ComfyUI Node
BSAI Face Refine (one-click HD face repair)
One-click repair of deformed/broken/blurry faces in video; single and multi-person handled automatically, outputs high-definition faces.
BSAI Face Refine (one-click HD face repair)
- images
- model
- clip
- vae
- audio_vae
- audio
- identity_ref_1
- identity_ref_2
- identity_ref_3
- identity_ref_4
- images
- report
- face_count
◄detectorface_yolov8m.pt►
◄steps8►
◄denoise0.35►
◄canvas_size768►
◄small_face_denoise1.00►
◄confidence0.35►
◄crop_factor2.5►
◄smooth_window21►
◄schedulerbeta►
◄sampler_nameeuler►
◄seed0►
◄promptultra high definition close-up of a human face, crystal clear facial details, natural realistic skin texture, sharp detailed eyes, correct natural facial features, photorealistic, professional portrait quality►
◄enable_audio_locktrue►
◄identity_threshold0.28►
◄paste_regionface_only►
◄mask_dilation24►
◄feather24►
◄colour_match1.00►
◄blend1.00►
◄stage2_enablefalse►
◄stage2_engine1xSkinContrast (skin detail, fastest)►
◄stage2_scale2.0►
◄stage2_cfg0.5►
◄stage2_steps1►
◄stage2_fidelity0.60►
◄stage2_blend0.70►
◄stage2_temporal0.50►
◄undetected_framesfade_out►
◄person_fallbackfalse►
◄use_npu_face_detecttrue►
CategoryBSAI/H3 Face Refine
Inputs (40)
| Name | Type | Default | Description |
|---|---|---|---|
| images | IMAGE | Source video frames (any length; full frame rate recommended) | |
| model | MODEL | MiniMax H3 video model (TURBO 4/8-step model recommended) | |
| clip | CLIP | H3 text encoder (Qwen3VL-MiniMax-H3) | |
| vae | VAE | H3 video VAE | |
| audio_vae | VAE | H3 audio VAE (required when audio lock is enabled) | |
| detector | COMBO | face_yolov8m.pt | Face detection model picked from models/ultralytics/bbox. Only files registered with folder_paths are selectable; free-text paths are not accepted. |
| steps | INT | 81–32 | H3 redraw sampling steps (reference workflow uses 8) |
| denoise | FLOAT | 0.350–1 | Face redraw strength applied to large-face frames. Small-face frames automatically use full strength (reference workflow: large face 0.35, small face 1.0). Raise to 0.5-0.55 for badly broken faces. |
| canvas_size | INT | 768512–2048 | Square face-crop canvas size. 768 is H3's native recommendation; larger is sharper but uses more VRAM |
| small_face_denoise | FLOAT | 1.000–1 | Redraw strength for small-face frames (reference workflow 1.0). Smoothly interpolated by face size between frames |
| confidence | FLOAT | 0.350.05–1 | — |
| crop_factor | FLOAT | 2.51.2–6 | Expansion factor around the face box to give the redraw context |
| smooth_window | INT | 211–121 | Temporal smoothing window for crop position |
| scheduler | COMBO | beta | 5 options: beta, simple, karras, sgm_uniform, exponential |
| sampler_name | COMBO | euler | 6 options: euler, euler_ancestral, dpmpp_2m, dpmpp_sde, uni_pc, heun |
| seed | INT | 00–9007199254740991 | — |
| prompt | STRING | ultra high definition close-up of a human face, crystal clear facial details, natural realistic skin texture, sharp detailed eyes, correct natural facial features, photorealistic, professional portrait quality | Redraw prompt. Default is a generic HD portrait enhancement prompt; override as needed |
| audioopt | AUDIO | Original video audio (used for lip/audio lock when enabled, keeps audio in sync) | |
| identity_ref_1opt | IMAGE | Identity reference for person 1 (leave blank for single-person auto-largest-face tracking) | |
| identity_ref_2opt | IMAGE | Identity reference for person 2 (multi-person) | |
| identity_ref_3opt | IMAGE | Identity reference for person 3 (multi-person) | |
| identity_ref_4opt | IMAGE | Identity reference for person 4 (multi-person) | |
| enable_audio_lockopt | BOOLEAN | true | Lock lip sync to the original video audio (requires audio_vae + audio) |
| identity_thresholdopt | FLOAT | 0.280–1 | — |
| paste_regionopt | COMBO | face_only | 3 options: face_only, face_ellipse, full_crop |
| mask_dilationopt | INT | 240–256 | — |
| featheropt | INT | 240–256 | — |
| colour_matchopt | FLOAT | 1.000–1 | — |
| blendopt | FLOAT | 1.000–1 | — |
| stage2_enableopt | BOOLEAN | false | Enable stage-2 detail enhancement (reuses the local BSAI-H3-upscale-4K engine; corresponds to the VOSR2 step in the reference video) |
| stage2_engineopt | COMBO | 1xSkinContrast (skin detail, fastest) | Stage-2 engine: fast tier (1xSkinContrast skin detail / RealESRGAN / DLSS5 / FlashVSR / SeedVR2-INT8), extreme tier (VOSR generative detail), or face restoration (CodeFormer / GFPGAN) |
| stage2_scaleopt | FLOAT | 2.01–4 | Stage-2 sampling scale. 1x engines (1xSkinContrast / 1xDeJPG) are pinned to 1x repair-only; other engines generate at this scale and the stitch layer maps back to original coordinates. VOSR needs at least 2x |
| stage2_cfgopt | FLOAT | 0.5-2–2 | VOSR / SeedVR2 fidelity (higher = closer to the original face) |
| stage2_stepsopt | INT | 11–25 | VOSR inference steps (1 = fastest single step; raise to 2-4 if texture is unsatisfactory) |
| stage2_fidelityopt | FLOAT | 0.600–1 | CodeFormer fidelity (CodeFormer engine only) |
| stage2_blendopt | FLOAT | 0.700.1–1 | CodeFormer blend strength (CodeFormer engine only) |
| stage2_temporalopt | FLOAT | 0.500–1 | Face temporal stability (CodeFormer engine only) |
| undetected_framesopt | COMBO | fade_out | 3 options: fade_out, skip, composite_anyway |
| person_fallbackopt | BOOLEAN | false | Wide-shot small-face fallback: only on frames where the face detector completely loses the face, use the person whole-body segmentation model to back-derive the head from the top of the body box. Note: it participates in trajectory smoothing, so in medium/night shots a low estimate can drag the crop box onto the neck and cause noise. Off by default; enable only when you confirm pure wide-shot small faces with frequent lost face boxes. Requires models/ultralytics/segm/person_yolov8m-seg.pt. |
| use_npu_face_detectopt | BOOLEAN | true | Enable NPU face detection: auto-starts the NPU service (8191); face detection / landmarks / age recognition run on the Intel AI Boost NPU and do not use the GPU |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| images | IMAGE | — |
| report | STRING | — |
| face_count | INT | — |