n0p/mlx-video
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

format

Browse Source
This commit is contained in:
Prince Canuma
2026-03-18 17:40:05 +01:00
parent 78bcfba31b
commit 17397da70c
77 changed files with 4125 additions and 1655 deletions
Download Patch File Download Diff File Expand all files Collapse all files

85
scripts/video/compare_videos.py
View File

@@ -170,19 +170,33 @@ def print_report(results, ref_path, test_path):
print("AGGREGATE METRICS")
print("-" * 40)
print(f" PSNR (dB): mean={np.mean(psnr):6.2f} min={np.min(psnr):6.2f} max={np.max(psnr):6.2f}")
print(f" SSIM: mean={np.mean(ssim):.4f} min={np.min(ssim):.4f} max={np.max(ssim):.4f}")
print(f" Mean diff: mean={np.mean(md):6.2f} min={np.min(md):6.2f} max={np.max(md):6.2f}")
print(f" Max diff: mean={np.mean(mx):6.1f} min={np.min(mx):6.1f} max={np.max(mx):6.1f}")
print(f" Color dist: mean={np.mean(cd):.4f} min={np.min(cd):.4f} max={np.max(cd):.4f}")
print(
f" PSNR (dB): mean={np.mean(psnr):6.2f} min={np.min(psnr):6.2f} max={np.max(psnr):6.2f}"
)
print(
f" SSIM: mean={np.mean(ssim):.4f} min={np.min(ssim):.4f} max={np.max(ssim):.4f}"
)
print(
f" Mean diff: mean={np.mean(md):6.2f} min={np.min(md):6.2f} max={np.max(md):6.2f}"
)
print(
f" Max diff: mean={np.mean(mx):6.1f} min={np.min(mx):6.1f} max={np.max(mx):6.1f}"
)
print(
f" Color dist: mean={np.mean(cd):.4f} min={np.min(cd):.4f} max={np.max(cd):.4f}"
)
print()
print("TEMPORAL COHERENCE (mean frame-to-frame diff, lower = smoother)")
print("-" * 40)
print(f" Reference: {results['ref_temporal_coherence']:.2f}")
print(f" Test: {results['test_temporal_coherence']:.2f}")
ratio = results["test_temporal_coherence"] / (results["ref_temporal_coherence"] + 1e-10)
print(f" Ratio: {ratio:.2f}x {'(test is smoother)' if ratio < 1 else '(test is jerkier)' if ratio > 1.05 else '(similar)'}")
ratio = results["test_temporal_coherence"] / (
results["ref_temporal_coherence"] + 1e-10
)
print(
f" Ratio: {ratio:.2f}x {'(test is smoother)' if ratio < 1 else '(test is jerkier)' if ratio > 1.05 else '(similar)'}"
)
print()
# Identify worst frames
@@ -190,7 +204,9 @@ def print_report(results, ref_path, test_path):
print("-" * 40)
worst_idx = np.argsort(psnr)[:5]
for i in worst_idx:
print(f" Frame {i:4d}: PSNR={psnr[i]:6.2f} dB SSIM={ssim[i]:.4f} mean_diff={md[i]:.2f}")
print(
f" Frame {i:4d}: PSNR={psnr[i]:6.2f} dB SSIM={ssim[i]:.4f} mean_diff={md[i]:.2f}"
)
print()
# Quality assessment
@@ -210,7 +226,9 @@ def print_report(results, ref_path, test_path):
grade = "Very different"
print(f" Overall: {grade} (PSNR={mean_psnr:.1f} dB, SSIM={mean_ssim:.4f})")
if mean_psnr < 30:
print(" ⚠ Videos differ significantly — likely a bug or different generation seed")
print(
" ⚠ Videos differ significantly — likely a bug or different generation seed"
)
print("=" * 72)
@@ -242,9 +260,7 @@ def main():
parser.add_argument(
"--diff-video", help="Save side-by-side diff visualization to this path"
)
parser.add_argument(
"--max-frames", type=int, help="Compare only first N frames"
)
parser.add_argument("--max-frames", type=int, help="Compare only first N frames")
parser.add_argument(
"--ssim-win", type=int, default=7, help="SSIM window size (default: 7)"
)
@@ -254,26 +270,29 @@ def main():
default=5.0,
help="Diff heatmap amplification (default: 5.0)",
)
parser.add_argument(
"--csv", help="Export per-frame metrics to CSV file"
)
parser.add_argument("--csv", help="Export per-frame metrics to CSV file")
args = parser.parse_args()
print(f"Loading reference: {args.reference}")
ref_frames, ref_fps = load_video(args.reference, args.max_frames)
print(f" → {len(ref_frames)} frames, {ref_fps:.1f} fps, {ref_frames[0].shape[1]}x{ref_frames[0].shape[0]}")
print(
f" → {len(ref_frames)} frames, {ref_fps:.1f} fps, {ref_frames[0].shape[1]}x{ref_frames[0].shape[0]}"
)
print(f"Loading test: {args.test}")
test_frames, test_fps = load_video(args.test, args.max_frames)
print(f" → {len(test_frames)} frames, {test_fps:.1f} fps, {test_frames[0].shape[1]}x{test_frames[0].shape[0]}")
print(
f" → {len(test_frames)} frames, {test_fps:.1f} fps, {test_frames[0].shape[1]}x{test_frames[0].shape[0]}"
)
if ref_frames[0].shape != test_frames[0].shape:
print(f"Warning: resolution mismatch {ref_frames[0].shape} vs {test_frames[0].shape}")
print(
f"Warning: resolution mismatch {ref_frames[0].shape} vs {test_frames[0].shape}"
)
print("Resizing test frames to match reference...")
h, w = ref_frames[0].shape[:2]
test_frames = [
cv2.resize(f, (w, h), interpolation=cv2.INTER_LANCZOS4)
for f in test_frames
cv2.resize(f, (w, h), interpolation=cv2.INTER_LANCZOS4) for f in test_frames
]
print("Computing metrics...")
@@ -282,23 +301,29 @@ def main():
print_report(results, args.reference, args.test)
if args.diff_video:
save_diff_video(ref_frames, test_frames, args.diff_video, ref_fps, args.diff_scale)
save_diff_video(
ref_frames, test_frames, args.diff_video, ref_fps, args.diff_scale
)
if args.csv:
import csv
with open(args.csv, "w", newline="") as f:
writer = csv.writer(f)
writer.writerow(["frame", "psnr", "ssim", "mean_diff", "max_diff", "color_dist"])
writer.writerow(
["frame", "psnr", "ssim", "mean_diff", "max_diff", "color_dist"]
)
for i in range(results["num_frames"]):
writer.writerow([
i,
f"{results['psnr'][i]:.4f}",
f"{results['ssim'][i]:.6f}",
f"{results['mean_diff'][i]:.4f}",
f"{results['max_diff'][i]:.1f}",
f"{results['color_dist'][i]:.6f}",
])
writer.writerow(
[
i,
f"{results['psnr'][i]:.4f}",
f"{results['ssim'][i]:.6f}",
f"{results['mean_diff'][i]:.4f}",
f"{results['max_diff'][i]:.1f}",
f"{results['color_dist'][i]:.6f}",
]
)
print(f"Per-frame metrics saved to {args.csv}")

