Nodes/comfyui-mudknight-utils/FastDetailer (full-pipe)
ComfyUI Node

FastDetailer (full-pipe)

The same face fixer, now with one wire instead of five

By mudknight·Created 8 months ago·Updated 2 months ago· 1
FastDetailer (full-pipe)
  • full_pipe
  • full_pipe
  • image
  • cropped_image
bbox_model
fallback_model
threshold0.50
steps20
cfg1.5
samplereuler_ancestral_cfg_pp
scheduleralign_your_steps
denoise0.40
upscale_method
upscale_model
feather0.20
context_padding0.10

This is the pipe-flavored twin of FastDetailer, and it's the one you want if the mudknight pack's full-pipe convention is doing the wiring for you. Instead of dragging in a model, a VAE, positive and negative conditioning, and a seed, you feed it one full_pipe - the single bundle this pack passes around containing all of those - and it hands the whole thing back, updated.

It does exactly the same job as the plain DetailerNode: detect a face with a YOLO model, crop it, upscale to ~1MP, run a fresh sampling pass, downscale, and feather it back into the original. Same face_yolov8m.pt detector, same euler_ancestral_cfg_pp / align_your_steps defaults, same denoise, feather, and context_padding dials. What changes is the interface.

The inputs that matter

Required inputs are bbox_model, fallback_model, threshold, the full_pipe, and the sampler controls (steps, cfg, sampler, scheduler, denoise, upscale_method, upscale_model, feather, context_padding). The one that trips people up is that there's no separate seed or image input - those ride inside the pipe. If your pipe doesn't have an image in it, the node raises an error telling you exactly that, so make sure something earlier in the chain put an image into the pipe (a FullPipeIn or a loader that sets one).

Outputs are full_pipe (with the detailed image written back into it), image (the finished composite, so you can preview without unpacking), and cropped_image for debugging your detector.

Installing it

Same pack, same requirement: ComfyUI Manager, search comfyui-mudknight-utils, restart - but you also need Impact Pack and Impact Subpack installed, because the detection runs on UltralyticsDetectorProvider and BboxDetectorSEGS from that ecosystem. The README says this outright, and it's the #1 reason people think this node is broken when it just has no detector to call.

cd ComfyUI/custom_nodes
git clone https://github.com/mudknight/comfyui-mudknight-utils

Where people get burned

Pipe-detailing has one gotcha the standalone node doesn't: order of operations. If you detail after upscaling, the crop is already big and the 1MP upscale is a no-op; if you detail before a final upscale, the detail can get blurred away by the upscale afterward. Detail late, but not last, and keep the denoise around 0.4 unless the face needs a rebuild. And if the pipe comes back with a face that's wandered off-model, that's the 1MP-cohesion tradeoff from the FastDetailer docs doing its thing - dial denoise back rather than blaming the node.

Categorydetailer

Inputs (13)

NameTypeDefaultDescription
bbox_modelCOMBO1 options: bbox/face_yolov8m.pt
fallback_modelCOMBO2 options: none, bbox/face_yolov8m.pt
thresholdFLOAT0.500–1Detection threshold
full_pipeFULL_PIPE
stepsINT201–10000
cfgFLOAT1.50–100
samplerCOMBOeuler_ancestral_cfg_pp44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38
schedulerCOMBOalign_your_steps10 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +4
denoiseFLOAT0.400–1
upscale_methodCOMBO5 options: lanczos, bilinear, bicubic, area, nearest-exact
upscale_modelCOMBO1 options: none
featherFLOAT0.200–1Percentage of image to feather when uncropping
context_paddingFLOAT0.100–1Percentage of image to use for context from edge

Outputs (3)

NameTypeDescription
full_pipeFULL_PIPE
imageIMAGE
cropped_imageIMAGE