Faces are where AI images most often go wrong. In a full-body shot or a group photo, each face only gets a small patch of pixels, so you end up with blurry eyes, smudged teeth or slightly melted features, even when the rest of the image looks great.
The fix in ComfyUI is FaceDetailer from the Impact Pack. It finds every face, crops it, redraws it at high resolution, and pastes it back seamlessly. It is the ComfyUI version of ADetailer from Automatic1111 and Forge, and once it is in your workflow you will rarely ship a bad face again.
This guide covers installing the Impact Pack and Subpack, downloading the right detection models, wiring FaceDetailer into a workflow, what every important setting does, recipes for common problems, using it for hands and people, and the errors people hit most.
Node names and defaults are checked against the Impact Pack source code as of November 2026.
- Install two packs from ComfyUI Manager: ComfyUI Impact Pack and ComfyUI Impact Subpack. Since Impact Pack V8.0 the Subpack (which holds
UltralyticsDetectorProvider) is no longer installed automatically. - Download a face model such as
face_yolov8m.ptintoComfyUI/models/ultralytics/bbox/. - Wire it: your image + model + CLIP + VAE + prompts > FaceDetailer, with UltralyticsDetectorProvider feeding
bbox_detector. - Key settings: denoise 0.35 to 0.5, guide_size 512 (768 for SDXL/Flux), bbox_threshold 0.5, feather 5 or more.
- Faces look pasted on? Lower denoise, raise feather and bbox_crop_factor.
- Same node, other parts: swap in
hand_yolov8n.ptfor hands or a person model for whole bodies.
How FaceDetailer Works
The Impact Pack README describes FaceDetailer as a combination of a detector node for finding faces and a detailer node for enhancing them. Under the hood it does this for each face:
- Detect: a YOLO model draws a box around every face it finds.
- Crop: it cuts out the face plus some surrounding area for context.
- Upscale: the crop is enlarged so the face gets plenty of pixels (controlled by guide_size).
- Redraw: it runs an img2img pass on the crop using your model and prompt, at a partial denoise.
- Paste back: the new face is scaled down and blended into the original with a feathered mask.
Because step 4 is really img2img, everything about denoise from our ComfyUI img2img guide applies here too: lower values keep the original face, higher values redraw it more.
Install Impact Pack and Subpack
With ComfyUI Manager (recommended)
- Open Manager > Custom Nodes Manager.
- Search ComfyUI Impact Pack and click Install.
- Search ComfyUI Impact Subpack and click Install.
- Restart ComfyUI and refresh the browser.
- Double-click the canvas and search
FaceDetailerandUltralyticsDetectorProviderto confirm both appear.
Manual install
cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack
git clone https://github.com/ltdrdata/ComfyUI-Impact-Subpack
cd ComfyUI-Impact-Pack
pip install -r requirements.txt
cd ../ComfyUI-Impact-Subpack
pip install -r requirements.txt
Run pip inside the same Python that runs ComfyUI. For the Windows portable build, the Subpack README says to use <installed path>\python_embeded\python.exe -m pip instead of plain pip. For a venv, activate it first.
Not sure which other node packs are worth installing? See our list of must-have ComfyUI custom nodes.
Download Detection Models
The detector needs a YOLO model. The Subpack README points to the Bingsu/adetailer repository on Hugging Face for face, hand and person models, and many can be downloaded by searching “ultralytics” in ComfyUI Manager’s Model Manager.
| Model | Type | Folder | Use it for |
|---|---|---|---|
face_yolov8n.pt |
bbox | models/ultralytics/bbox/ |
Fastest face detection |
face_yolov8s.pt / face_yolov8m.pt |
bbox | models/ultralytics/bbox/ |
More accurate face detection (good default) |
hand_yolov8n.pt |
bbox | models/ultralytics/bbox/ |
Fixing hands |
person_yolov8n-seg.pt / person_yolov8s-seg.pt |
segm | models/ultralytics/segm/ |
Whole people, and silhouette masks |
sam_vit_b_01ec64.pth |
SAM | models/sams/ |
Optional tighter masks (auto-downloaded at install) |
Bbox models go in bbox, segmentation models (names ending in -seg) go in segm. Putting a model in the wrong folder is one of the most common reasons it does not show up. Our ComfyUI model folders cheat sheet lists every other folder.
