scripts/_make_cmp_grid.py

import os, glob, sys
import torch
from torchvision.io import read_video
from torchvision.utils import save_image

ROOT = "workspace/outputs/gen_4chunk_motion_delta_vs_multires"
OUT = "workspace/outputs/gen_4chunk_motion_delta_vs_multires/_grids"
os.makedirs(OUT, exist_ok=True)
FRAMES = [0, 6, 12, 18]  # sample across the 21-frame clip

def load(path):
    v, _, _ = read_video(path, pts_unit="sec", output_format="TCHW")
    return v.float() / 255.0

# key prompts (substring match), across train+ood
keys = ["skier", "golden retriever", "Mercedes", "flyby video", "Aerial drone"]
pairs = []
for s in ["train", "ood"]:
    for mp in sorted(glob.glob(os.path.join(ROOT, "motion_delta", s, "*.mp4"))):
        name = os.path.basename(mp)
        if any(k.lower() in name.lower() for k in keys):
            rp = os.path.join(ROOT, "multires", s, name)
            if os.path.exists(rp):
                pairs.append((s, name, mp, rp))

for s, name, mp, rp in pairs:
    vm, vr = load(mp), load(rp)
    n = min(vm.shape[0], vr.shape[0])
    idx = [i for i in FRAMES if i < n]
    rows = torch.cat([vm[idx], vr[idx]], dim=0)  # top: motion_delta, bottom: multires
    tag = (s + "_" + name[:32]).replace(" ", "_").replace("/", "_").replace(".", "")
    outp = os.path.join(OUT, tag + ".png")
    save_image(rows, outp, nrow=len(idx))
    print(f"{outp}  frames={idx}  (top=motion_delta bottom=multires)")