H3 Load Video + Face Select

MiniMax H3 is genuinely good at video, and genuinely bad at faces that take up a small slice of the frame. That's not a resolution problem - a 720p clip still has the same weak head-size-to-frame ratio - so no upscaler fixes it. The ComfyUI-H3-FaceRefine pack takes the FaceDetailer trick from Impact Pack (detect, crop, re-render at real scale, paste back) and stretches it across every frame. H3 Load Video + Face Select is the front end of that: it loads your video, finds every face and every hard cut in one pass, and decides which face is the subject before any sampling happens.

The pitch is "one pass, one decision." This node owns detection when it's wired in. It scans the clip once, hands its detected boxes and shot boundaries to H3 Face Track + Crop through the face_pick output, and that tracker then greys out its own detection settings because the work is already done. The tracker never decodes or detects the same video twice.

The select input is where you say who the subject is. The default is manual: click the Pick faces button on the node, and it shows one frame per shot with every face outlined and numbered. Choose one per shot and press Use these - the answer lands in confirmed_pick as comma-separated indices like 0,3,4, one per shot, saved with the workflow. There's a reason it asks per shot rather than once: at a hard cut every face gets renumbered, so the person who was face 1 before the cut may be face 3 after. That's also how you tell the graph that two differently-numbered faces are the same person.

Don't want to click? select also takes identity_reference (wire in a portrait and each shot picks the matching face), or a ranking rule: largest_face, centre_most, closest_to_xy and friends, with select_index taking the nth face out of that ranking. Two details worth knowing: the subject is chosen once per shot, then continuity follows it - it isn't re-ranked every frame, so a subject who crosses the frame past someone else is still tracked. And manual will refuse to run until a pick exists, failing fast rather than silently refining whoever ranks first.

Set cut_detection to auto (pyscenedetect) on clips with hard cuts, so the subject is chosen per shot and the tracker doesn't smooth the crop across a join. That's the one dependency this node adds over the pack's others: scenedetect>=0.7, installed automatically with requirements.txt along with ultralytics, scipy and insightface.

The rest of the inputs are trim: video accepts a pasted path (read in place, which is what you want for a big clip) or a Browse-picked file copied into ComfyUI's input folder; confidence is the detector's score floor, where lower catches profiles and small faces at the cost of false positives; skip_first_frames and frame_load_cap trim the clip (the audio is cut to match so lipsync stays aligned). X/Y/frame_index only appear with closest_to_xy - the Preview coordinates button shows you where your numbers land before you render anything.

Outputs worth wiring: images and audio go where your old video loader went (the node replaces it, it doesn't sit beside it); frame_count feeds the H3 node's length so it follows the clip; face_pick goes to the tracker; preview to a PreviewImage shows one card per shot with faces numbered. The report output tells you how many shots were found and which frame each pick locked onto - glance at it, because a false cut costs smoothing on both sides of it.

Install: ComfyUI Manager (search "ComfyUI-H3-FaceRefine") or

cd ComfyUI/custom_nodes
git clone https://github.com/Carasibana/ComfyUI-H3-FaceRefine.git

then restart ComfyUI. You must supply a face detector yourself - face_yolov8m.pt from Bingsu/adetailer into models/ultralytics/bbox/ - and the MiniMax H3 model, VAE and Qwen3-VL text encoder if you're not already generating H3 video. The Manual Select example workflow additionally needs VideoHelperSuite (to save) and ComfyUI-H3-NativeAudioLock (for lipsync). One trap the README calls out: don't install onnxruntime-gpu alongside the CPU onnxruntime - the CPU build shadows it, and identity matching silently drops to CPU. On anime or stylized footage, swap in an anime face detector and set identity_model to clip_vision (or ccip) instead of the insightface default, which is trained on photographed faces.

For a niche of a niche, the README is unusually honest about the pain points, and most of them live in this node. confirmed_pick is dropped automatically if you change the video, detector, frame range or cut settings - annoying until you realize those changes renumber the faces your picks described. And remember H3 wants frame counts on its 17k+5 grid, so let frame_load_cap land on one of those if you're trimming. Wire the preview up, watch the report, and you'll know exactly who's getting the facelift.