Video summary

AI Face & Hand Replacement Tutorial: Mastering Comfy UI Impact Pack Part 1

Main summary

Key takeaways

Technology

Tech / product overview

  • This tutorial is part of a ComfyUI series focused on replacing/fixing faces and hands using the ComfyUI Impact Pack.
  • Installation is done via ComfyUI Manager (custom node loader), specifically from “Dr LT data / Dr Lieutenant data” (same creator as ComfyUI Manager).
    • After the first install, restart ComfyUI.
  • The main node shown is Impact Pack → Face Detailer, which replaces faces by:
    • Detecting the edit region using a bbox detector (bounding box) model
    • Running an inpainting-like refinement focused on the cropped region

Base workflow shown (SDXL)

  • Starts from a basic SDXL + refiner KSampler workflow (using positive/negative prompts, base checkpoint + refiner checkpoint).
  • Adds Face Detailer using a detection model and (initially) the same prompts.
  • Internal components reviewed:
    • bbox detector (required to locate the face/hands region)
    • Optional models/settings, including bbox/segmentation variants, detector provider, segmentation models, etc.
  • Bounding-box detection models provided after installing the Impact Pack:
    • bbox face YOLO
    • bbox hand YOLO
    • bbox person YOLO
  • BBox vs segmentation:
    • “bbox” models use bounding boxes
    • “seum”/segmentation models use masks (for person)
    • For face/hands replacement, the tutorial uses bbox-face/bbox-hand.

Pipeline simplification (“pipe” concept)

  • Introduces Pipe nodes as an input combiner/transport mechanism to avoid manually connecting many parameters.
  • Uses:
    • Basic Pipe to bundle model + clip + vae + prompts
    • Basic Pipe → Detailer Pipe and/or Detailer Pipe nodes to wire quickly into Face Detailer

First-pass results + visual debugging

  • Demonstrates running Face Detailer on the generated SDXL image and comparing:
    • Base face vs Face Detailer output
  • Common quality/artifact issues reported:
    • Replacement can look “airbrushed” / lacking finished detail
    • Some areas (e.g., hair corners/forehead strands) may be altered slightly
  • Uses Face Detailer preview outputs such as:
    • crop refined
    • crop enhanced alpha to understand what region was detected and how it’s refined before replacement.

Multi-pass strategy (iterative face replacement)

  • Improves results by running Face Detailer in a second round:
    • Clone/reapply the pipeline so Face Detailer output is fed back into Face Detailer.
  • Observed effect:
    • The second pass often increases mouth/nose/feature clarity
    • But can still keep the airbrush look
    • Sometimes introduces artifacts

Advanced knobs tested (SDXL base/refiner behavior)

  • Uses an SDXL-specific Face Detailer pipe: basic pipe to detailer sdxl, supporting separate:
    • base pipe
    • refiner pipe
  • Key setting:
    • Refiner ratio: controls the step fraction where refinement switches from base to refiner
    • The tutorial explains step math (e.g., 20 steps with 10% → 2 steps in the refiner portion).
  • Experiments performed:
    • Adjusting refiner ratio (e.g., 20% then 30%)
    • Increasing step count (up to ~35)
    • Sampler choice (notably uler, suggested to work better with more steps)
  • Outcome:
    • Quality improved for the first pass around ~30 steps + uler
    • Further increases can overbake, causing deterioration or eye/teeth artifacts in later passes.

Model swap: moving away from the SDXL Detailer pipe

  • A major pivot happens because SDXL results are inconsistent.
  • Switch to DreamShaper (Excel) (download referenced via a CivitAI link in the description).
  • Reasoning:
    • DreamShaper is described as not needing a separate SDXL refiner (refiner “built in”)
    • So the tutorial removes the SDXL detailer pipe and uses basic pipe → face detailer pipe
  • Prompt wiring fix:
    • Text encoders were originally tied to the SDXL base CLIP, so the tutorial clones and reconnects CLIP text encoders to ensure prompts align with DreamShaper.

Two-pass DreamShaper Face Detailer run

  • Reported as the best face quality so far
  • Good facial structure and eyes, though still not perfect
  • Overall improvement compared to the SDXL base/refiner pipeline

Prompting refinements (“Primitives” concept)

  • Tries to improve face realism by changing conditioning:
    • Uses more specific “Primitive” prompts rather than general prompts for the entire image.
  • Example prompt strategy:
    • Positive: choose a target ethnicity identity like “Indian woman” (plus details such as “8K detail, digital photography, realistic”)
    • Negative: avoid cartoon/3D/disfigured/bad art, plus additional undesired outputs (e.g., painting/watercolor/blurry)
  • Observations:
    • Ethnicity conditioning can shift facial feature strengths
    • Second pass can amplify prompt strength
    • The rest of the image remains largely unchanged because Face Detailer applies modifications only within the detected region

Suggested next steps / limits of this video

  • The creator says they ran out of time to explain each Face Detailer option deeply.
  • Promises a follow-up video covering remaining features one by one.
  • Encourages viewers to share their best combinations:
    • model + prompt + sampler + step/refiner settings

Main speakers / sources

  • Primary speaker: The YouTube channel host/instructor (unnamed in subtitles).
  • Tool/authors referenced:
    • Dr LT data / Dr Lieutenant data (creator associated with the Impact Pack and ComfyUI Manager).

Original video