78
scripts/video/video_quality.py
View File

@@ -158,10 +158,14 @@ def analyze_video(frames, chunk_size=None, compute_flow=False):
boundary_metrics = []
for b in boundaries:
if b < n and b > 0:
pre = metrics["frame_diff"][b - 1] if b > 1 else metrics["frame_diff"][1]
pre = (
metrics["frame_diff"][b - 1] if b > 1 else metrics["frame_diff"][1]
)
at = metrics["frame_diff"][b]
ratio = at / (pre + 1e-10)
brightness_jump = metrics["brightness"][b] - metrics["brightness"][b - 1]
brightness_jump = (
metrics["brightness"][b] - metrics["brightness"][b - 1]
)
contrast_jump = (
(metrics["contrast"][b] - metrics["contrast"][b - 1])
/ (metrics["contrast"][b - 1] + 1e-10)
@@ -198,7 +202,9 @@ def print_report(metrics, path, fps, total_frames, frames_analyzed):
print("VIDEO QUALITY REPORT")
print("=" * 72)
print(f" File: {path}")
print(f" Total frames: {total_frames} Analyzed: {frames_analyzed} FPS: {fps:.1f}")
print(
f" Total frames: {total_frames} Analyzed: {frames_analyzed} FPS: {fps:.1f}"
)
duration = total_frames / fps if fps > 0 else 0
print(f" Duration: {duration:.1f}s")
print()
@@ -211,52 +217,76 @@ def print_report(metrics, path, fps, total_frames, frames_analyzed):
print("-" * 40)
if n_uniform:
frames_list = np.where(metrics["is_uniform"])[0][:10]
print(f" Uniform/blank frames: {n_uniform} — frames {list(frames_list)}{'...' if n_uniform > 10 else ''}")
print(
f" Uniform/blank frames: {n_uniform} — frames {list(frames_list)}{'...' if n_uniform > 10 else ''}"
)
if n_noisy:
frames_list = np.where(metrics["is_noisy"])[0][:10]
print(f" Noisy frames: {n_noisy} — frames {list(frames_list)}{'...' if n_noisy > 10 else ''}")
print(
f" Noisy frames: {n_noisy} — frames {list(frames_list)}{'...' if n_noisy > 10 else ''}"
)
print()
print("SHARPNESS")
print("-" * 40)
print(f" Laplacian var: mean={np.mean(sl):8.1f} min={np.min(sl):8.1f} max={np.max(sl):8.1f} std={np.std(sl):.1f}")
print(f" Gradient mag: mean={np.mean(sg):8.2f} min={np.min(sg):8.2f} max={np.max(sg):8.2f} std={np.std(sg):.2f}")
print(
f" Laplacian var: mean={np.mean(sl):8.1f} min={np.min(sl):8.1f} max={np.max(sl):8.1f} std={np.std(sl):.1f}"
)
print(
f" Gradient mag: mean={np.mean(sg):8.2f} min={np.min(sg):8.2f} max={np.max(sg):8.2f} std={np.std(sg):.2f}"
)
if np.std(sl) / (np.mean(sl) + 1e-10) > 0.3:
print(" ⚠ High sharpness variation — possible blur artifacts")
print()
print("BRIGHTNESS & CONTRAST")
print("-" * 40)
print(f" Brightness: mean={np.mean(br):6.1f} min={np.min(br):6.1f} max={np.max(br):6.1f} std={np.std(br):.2f}")
print(f" Contrast (std): mean={np.mean(ct):6.1f} min={np.min(ct):6.1f} max={np.max(ct):6.1f} std={np.std(ct):.2f}")
print(
f" Brightness: mean={np.mean(br):6.1f} min={np.min(br):6.1f} max={np.max(br):6.1f} std={np.std(br):.2f}"
)
print(
f" Contrast (std): mean={np.mean(ct):6.1f} min={np.min(ct):6.1f} max={np.max(ct):6.1f} std={np.std(ct):.2f}"
)
if np.std(br) > 3.0:
print(" ⚠ Brightness instability — may indicate chunk boundary artifacts")
