % ncol: |
| tiles.append(("", np.zeros_like(tiles[0][1]))) |
| frames = [] |
| for t in range(T): |
| rows = [np.concatenate([label(v[t], n) for n, v in tiles[r:r + ncol]], 1) for r in range(0, len(tiles), ncol)] |
| frames.append(np.concatenate(rows, 0)) |
| return np.stack(frames) |
|
|
|
|
| def write(path, frames): |
| iio.imwrite(path, frames, fps=16, codec="libx264", quality=None, macro_block_size=16, |
| ffmpeg_params=["-crf", "23", "-pix_fmt", "yuv420p", "-movflags", "+faststart"]) |
| print(path, frames.shape, round(os.path.getsize(path) / 1e6, 1), "MB") |
|
|
|
|
| EN = {"orig": "original first frame", "inkwash": "UNSEEN style: ink-wash", "clay": "UNSEEN style: claymation", |
| "papercraft": "UNSEEN style: papercraft"} |
|
|
|
|
| def l2(tag): |
| d = f"{G}/L2/{tag}" |
| tiles = [] |
| any_mp4 = next(f for f in sorted(os.listdir(d)) if f.endswith(".mp4") and not f.endswith("_gen.mp4")) |
| tiles.append(("input map (numeric ids, parts coloured)", panel(f"{d}/{any_mp4}", 1))) |
| if tag.startswith("val"): |
| cid = open(f"{d}/src/done").read().strip() |
| tiles.append(("GT (held-out run)", panel(f"{L1}/{cid}.mp4", 0))) |
| tiles.append(("generated: original first frame", panel(f"{L1}/{cid}.mp4", 1))) |
| for s in ("orig", "inkwash", "clay", "papercraft"): |
| p = f"{d}/{tag}_{s}.mp4" |
| if os.path.exists(p): |
| tiles.append((f"generated: {EN[s]}", panel(p, 0))) |
| write(f"{G}/compare_{tag}.mp4", grid(tiles)) |
|
|
|
|
| def l3(): |
| tiles = [] |
| for c in sorted(os.listdir(f"{G}/L3")): |
| p = f"{G}/L3/{c}/{c}.mp4" |
| if not os.path.exists(p): |
| continue |
| theme = c.split("__")[0] |
| tiles.append((f"UNSEEN map [{theme}]: numeric ids", panel(p, 1))) |
| tiles.append((f"generated from FLUX first frame [{theme}]", panel(p, 0))) |
| write(f"{G}/compare_L3_unseen_maps.mp4", grid(tiles)) |
|
|
|
|
| if __name__ == "__main__": |
| for tag in sys.argv[1:]: |
| l3() if tag == "L3" else l2(tag) |
|
|
