ComfyUI Node
🐇 Autocrop to Loop
Auto-crops the clip to create a smoother loop: tests crops from the end and scores the seam so it feels like a normal step.
🐇 Autocrop to Loop
- clip_frames
- cropped_clip
- end_crop_frames
- cropped_length
- score
- diagnostics_csv
◄max_end_crop_frames12►
◄include_first_steptrue►
◄include_last_steptrue►
◄include_global_median_stepfalse►
◄seam_window_frames2►
◄distance_metricL1►
◄score_in_8bitfalse►
◄use_ssim_similaritytrue►
◄use_exposure_guardtrue►
◄use_flow_guardfalse►
◄weight_step_size0.55►
◄weight_similarity0.30►
◄weight_exposure0.10►
◄weight_flow0.05►
◄ssim_downsample_scales1,2►
◄accelerate_with_gputrue►
◄use_mixed_precisiontrue►
Categoryvideo utils
Inputs (18)
| Name | Type | Default | Description |
|---|---|---|---|
| clip_frames | IMAGE | Your full clip (NHWC, 0–1). Tries every crop from 0..max_end_crop_frames and returns the best loop. | |
| max_end_crop_frames | INT | 120–10000 | Largest crop to test at the END. Higher = more candidates (slower), but potentially better. |
| include_first_step | BOOLEAN | true | Use the first neighbor pair (frame 0→1) as a target step size/similarity. |
| include_last_step | BOOLEAN | true | Use the last neighbor pair inside the KEPT region as a target. |
| include_global_median_step | BOOLEAN | false | Also use the median step across the KEPT region (needs ≥3 frames). Helps ignore outliers. |
| seam_window_frames | INT | 21–6 | Average over multiple aligned pairs across the seam. Larger = more robust. |
| distance_metric | COMBO | L1 | How to measure step size for matching. L1 is usually more forgiving; MSE penalizes big errors more. |
| score_in_8bit | BOOLEAN | false | Score with an 8-bit view (simulate export). Output video still stays float. |
| use_ssim_similarity | BOOLEAN | true | Include SSIM so the seam ‘looks’ like a normal neighbor—avoid freeze or jump. |
| use_exposure_guard | BOOLEAN | true | Promote smooth brightness across the seam (reduces flicker pops). |
| use_flow_guard | BOOLEAN | false | Encourage consistent motion across the seam (needs OpenCV; slower). |
| weight_step_size | FLOAT | 0.550–1 | Importance of matching step size. Higher = less freeze/jump risk. |
| weight_similarity | FLOAT | 0.300–1 | Importance of visual similarity (SSIM). Helps avoid a frozen-looking seam. |
| weight_exposure | FLOAT | 0.100–1 | Importance of even brightness across the seam. |
| weight_flow | FLOAT | 0.050–1 | Importance of motion continuity across the seam. |
| ssim_downsample_scales | STRING | 1,2 | SSIM scales to average, as a comma list. Example: 1,2 = full-res and half-res. |
| accelerate_with_gpu | BOOLEAN | true | If ON and CUDA is available, run scoring on GPU for a big speedup (same results). |
| use_mixed_precision | BOOLEAN | true | If ON (with GPU), use mixed precision for SSIM/conv math (faster on larger clips). |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| cropped_clip | IMAGE | — |
| end_crop_frames | INT | — |
| cropped_length | INT | — |
| score | FLOAT | — |
| diagnostics_csv | STRING | — |