FUSE KSampler (Video)

The still-image FUSE KSampler works frame by frame, which is the whole problem with video. Run it on each frame independently and faces flicker between "fixed" and "not fixed," boxes jitter, and the result is unwatchable even though every individual frame looks fine - the KB's own upscaling doc calls that the definition of the video problem. FUSE KSampler (Video) exists to stop that. It takes a 4D image batch or a 5D video tensor, tracks each face across frames as one consistent identity, and processes each person as their own coherent video batch so the fix holds together over time.

You want this when you have generated video with mushy faces (LTX, Wan, Hunyuan - pick your model) and you want one node that does the face pass with temporal sense. It's also happy doing general-object enhancement if you point it at a non-face YOLO model.

How it works

The core loop is the FUSE machinery you already know - YOLO detect, crop, VAE-encode, sample, decode, blend back - but wrapped in a temporal layer. On a 5D input it runs detection per frame, then matches boxes across frames by IoU to assign each one a track id, so the same person stays the same person. Each track becomes its own video batch and is sampled as a unit, which is what gives you true temporal coherence instead of a per-frame gamble. Bounding boxes get temporally smoothed (so crops don't dance around the face) and you can tune all of it through the yolo_optionals input.

A few differences from the still version:

• cfg defaults to 3.5, not 8.0 - video models (especially the 5D ones) don't want SD-class CFG.
• model and vae should support video if your input is 5D. The node doesn't convert; it passes the tensor through to the sampler.
• yolo_detector accepts prefixes - face/ for face models, yolo/ for general ones - which is how one node handles both jobs.
• temporal_tracking (default on) is the kill switch. Turn it off and it behaves like the still version applied per frame, which you'll only want for debugging.

Inputs that matter

• images - 4D image batch or 5D video tensor.
• model / vae - video-capable for 5D input.
• positive / negative, seed, steps, cfg, sampler_name, scheduler - standard sampler controls, with the low CFG default noted above.
• denoise - 0.5 default; lower keeps the original actor's face, which matters more in video where you can't hide identity drift behind a single pretty frame.
• face_id (-1 = all faces), face_order, face_size (512–1536), face_padding, force_square - the face-crop family, same as the still node.
• yolo_detector / sam_segmenter / sam_model_type - detection and optional SAM masking.

Optional mask_optionals (blending/mask) and yolo_optionals (confidence + the tracking knobs like tracking_smooth_window) plug in as DICTs. Outputs: images and masks.

Installing it

Same pack, same story: FUSE Face Enhancer by WASasquatch, install via ComfyUI Manager or:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/face-upscaling-and-seamless-embedding
cd face-upscaling-and-seamless-embedding
pip install -r requirements.txt

Restart, then face YOLO models into ComfyUI/models/yolo-face/, general ones into models/yolo/, SAM checkpoints into models/sams/.

Where people get burned

• Still-image model on 5D input. If the model doesn't handle video tensors, sampling a 5D batch fails or silently degrades. Match the model to the input dimension.
• Tracking tuning. Faces that move fast can jump between track ids - raise tracking_iou_threshold a little, and widen tracking_smooth_window only if crops look jittery (it must stay odd, 1–11). Too wide and crops lag on motion.
• Expecting it to be cheap. This is the FUSE loop times the frame count. A 100-frame clip at face_size 512 is a real wait; drop to 512 and use FP16 YOLO (half_precision) before giving up on speed.