Safety note: .pt files can contain code. Download detection models only from the sources above and avoid random re-uploads.
Build a FaceDetailer Workflow
Start with any working text-to-image workflow, then add these nodes after VAE Decode:
- Add
UltralyticsDetectorProviderand pickbbox/face_yolov8m.pt. - Add
FaceDetailer. - Connect the
VAE DecodeIMAGE to FaceDetailerimage. - Connect the same MODEL, CLIP and VAE you used for generation, plus your positive and negative CONDITIONING.
- Connect the provider’s
BBOX_DETECTORto FaceDetailerbbox_detector. - Optional: add
SAMLoader (Impact)and connect it tosam_model_optfor cleaner face masks. - Add a
Save Imagenode on FaceDetailer’simageoutput, and aPreview Imageoncropped_refinedto see each fixed face. - Run the workflow.
You can also run FaceDetailer on an existing picture: use Load Image instead of a generation chain, and load a checkpoint just for the detailing pass.
The mask output shows exactly which areas were detected and redrawn, which is handy when a face gets missed.
FaceDetailer Settings Explained
| Setting | Default | What it does | Tip |
|---|---|---|---|
guide_size |
512 | Target size the face crop is upscaled to before redrawing | 512 for SD 1.5, 768 to 1024 for SDXL and Flux |
guide_size_for |
bbox | Whether guide_size applies to the face box or the whole crop region | Leave on bbox |
max_size |
1024 | Upper limit on the upscaled crop | Raise to 1536 for big models if VRAM allows |
steps, cfg, sampler_name, scheduler |
20, 8.0 | Same as KSampler, for the face pass | Match your main generation settings (Flux usually needs cfg 1) |
denoise |
0.5 | How much the face is redrawn | 0.3 to 0.4 to keep identity, 0.5 to 0.6 to fix badly broken faces |
feather |
5 | Blur on the paste-back edge | Raise to 10 to 20 if you see a visible box |
noise_mask |
enabled | Only adds noise inside the mask | Keep enabled |
force_inpaint |
enabled | Redraws even when the face is already large | Keep enabled |
bbox_threshold |
0.5 | Detection confidence needed | Lower to 0.3 for small or side-on faces, raise to stop false detections |
bbox_dilation |
10 | Grows the face box by this many pixels | Raise to include hair and jawline |
bbox_crop_factor |
3.0 | How much surrounding area is included as context | 2.5 to 3.5 is a good range; too low gives mismatched lighting |
drop_size |
10 | Ignores detections smaller than this many pixels | Raise to skip tiny background faces |
sam_threshold |
0.93 | SAM mask confidence (only with a SAM model) | Leave as is |
wildcard |
empty | Optional separate prompt for faces | Use for face-specific details like “freckles, green eyes” |
cycle |
1 | Number of detailing passes | 2 helps very damaged faces |
Recipes for Common Problems
| Goal | Change these |
|---|---|
| Sharper faces in full-body shots | guide_size 768, denoise 0.4, bbox_crop_factor 3 |
| Keep the same face, just cleaner | denoise 0.3, same seed and prompt as the original |
| Fix a badly melted face | denoise 0.55, cycle 2 or a two-pass setup (below) |
| Group photo with many small faces | bbox_threshold 0.35, drop_size 20, guide_size 512 |
| Visible square around the face | feather 15, bbox_crop_factor 3.5, lower denoise |
| Face changes skin tone or lighting | Raise bbox_crop_factor and use the same prompt as the main image |
| Faster runs | face_yolov8n, guide_size 512, steps 15 |
SDXL, Flux and Other Models
FaceDetailer works with any model that plugs into a normal KSampler, because you pass in the MODEL, CLIP and VAE yourself.
