Skip to content

Cull Score

Drop in a whole shoot. The node scores every frame, skips blurry, badly exposed and near-duplicate shots, and hands back the keepers, best first.

Cull Score — a batch with the sharp frame kept and the blurred and blown-out frames dropped
In ComfyUI Add Node → ComfyUI-JH-PixelPro → qa → QA: Cull Score
Good for Shortlisting a burst or a portrait session, removing blurry and blown-out frames, keeping one of several near-identical shots
Speed Each photo is measured on a 1024 px copy, so the cost does not depend on the original size

When to use it

After a shoot you often have dozens of near-identical frames. This node is a helper to shortlist them. It is not a judge of expression or composition, so look at the picks yourself before you deliver.

It checks four things: sharpness, exposure, closed eyes (when it finds faces) and duplicates. Sharpness is compared with the sharpest shot of the same scene, so a soft-looking portrait does not lose to a busy street photo that happens to be in the same batch.

Quick start

  1. Load a batch of same-size photos (for example with Load Image Batch).
  2. Connect it to QA: Cull Score.
  3. Send selected to Preview Image or Save Image. Read report with a text display node if you want the numbers.
  4. Set top_k to the number of photos you want, or leave it at 0 to keep every photo that passes min_score.

Settings

Setting Default What it does Tip
images A batch of photos, all the same size.
top_k 0 0 to 1000. Keep at most this many photos. 0 keeps every photo that passes. Set it to 1 to get only the best frame.
min_score 0.5 0 to 1. Drops photos scoring below this. Raise it for a stricter cull, lower it to keep more.
check_eyes on Lowers the score of photos where a face has closed eyes. Needs MediaPipe, and is skipped if it is missing. Turn off for scenes without people.

Outputs

Output What it is
selected The kept photos, best first. It is never empty: if nothing passes, the best photo is returned.
report JSON with one entry per input photo: index, score, sharpness, exposure, eyes_closed, faces, duplicate_of, keep and energy.
kept How many photos were kept (INT).

Tips from retouching

  • The score is 0.6 for sharpness and 0.4 for exposure, minus 0.4 if eyes are closed. Use report to see why a frame was dropped.
  • In report, duplicate_of shows which frame a near-identical shot was folded into. Only the best of each group is kept.
  • Photos in one batch must be the same size. Resize first if your shoot mixes sizes.
  • A blurred background with a sharp subject still counts as sharp. When faces are found, the face areas are measured instead.
  • Send selected on to the retouching chain, for example Auto Tone, so you only edit the keepers.
  • eyes_closed is null in the report when the eye check is off or MediaPipe is not available.
Under the hood

Every photo is measured on a 1024 px copy. Sharpness is Laplacian energy per tile of an 8 by 8 grid, taking the 90th percentile, so a sharp subject on a soft background still scores high. When faces are found (YuNet), the face boxes are measured instead. The value goes on a log scale and is compared with the sharpest shot of the same scene, found by a similar difference hash. A shot 10 times blurrier than that scores 0 for the relative part.

Exposure is penalised for clipped highlights, crushed shadows and an average far from mid-grey. Closed eyes use the eye aspect ratio (Soukupova and Cech, 2016) from MediaPipe landmarks, with 0.18 as the limit.

Duplicates are grouped with a 64-bit difference hash. Photos within 10 bits of each other form one group and only the best is kept. Source: core/cull.py, wrapper nodes/cull_node.py.

Example workflow

Cull Score example workflow builds a small batch from the sample portrait (sharp, blurred and blown-out versions) and shows the selected photo and the report. Copy workflows/sample_portrait.jpg into ComfyUI/input/ (Pexels License).