DMC Face Detailer · Ordered Characters

A face detailer fixes smeared faces by cropping each one, re-rendering it bigger, and pasting it back. The catch nobody warns you about: every crop gets re-rendered with the same prompt, so a group shot comes back as a row of siblings. Hence the perennial "room full of clones" complaint.

DMCFaceDetailer is a specific fix for that. Give it a JSON list of characters - name, position, face-only prompt - and each detected face gets re-rendered with its own prompt. Same seed, same checkpoint, two genuinely different people.

How the routing actually works

It wraps the pack's own DonutFaceDetailer, so it inherits the whole Impact-style loop: detect boxes, optionally tighten them with SAM, crop, scale toward a target pixel budget, run a fresh masked sampling pass, paste back.

The wrapper adds two things. It detects faces once per image, then has to prove a matching between those faces and the characters you described - it doesn't guess. Then it runs the refinement pass one face at a time, fresh-encoding that character's face_prompt and swapping it in for the frame-level positive on that pass only. Your negative stays shared.

Ordering is the clever bit. Give positions as left/center/right or ordinals 1st–16th and it sorts by horizontal center - but only after checking each gap is wider than a quarter of the smaller face's width. Faces at nearly the same x get a refusal, not a coin flip. For awkward layouts, position can be an object with left_of and above arrays naming other characters, solved as a constraint problem.

So it won't silently process the wrong face: get the count wrong and you get Face routing mismatch: expected 2 characters but detected 2 faces, with no character prompts applied.

Note that the per-character prompt is the positive for that pass - not merged with your frame prompt, so keep that one about the scene.

The inputs worth touching

face_prompts_json is the point of the node, and it's a bare multiline text box - no rows, no picker. It takes 2 to 16 objects, each with exactly character, position and face_prompt:

[
  {"character": "Ada", "position": "left", "face_prompt": "Ada, green eyes, freckles, faint smile"},
  {"character": "Bo", "position": "right", "face_prompt": "Bo, thick beard, weathered skin, scowling"}
]

Duplicate names, a missing key, an empty prompt, or quotes pasted from a formatted doc are all rejected before anything samples. A single entry is an error too - use the plain DonutFaceDetailer for one face.

max_faces (default 2) caps how long the list may be; describe three characters with it at 2 and the node refuses rather than truncating, because dropping a face would shift everybody's identity. resolution (default 1024) is the pixel budget per crop's sampling canvas - the real cost dial - with max_resolution clamping the long edge. bbox_threshold (0.5) is what you lower when a face in a crowd isn't caught, and bbox_crop_factor (3) is how much context rides along. Flick on vary_seed_per_face for crowds.

Outputs: image is the frame with refined faces pasted in - wire that onward. mask is the combined mask of the matched faces. cropped_refined and cropped_enhanced_alpha are lists of per-face crops: fine for a preview grid, awkward anywhere expecting a batch. detailer_pipe feeds Impact's pipe nodes; cnet_images only fills if a hook emits control images.

edit_mode, face_reference and grounding_px do Krea 2 identity-edit passes, but only in the single-face path - with multi-character routing you get Multi-face identity editing and identity references are not supported.

Installing it

Easiest route is ComfyUI Manager: search DonutNodes, install or update, restart. The README is blunt that your installed package has to actually contain the node you're looking at - the "Donut Workflow V5" release name and the package version are different things. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/DonutsDelivery/ComfyUI-DonutNodes.git donutnodes
cd donutnodes
python -m pip install -r requirements.txt

Use the interpreter that launches ComfyUI. Requirements are light (opencv-python-headless, scipy, matplotlib, psutil, tqdm, requests), and the file warns against pinning NumPy globally. Only one OpenCV variant should be present.

What it genuinely needs on top is the Impact Pack - resolved at execution time, and it throws DonutFaceDetailer requires a working ComfyUI Impact Pack if missing. You also need a detector in bbox_detector: Ultralytics plus a face_yolov8-family model, typically. SAM is optional; sam_model_opt just tightens the masks.

Where people get burned

The count has to match. A profile shot or a hand across a face gives you a routing mismatch instead of an image. Drop bbox_threshold, nudge bbox_dilation up, and check your detector before blaming the JSON.

It's slower than it looks. Each character is its own sampling pass per image at resolution squared - four people across four images is sixteen passes. Drop resolution to 768 for group shots; the crops are already upscaled.

Don't pair it with wildcard or detailer_hook in multi-character mode. Both can replace the conditioning the routing just built, so the node rejects them rather than let a wildcard override your per-character prompts.

Mind your CFG. On a distilled Krea 2 Turbo at CFG 1 the negative is inert - true of every guidance-distilled model, not a bug here. Write constraints positively instead, or enable nag_enabled (needs the separate krea2-nag pack).

It routes, it doesn't recognize. Nothing checks that the face it refined belongs to the person you described - staggered faces can flip a left/right ordering and put correct prompts on the wrong people. Use the relative-position form there, and eyeball one render before queueing fifty.