CoachBate H3 Subject Stitch Back

Paste is where refine workflows die

Detecting the subject and re-rendering it big is the easy part. Putting it back is where every face-detailer workflow either looks like nothing happened or looks like a sticker. CoachBate H3 Subject Stitch Back is the paste, tuned specifically for the awkward case: the crop you paste was generated at 768 and lands on a box maybe 200 pixels wide, so the composite is a downscale, and it happens through a mask whose subject changes size every frame.

It takes base_images (the original clip), refined_crops (the decoded H3 output of the crops), and the transform from CoachBate H3 Subject Track + Crop, and returns images. That's the loop: track, inject, denoise, sample, decode, mask, stitch.

The settings that decide whether it looks real

Only one output, so all the decisions live in the inputs.

paste_region is what actually gets composited: subject_only or subject_ellipse pastes just the detected subject box - the wider crop exists to give the sampler context, not to be pasted - while full_crop pastes the whole crop including surroundings and background, which risks a visible rectangle if H3 altered them.

feather is where a lot of people go wrong, and the tooltip explains why at length. It's a Gaussian blur radius in source pixels, measured against the final frame rather than the canvas. Canvas-relative feather is a trap: a 75px crop blown up to a 512 canvas makes a 40px canvas feather only ~6 source pixels, while a 720px crop makes it ~56. The blend ends up tightest exactly where the subject is smallest and you need the most help - which reads as a hard edge appearing as the shot zooms out. There's an optional feather_scales_with_crop that restores the old canvas-relative behaviour; leave it off.

resample defaults to bicubic_antialias, and this is not a detail. The paste is a downscale, and plain bilinear with no antialiasing softens and aliases exactly the detail the pass just added. The antialiased path shrinks the crop to the box size with an antialiased bicubic filter first, then warps it. bilinear is the original behaviour and exists because forks change defaults carefully.

colour_match (default 1.0) matches the refined crop's per-channel mean and standard deviation to the region it replaces. The crop and the full frame went through independent passes, so without this the subject comes back subtly brighter or differently tinted, and your eye reads that as "pasted" even when the detail is perfect. blend is a plain global opacity on the refined subject - knock it to 0.7 or 0.8 if the pass came back over-sharpened.

mask_dilation (default 16) grows the box or mask before blurring, in canvas pixels, so the blur has room and doesn't eat into the subject. Impact Pack dilates for the same reason.

Frames where the subject isn't there

undetected_frames is the one people find surprising, because all frames go through H3 either way - that's what keeps the clip temporally consistent. This only controls what gets pasted. fade_out ramps the composite to zero across the gap (smooth, no pop, and the default), skip hard-cuts so those frames keep their original pixels exactly, and composite_anyway pastes regardless - which risks pasting a subject H3 hallucinated where there wasn't one.

The optional masks input overrides paste_region entirely. Feed it from CoachBate H3 Segm Mask (YOLO) and the blend falls on the actual silhouette instead of an arbitrary rectangle - this is the bbox-vs-polygon lesson from face detailing, applied to video. Pair a silhouette with a small feather (2–8); a rectangle needs much more.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/CoachBate/ComfyUI-CoachBate.git

Or ComfyUI Manager, search "CoachBate". Restart, then look under CoachBate / H3 Subject Refine. No pip dependencies of its own - but the pipeline around it needs ultralytics for detection, the Video Helper Suite, and, if you use the blueprints, ComfyUI-Pixaroma's Loop Start / Loop End.

Where people get burned

• A sharp rectangle appears as the camera pulls back. Canvas-relative feather, most likely via feather_scales_with_crop, or a feather value that was tuned for a fixed-size subject. Feather is in source pixels precisely so it survives magnification changes.
• The subject looks slightly wrong-tinted. colour_match is at 0, or the base images and the crops were colour-managed differently.
• The subject pops in and out of existence. undetected_frames at skip on a clip with detection dropouts. fade_out exists for this.
• Soft, mushy result. resample set to bilinear.