comfyui-mcp/packs/krea2-txt2img-json/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-txt2img-json
display_name: KREA 2 — Text to Image (JSON Prompt Builder)
family: krea2
kind: txt2img
description: >-
  KREA 2 (krea2_turbo) text-to-image driven by an Ideogram-4-style structured
  JSON prompt-builder (Ideogram4PromptBuilderKJ) — per-region descriptions and
  bounding boxes for precise area prompting, on the same krea2 turbo stack
  (native CLIPLoader type=krea2 + Qwen3-VL 4B text encoder + Qwen image VAE, at
  8 steps / cfg 1). Standalone single-pipeline graph sliced from the KREA2 ULTRA
  V2 monolith — no group toggles. V2 adds the Krea2T-Enhancer model detail-boost
  patch (ships ACTIVE). For a plain prose prompt use krea2-txt2img-manual; for the
  two-pass detail-boost combo use krea2-combo.
vram: "<24GB"
workflow: workflow.json
skill: krea2-txt2img
launch_args: []
sources:
  mirror: https://huggingface.co/Aitrepreneur/FLX
notes:
  - Needs ComfyUI with native KREA 2 support (comfy/text_encoders/krea2.py,
    ComfyUI >= v0.26.0) — the CLIPLoader uses type=krea2.
  - Run the generated installer from your ComfyUI root (folder with custom_nodes/
    and models/).
  - Default diffusion model is krea2_turbo_fp8 (RTX 4000/3000/2000). On RTX 5000
    (Blackwell) swap in krea2_turbo_mxfp8.safetensors for native fp8.
  - One clean pipeline, no group toggles; the Ideogram4PromptBuilderKJ drives the
    prompt. The manual-prompt branch lives in krea2-txt2img-manual; the two-pass
    detail-boost combo is krea2-combo.
  - V2 Krea2T-Enhancer (a MODEL detail-boost patch) ships ACTIVE in the model path
    (PowerLora -> Krea2T-Enhancer -> sampler). Bypass it to compare against the
    un-boosted result.
  - Smart Seed Variance (RBG_Smart_Seed_Variance) ships BYPASSED (optional,
    passthrough in the positive-conditioning loop) — un-bypass to explore
    controlled variations of the same prompt; set its seed mode to "randomize" and
    tune the variance mode/strength widgets.
  - JSON prompting tips — set ALL builder fields (prompt/background/technical/
    style/lighting), keep palettes EMPTY (a rich palette renders as a literal
    color-swatch strip), and add "no people / single full-frame photograph" for
    object/landscape scenes.
  - Render-verified — 1920x1080 tea still-life (teapot/cup/figs each placed by
    its bbox) at 8 steps / cfg 1 / er_sde.
post_install:
  - Restart ComfyUI (or ComfyUI-Manager -> Install missing custom nodes for any
    node Python deps).
  - Load workflow.json.