ComfyUI Node
Temporal Face Detailer (SDXL video)
A ComfyUI node in TemporalFaceDetailer with 45 inputs and 4 outputs.
Temporal Face Detailer (SDXL video)
- image
- model
- clip
- vae
- positive
- negative
- lora_stack
- reference_image
- image
- face_masks
- debug_overlay
- face_tracks
◄positive_textdetailed face, sharp eyes, high quality skin texture►
◄negative_textblurry, deformed, low quality►
◄detectorinsightface►
◄det_threshold0.50►
◄min_face_size24►
◄max_faces4►
◄detector_devicecuda►
◄iou_threshold0.25►
◄max_track_gap10►
◄min_track_length2►
◄crop_factor1.7►
◄crop_smoothing0.80►
◄crop_anchorlandmarks►
◄align_rotationfalse►
◄guide_size768►
◄max_size1024►
◄seed0►
◄steps20►
◄cfg7.0►
◄sampler_name▾►
◄scheduler▾►
◄denoise0.35►
◄denoise_max0.00►
◄noise_modefixed_per_track►
◄detail_modeimg2img►
◄mask_dilation8►
◄feather15►
◄temporal_strength1.00►
◄flow_strength0.50►
◄latent_blend0.00►
◄flow_bidirectionaltrue►
◄flow_backendfarneback►
◄color_match0.50►
◄chunk_size4►
◄detail_every1►
◄track_prompts►
◄reference_strength0.35►
CategoryTemporalFaceDetailer
Inputs (45)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | video frame batch | |
| model | MODEL | — | |
| clip | CLIP | — | |
| vae | VAE | — | |
| positive_text | STRING | detailed face, sharp eyes, high quality skin texture | — |
| negative_text | STRING | blurry, deformed, low quality | — |
| detector | COMBO | insightface | yolo:* entries are ultralytics models found in models/ultralytics — use an anime face model (e.g. bbox/face_yolov8m_anime.pt) for stylized characters |
| det_threshold | FLOAT | 0.500.05–1 | — |
| min_face_size | INT | 248–1024 | — |
| max_faces | INT | 41–16 | — |
| detector_device | COMBO | cuda | 2 options: cuda, cpu |
| iou_threshold | FLOAT | 0.250.05–0.95 | min overlap to link a detection to a track |
| max_track_gap | INT | 100–120 | frames a face may vanish (occlusion) before its track ends; gaps are interpolated |
| min_track_length | INT | 21–120 | — |
| crop_factor | FLOAT | 1.71–4 | — |
| crop_smoothing | FLOAT | 0.800–1 | temporal smoothing of the crop window (anti-jitter) |
| crop_anchor | COMBO | landmarks | landmarks centers the crop on the facial-landmark centroid (far more stable than the detection box); falls back to bbox when the detector yields no landmarks |
| align_rotation | BOOLEAN | false | rotation-register crops so the eye line is horizontal in every sampled crop (similarity transform, inverse-warped on paste-back); needs a landmark-capable detector |
| guide_size | INT | 768256–2048 | resolution faces are resampled at |
| max_size | INT | 1024256–2048 | — |
| seed | INT | 00–18446744073709550000 | — |
| steps | INT | 201–100 | — |
| cfg | FLOAT | 7.00–30 | — |
| sampler_name | COMBO | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 | |
| scheduler | COMBO | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 | |
| denoise | FLOAT | 0.350–1 | main quality/consistency lever: higher = more detail but more flicker (0.3-0.45 recommended) |
| denoise_max | FLOAT | 0.000–1 | adaptive denoise: when above 'denoise', steady frames keep the low base value while high-motion frames ramp toward this. 0 = off (single global denoise) |
| noise_mode | COMBO | fixed_per_track | fixed_per_track reuses the same seed and noise on every frame of a track — the biggest identity-stability lever |
| detail_mode | COMBO | img2img | inpaint restricts sampling to the face mask in latent space; img2img resamples the whole crop (mask still limits the paste-back) |
| mask_dilation | INT | 80–128 | — |
| feather | INT | 150–128 | — |
| temporal_strength | FLOAT | 1.000–1 | master anti-flicker strength; scales flow blend + color match |
| flow_strength | FLOAT | 0.500–1 | optical-flow-guided temporal blend of detailed crops (pixel space) |
| latent_blend | FLOAT | 0.000–1 | flow-guided temporal blend in LATENT space before decode — smooths in the VAE's semantic space, letting you raise denoise with less flicker. Try 0.3-0.5 with denoise 0.25+; 0 = off |
| flow_bidirectional | BOOLEAN | true | — |
| flow_backend | COMBO | farneback | raft_small/raft_large (torchvision) give much cleaner flow on fast motion at some VRAM/time cost; auto-falls back to farneback on any failure |
| color_match | FLOAT | 0.500–1 | match each detailed crop's color to its source (stops brightness pulsing) |
| chunk_size | INT | 41–64 | crops sampled per batch; lower if you hit OOM |
| detail_every | INT | 11–30 | keyframe mode: sample every Nth frame, flow-propagate the rest (speed on long clips) |
| positiveopt | CONDITIONING | overrides positive_text | |
| negativeopt | CONDITIONING | overrides negative_text | |
| lora_stackopt | LORA_STACK | — | |
| track_promptsopt | STRING | per-track positive prompt overrides, one per line: 'track_id: prompt' (see the debug overlay for IDs) | |
| reference_imageopt | IMAGE | identity anchor: a reference face image the detailed faces are biased toward (init-latent nudge + color anchoring). For strong identity conditioning also patch the MODEL with IPAdapter FaceID upstream — crops are sampled with whatever model you feed in, so it composes | |
| reference_strengthopt | FLOAT | 0.350–1 | how hard to pull toward reference_image (no effect unless it is connected) |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |
| face_masks | MASK | — |
| debug_overlay | IMAGE | — |
| face_tracks | FACE_TRACKS | — |