- SD 1.5: guide_size 512, cfg 6 to 8.
- SDXL: guide_size 768 to 1024, the same cfg as your main pass.
- Flux: guide_size 1024, cfg 1, and put your guidance in the conditioning as you do in the main workflow. Use the same text encoders and VAE. See how to use Flux for the loader setup.
- LoRAs: feed FaceDetailer the MODEL and CLIP after your LoRA loader so a character or style LoRA also applies to the face. If the LoRA seems ignored, check our ComfyUI LoRA not working fix.
Using a different checkpoint just for faces is allowed, but the face may stop matching the body’s style. Start with the same model.
Hands, People and Two-Pass Detailing
Fix hands
Add a second UltralyticsDetectorProvider with bbox/hand_yolov8n.pt and a second FaceDetailer after the first. Use denoise 0.4 to 0.5. Hands are harder than faces, so try a few seeds.
Whole people
A person segmentation model such as segm/person_yolov8s-seg.pt lets you detail each person in a scene. The Impact Pack also shows a setup that combines a face bbox with a person silhouette mask to stop the background from being distorted.
Two-pass refine
For severely damaged faces, the README recommends a two-pass setup: a first pass at moderate resolution and settings to restore the rough shape, then a second pass to add fine detail. FaceDetailer (pipe) lets you share the common inputs between passes so you do not have to wire everything twice.
When the faces look right, finish with an upscale. Our ComfyUI upscale guide shows the best order of operations.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| UltralyticsDetectorProvider missing | Subpack not installed (not automatic since V8.0) | Install ComfyUI Impact Subpack, then restart |
| Model list is empty | Files in the wrong folder | Put bbox models in models/ultralytics/bbox and seg models in models/ultralytics/segm |
Error mentioning weights_only |
PyTorch 2.6+ safe loading blocks older .pt files | Use trusted models; the Subpack lets you whitelist safe files in ComfyUI/user/default/ComfyUI-Impact-Subpack/model-whitelist.txt |
| No faces detected | Threshold too high or faces too small | Lower bbox_threshold to 0.3, lower drop_size, try face_yolov8m |
| Background objects treated as faces | Threshold too low | Raise bbox_threshold to 0.6 |
| Face looks pasted on | Denoise too high or feather too low | Lower denoise, raise feather and bbox_crop_factor |
| Face looks the same as before | Denoise too low | Raise denoise to 0.45 to 0.5 |
| Black or grey face patch | VAE or precision issue | See our ComfyUI black image fix |
| Install fails or red nodes | Dependencies not in ComfyUI’s Python | Re-run pip with ComfyUI’s Python, then check the Impact Pack troubleshooting page |
| Out of memory | guide_size or max_size too high | Lower guide_size and max_size, or enable tiled_encode and tiled_decode |
FAQ
Is FaceDetailer the same as ADetailer?
It does the same job. ADetailer is the extension for Automatic1111 and Forge; FaceDetailer is the Impact Pack node for ComfyUI, and both use the same family of YOLO detection models.
Can FaceDetailer change who the person looks like?
At high denoise it can redraw the face noticeably. Keep denoise around 0.3 to 0.4 to preserve the original face. Do not use it to make images of real people without their consent.
Does FaceDetailer work on anime images?
Yes. The Bingsu face models are trained on both 2D and realistic faces.
Does it slow down generation?
Each face adds one extra sampling pass on a small crop. With one or two faces it usually adds a few seconds on a modern GPU.
Can I use FaceDetailer on photos I did not generate?
Yes, load the image and a checkpoint. Only edit photos you own or have permission to use.
What is cropped_refined?
It is a preview of each redrawn face before it is pasted back, useful for checking quality.




