ComfyUI Node
H3 LongTake FaceRefine (small faces)
Regenerates small faces clip by clip with ComfyUI-H3-FaceRefine and writes <project>/fr; assemble with H3 LongTake Stitch. Skips the clips whose face is already large.
H3 LongTake FaceRefine (small faces)
- model
- clip
- vae
- face_image
- last_clip
- project_dir
- report
◄project_namelongtake►
◄description►
◄denoise0.25►
◄steps4►
◄seed0►
◄max_face_px150►
◄modecontinue►
◄project_dir—►
◄large_face_multiplier0.35►
◄smooth_window11►
◄crop_factor2.5►
◄feather24►
◄chunk_crf10►
◄seam_blend_frames8►
CategoryH3 LongTake
Inputs (18)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | H3 ref2va with the Turbo (8-step v1.0 recommended): the ref2va is needed because the face goes in as <Picture 1>, even if the project was made with the fl2va. | |
| clip | CLIP | — | |
| vae | VAE | — | |
| face_image | IMAGE | A close-up of the face, not the whole photo. | |
| project_name | STRING | longtake | An already rendered project. The result goes to <project>/fr (in the Stitch: project_name = 'name/fr'). |
| description | STRING | The subject's look (hair, outfit): helps not to change person. | |
| denoise | FLOAT | 0.250.05–1 | 0.25-0.35. With H3's shift 12, 0.25 already rewrites the face; above it the head drifts against the body. |
| steps | INT | 41–50 | — |
| seed | INT | 00–18446744073709550000 | — |
| max_face_px | INT | 1500–2000 | Clips whose median face is taller than this (px) are copied without FaceRefine. 0 = refine every clip. |
| mode | COMBO | continue | continue: skips the clips already done. |
| project_diropt | STRING | Connect project_dir from the Render/I2V/Refine. | |
| large_face_multiplieropt | FLOAT | 0.350–1 | Fraction of the denoise on large faces (the author's default). |
| smooth_windowopt | INT | 111–201 | Box smoothing: lower if the camera moves. |
| crop_factoropt | FLOAT | 2.51.5–4 | — |
| featheropt | INT | 240–128 | — |
| chunk_crfopt | INT | 100–30 | — |
| seam_blend_framesopt | INT | 80–48 | At the inner joins the refined clip fades to the original over these frames, so the face does not pop when going from a copied (or separately refined) clip to a refined one. 0 = off. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| last_clip | IMAGE | — |
| project_dir | STRING | — |
| report | STRING | — |