ComfyUI Node
SeedVR2 Framing Downscale
Measures the tallest face (SAM 3) as a fraction of the image height and picks SeedVR2 Resize's downscale_factor: close-up, medium or far, each with its own factor. A lower factor gives more skin texture but changes a small face, so a far shot is not downscaled. The thresholds are uncalibrated estimates.
SeedVR2 Framing Downscale
- image
- downscale_factor
- face_fraction
◄sam3_modelsam3.1_multiplex_fp16.safetensors►
◄close_up_min_face0.30►
◄close_up_factor0.50►
◄medium_min_face0.18►
◄medium_factor0.75►
◄far_factor1.00►
◄no_face_factor1.00►
◄detection_threshold0.50►
CategoryBCNodes/seedvr2
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | The original image(s), the batch SeedVR2 Resize gets. One factor for the whole batch, from its tallest face (Resize takes one factor). | |
| sam3_model | COMBO | sam3.1_multiplex_fp16.safetensors | SAM 3 checkpoint under models/checkpoints; the default is downloaded when missing. |
| close_up_min_face | FLOAT | 0.300.01–1 | Face height / image height from which the shot is a close-up. Uncalibrated default (an estimate, not yet measured against SeedVR2 output). |
| close_up_factor | FLOAT | 0.500.01–1 | downscale_factor for a close-up: a large face keeps its shape at a strong downscale and gains the most skin texture. |
| medium_min_face | FLOAT | 0.180.01–1 | Face height / image height from which the shot is medium; below it, far. Must be below close_up_min_face. Uncalibrated default (an estimate, not yet measured against SeedVR2 output). |
| medium_factor | FLOAT | 0.750.01–1 | downscale_factor for a medium shot. |
| far_factor | FLOAT | 1.000.01–1 | downscale_factor for a far shot. 1 = no downscale, so SeedVR2 does not change a small face. |
| no_face_factor | FLOAT | 1.000.01–1 | downscale_factor when no face is found (logged). |
| detection_threshold | FLOAT | 0.500–1 | SAM 3 detection threshold for the face. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| downscale_factor | FLOAT | Wire to SeedVR2 Resize (or SeedVR2 Preprocess (Compact)) downscale_factor. |
| face_fraction | FLOAT | Height of the tallest face box / image height, over the whole batch; 0 when no face is found. |