Temporal Face Detailer (SDXL video)

If you've ever run Impact Pack's FaceDetailer on every frame of a video, you already know why this node exists. Each frame gets a fresh img2img pass with fresh noise, the face sharpens up nicely - and then morphs, shimmers, or straight-up changes identity a few frames later. Per-frame detailing fixes faces and creates flicker in the same gesture.

Temporal Face Detailer is the video-grade version of that loop. It detects and tracks faces across the whole frame batch, re-details each tracked face with a fixed seed and stabilized crops, and blends the results with optical flow, so you get the detail pass without the strobe light. One node does the entire pipeline: detect → track → detail → smooth → paste back. It's SDXL-only - you feed it a normal checkpoint's MODEL, CLIP, and VAE, and frames from a video loader like VideoHelperSuite's Load Video (Upload).

How it works

The anti-flicker magic is layered, and worth knowing - every lever is a widget on this node:

1. Tracked, stabilized crops. Faces are linked into identity tracks frame-to-frame (greedy association on IoU + center distance, with a constant-velocity prediction). Each track's crop is centered on the facial-landmark centroid instead of the wobbly detection box (crop_anchor: landmarks), and the window is smoothed over time - the sampled region barely moves, so neither does the result.
2. Fixed per-track seed and noise. Every frame in a track is sampled from the same noise - the source literally replicates one noise tensor across the track's frames. Single biggest identity-stability lever, and on by default (noise_mode: fixed_per_track).
3. Moderate denoise. Default 0.35, and the README's 0.3–0.45 sweet spot is right. denoise_max adds adaptive denoise: steady frames stay locked at the base value while high-motion frames ramp up, where correction is most needed and flicker is masked by the motion anyway.
4. Flow-guided temporal blending in pixel space (flow_strength) and optionally latent space (latent_blend). Optical flow from the originals warps the running result onto each new frame, occlusion-aware so motion doesn't smear. Latent blending smooths in the VAE's semantic space and lets you raise denoise before flicker shows.
5. Color match + feathered paste-back. Each crop's exposure is pinned to its source (no brightness pulsing), and only the face region is composited back, so backgrounds stay pixel-exact.

The inputs that matter

Wire image, model, clip, vae, type a positive/negative prompt (the widgets are fine - positive/negative conditioning inputs override them if you prefer encode nodes), and hit go. Defaults are genuinely good for a first pass. The ones you'll actually touch:

• denoise - the main quality-vs-consistency knob.
• noise_mode - leave it on fixed_per_track.
• detector - insightface for realistic faces, a yolo:... anime model for stylized ones.
• reference_image - a clean face photo to anchor identity against (see below).
• chunk_size - lower it (default 4) if you OOM.
• detail_every - set 2–4 on long clips; it samples keyframes and flow-propagates the rest for a big speedup.

Outputs: image goes to Video Combine (the final video), face_masks to a preview/save, debug_overlay to a second Video Combine if you want to inspect tracking, and face_tracks to Face Track Preview or a second-pass detail.

Installing it

ComfyUI Manager - search "Temporal Face Detailer" - or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/nikythebikky/ComfyUI-Temporal-Face-Detailer
pip install -r ComfyUI-Temporal-Face-Detailer/requirements.txt

Then restart. The heavy lift is insightface (RetinaFace) plus onnxruntime - swap for onnxruntime-gpu to run detection on CUDA. InsightFace's models auto-download into models/insightface, shared with ReActor and friends; if it isn't installed, the nodes fall back to OpenCV YuNet, then to a Haar cascade.

Where people get burned

Anime faces. Realistic-face detectors miss or jitter on stylized characters, and unstable detection is a root cause of flicker no downstream blending can fix. Install ultralytics and drop face_yolov8m_anime.pt (same weights Impact Pack uses) into models/ultralytics/bbox/, then pick yolo:bbox/face_yolov8m_anime.pt. The README's recipe for WAN Animate post-pass: keep denoise low (~0.15), add latent_blend 0.4, anchor with a reference_image, use raft_small flow on fast motion.

Identity drift over a long clip. Temporal blending keeps frames consistent with each other, but the whole sequence can drift away from who it should be. Connect a reference face to reference_image - it nudges each frame's init latent toward the reference (scaled by reference_strength × denoise, so it won't ghost at low denoise) and color-anchors the face region. For strong identity, patch the model with IPAdapter FaceID upstream and feed the patched model in.

Turing GPUs. Tested on a 22 GB 2080 Ti, the pack explicitly warns: don't enable SageAttention on Turing - it silently produces NaN/black frames. fp16/bf16 only, no fp8 needed.

Flicker you can't kill. Check debug_overlay first - if the boxes themselves flicker or drop out, no smoothing fixes the root cause. Switch detectors before you touch denoise.