print()
print("COLOR DISTRIBUTION (BGR)")
print("-" * 40)
print(f" Blue: mean={np.mean(metrics['color_mean_b']):6.1f} std={np.std(metrics['color_mean_b']):.2f}")
print(f" Green: mean={np.mean(metrics['color_mean_g']):6.1f} std={np.std(metrics['color_mean_g']):.2f}")
print(f" Red: mean={np.mean(metrics['color_mean_r']):6.1f} std={np.std(metrics['color_mean_r']):.2f}")
print(
f" Blue: mean={np.mean(metrics['color_mean_b']):6.1f} std={np.std(metrics['color_mean_b']):.2f}"
)
print(
f" Green: mean={np.mean(metrics['color_mean_g']):6.1f} std={np.std(metrics['color_mean_g']):.2f}"
)
print(
f" Red: mean={np.mean(metrics['color_mean_r']):6.1f} std={np.std(metrics['color_mean_r']):.2f}"
)
print()
print("TEMPORAL STABILITY")
print("-" * 40)
fd_nz = fd[1:] # skip first frame (always 0)
if len(fd_nz) > 0:
print(f" Frame diff: mean={np.mean(fd_nz):6.2f} min={np.min(fd_nz):6.2f} max={np.max(fd_nz):6.2f} std={np.std(fd_nz):.2f}")
print(
f" Frame diff: mean={np.mean(fd_nz):6.2f} min={np.min(fd_nz):6.2f} max={np.max(fd_nz):6.2f} std={np.std(fd_nz):.2f}"
)
if np.std(fd_nz) / (np.mean(fd_nz) + 1e-10) > 0.5:
print(" ⚠ High diff variance — jitter or discontinuities")
if "flow_mean" in metrics:
fm = metrics["flow_mean"][1:]
print(f" Optical flow: mean={np.mean(fm):6.2f} max_frame={np.max(metrics['flow_max'][1:]):.1f}")
print(
f" Optical flow: mean={np.mean(fm):6.2f} max_frame={np.max(metrics['flow_max'][1:]):.1f}"
)
print()
# Chunk boundaries
if "boundaries" in metrics and metrics["boundaries"]:
print("CHUNK BOUNDARIES")
print("-" * 40)
print(f" {'Frame':>6} {'Diff ratio':>10} {'Brightness':>10} {'Contrast %':>10} {'Sharpness %':>11}")
print(
f" {'Frame':>6} {'Diff ratio':>10} {'Brightness':>10} {'Contrast %':>10} {'Sharpness %':>11}"
)
for bm in metrics["boundaries"]:
print(
f" {bm['frame']:6d}"
@@ -267,7 +297,9 @@ def print_report(metrics, path, fps, total_frames, frames_analyzed):
)
avg_ratio = np.mean([b["diff_ratio"] for b in metrics["boundaries"]])
if avg_ratio > 2.0:
print(f" ⚠ Boundary diff ratio {avg_ratio:.1f}x — visible chunk transitions")
print(
f" ⚠ Boundary diff ratio {avg_ratio:.1f}x — visible chunk transitions"
)
print()
# Overall grade
@@ -303,9 +335,7 @@ def main():
type=int,
help="Frames per chunk for boundary analysis (e.g., 32)",
)
parser.add_argument(
"--start", type=int, default=0, help="Start frame (default: 0)"
)
parser.add_argument("--start", type=int, default=0, help="Start frame (default: 0)")
parser.add_argument("--end", type=int, help="End frame (default: all)")
parser.add_argument(
"--flow",
@@ -329,8 +359,14 @@ def main():
import csv
keys = [
"sharpness_lap", "sharpness_grad", "brightness", "contrast",
"color_mean_b", "color_mean_g", "color_mean_r", "frame_diff",
"sharpness_lap",
"sharpness_grad",
"brightness",
"contrast",
"color_mean_b",
"color_mean_g",
"color_mean_r",
"frame_diff",
]
if args.flow:
keys += ["flow_mean", "flow_max"]