comfyui-mcp/packs/krea2-identity-edit/pack.yaml

Local-first, agent-native control plane for ComfyUI — MCP server + autonomous sidebar agent that drives your live graph in natural language on ANY LLM: Claude/ChatGPT/Gemini on your subscription (no API key), free local models via Ollama (fully offline),

name: krea2-identity-edit
display_name: Krea 2 Identity Edit — outfit swap
family: krea2
kind: edit
description: >-
  Reference-driven OUTFIT SWAP and identity-preserving editing on KREA 2. Two images in:
  a PRIMARY reference (the person to keep — face, hair, body, pose) and a SECONDARY
  reference (the clothing or style to apply). Krea2EditGroundedEncode grounds the edit on
  both, so a plain-English instruction changes the outfit while the subject and background
  survive. Runs krea2_turbo fp8_scaled with the Qwen3-VL 4B encoder and the Krea 2
  Identity Edit v1.2 LoRA. Local, 10 steps, cfg 1 — no API node, no per-image cost.
vram: "16GB+ (fp8_scaled turbo; krea2_raw_bf16 wants 24GB+)"
workflow: workflow.json
skill: krea2-identity-edit
launch_args: []
sources:
  model: https://huggingface.co/Comfy-Org/Krea-2
  lora: https://huggingface.co/conradlocke/krea2-identity-edit
  nodes: https://github.com/lbouaraba/comfyui-krea2edit
  mirror: https://huggingface.co/Aitrepreneur/FLX
notes:
  - >-
    Run the generated installer from your ComfyUI root (the folder containing custom_nodes/
    and models/).
  - >-
    REFERENCE ORDER MATTERS, and it is the thing people get backwards. PRIMARY (group 1) is
    the SUBJECT — the output always matches this person. SECONDARY (group 2) is the OUTFIT.
    Swapping them swaps which identity survives.
  - >-
    The LoRA must land in loras/Krea2/ — the workflow loads it as
    "Krea2/krea2_identity_edit_v1_2.safetensors". A flat loras/ install leaves the loader
    pointing at a file that is present but unfindable.
  - >-
    THE FIDELITY DIALS FOLLOW THE IMAGE SLOTS, NOT THE ROLES. Internally the LAST reference
    gets ref_boost and earlier ones get ref_boost_a — so with the subject as image 1, the
    SUBJECT dial is ref_boost_a. Raise it if the face drifts.
  - >-
    Steps: 10 / cfg 1 on turbo. 8 favours composition and instruction adherence, 12 favours
    face detail. Swapping in krea2_raw_bf16 instead wants 40 steps and cfg 3-4, where the
    negative prompt starts to matter.
  - >-
    Set the output resolution to MATCH THE PRIMARY (subject) image — the ResolutionSelector
    drives EmptySD3LatentImage, and a mismatched aspect crops or letterboxes the subject.
  - >-
    Prompt in plain English and name both roles, e.g. "Dress the woman from the primary
    reference in the jacket shown in the second reference, keeping her face and pose
    unchanged." Naming what to PRESERVE is as load-bearing as naming what to change.
post_install:
  - >-
    Restart ComfyUI (or reload the browser tab) so comfyui-krea2edit registers
    Krea2EditGroundedEncode / Krea2EditModelPatch, then load workflow.json.
  - >-
    CLIPLoader type must be "krea2" — that requires a ComfyUI with native KREA 2 support.
    If the type is missing from the dropdown, update ComfyUI first.