⚡ FLUX 2 GODMOD — by VeloPilot

👋 Created by VeloPilot — one-graph FLUX.2 Klein editor: text-to-image · editing · inpaint · outpaint · multi-reference

• 🌐 CivitAI Profile

• 🐙 GitHub: ComfyUI

• 🤗 FLUX.2 Klein 9B (Comfy-Org)

• 📖 Official ComfyUI FLUX.2 Klein guide

📋 Overview

One workflow that covers every FLUX.2 Klein task. You never rewire anything — you flip toggles.

1. Step 1 — Load your photo into LOAD IMAGE 1 (plus optional reference images 2 and 3)

2. Step 2 — Prompt — pick a ready task preset (character swap / clothing / pose) or write your own, optionally rewritten by a local LLM

3. Step 3 — Choose a mode — T2I, edit, inpaint or outpaint, all on one node

4. Step 4 — Run — masks can be painted by hand, generated by SAM3 text prompt, or both at once

Mode Toggles on the Sampler node What it does 🖼️ Text → Image all toggles OFF clean generation at the selected aspect ratio ✏️ Edit ENABLE EDIT MODE character swap, clothing, pose, object removal 🖌️ Inpaint ENABLE EDIT MODE + ENABLE INPAINT MODE + mask repaint only the masked area, denoise slider active 🔲 Outpaint ENABLE EDIT MODE + ENABLE OUTPAINTING MODE extend the canvas by N pixels on any side

⚠️ FIRST: install all custom nodes via ComfyUI Manager, then load the workflow.

✨ Features

• 🎚️ Everything on one node — modes, steps, denoise, expand values live on the Sampler subgraph, no digging around the canvas

• 🧩 Three image slots — main photo + two reference images (style, clothing, object) fed through ReferenceLatent

• 🎯 SAM3 masking by text — type jacket, hair, background and get a mask, no brush needed

• 🖌️ Hand mask + SAM3 combined — the two masks are added together

• 🧠 Optional LLM prompt enhancer — rewrites a short request into a full FLUX.2 prompt, can read the input image

• 📝 Task presets — character swap / clothing change / pose change templates that keep identity and framing locked

• 🧰 Preprocessors on board — DWPose, Depth Anything V2, Canny, HED, M-LSD, LineArt, auto face crop

• 🔀 One-click bypass panels — disable image slots or the LLM group without touching links

🔧 Required Custom Nodes

Node pack GitHub Used for rgthree-comfy rgthree/rgthree-comfy Power Lora Loader, Fast Groups Bypasser, Any Switch, Image Comparer, Image Resize ComfyUI-KJNodes kijai/ComfyUI-KJNodes JoinStrings, GrowMaskWithBlur, Sage Attention patch ComfyUI-Easy-Use yolain/ComfyUI-Easy-Use easy boolean / float / imageSize / isMaskEmpty, cache & VRAM cleanup ComfyUI-Easy-Sam3 yolain/ComfyUI-Easy-Sam3 SAM3 model loader, SAM3 image segmentation (text-prompt masks) ComfyUI-Crystools crystian/ComfyUI-Crystools Switch any / Switch conditioning (all mode toggles) comfyui_controlnet_aux Fannovel16/comfyui_controlnet_aux DWPose, Depth Anything V2, Canny, HED, M-LSD, LineArt ComfyUI_Comfyroll_CustomNodes Suzie1/ComfyUI_Comfyroll_CustomNodes CR Image Border (outpaint canvas) ComfyUI-AutoCropFaces liusida/ComfyUI-AutoCropFaces automatic face crop for reference images was-node-suite-comfyui WASasquatch/was-node-suite-comfyui Image Blank ComfyLiterals M1kep/ComfyLiterals Int

✅ TextGenerate, RegexReplace, ReferenceLatent, ResolutionSelector, SetLatentNoiseMask, PrimitiveStringMultiline and MarkdownNote are built into ComfyUI core — just keep ComfyUI up to date.

📦 Required Models & Files

🧠 diffusion_models

• flux-2-klein-9b-fp8.safetensors — distilled, the one this workflow is tuned for → models/diffusion_models/

• flux-2-klein-base-9b-fp8.safetensors — optional base version → models/diffusion_models/

🔒 The BFL repos are gated — accept the license on HuggingFace before downloading.

🔤 text_encoders

• qwen_3_8b_fp8mixed.safetensors — main text encoder → models/text_encoders/

• (optional) any uncensored Qwen3 / Gemma text encoder for the LLM prompt enhancer group → models/text_encoders/

💡 The screenshots use finetuned/abliterated text encoders. They are not required — point both loaders at the stock qwen_3_8b and everything works.

🌀 vae

• flux2-vae.safetensors → models/vae/

🎯 sam3 (text-prompt masking)

• sam3-fp16.safetensors or the original facebook/sam3 → models/sam3/

🧍 Preprocessors (auto-download)

• yolox_l.onnx and dw-ll_ucoco_384_bs5.torchscript.pt — pulled automatically by comfyui_controlnet_aux on first DWPose run

• Depth Anything V2 / HED / LineArt / M-LSD checkpoints — same, downloaded on demand

🎭 loras

• Any FLUX.2 Klein LoRA of your choice — drop it into the Power Lora Loader, and put its trigger words into the Lora keywords node. Nothing is required to run the workflow.

📁 Model Folder Structure

📂 ComfyUI/
├── 📂 custom_nodes/
│   ├── rgthree-comfy/
│   ├── ComfyUI-KJNodes/
│   ├── ComfyUI-Easy-Use/
│   ├── ComfyUI-Easy-Sam3/
│   ├── ComfyUI-Crystools/
│   ├── comfyui_controlnet_aux/
│   ├── ComfyUI_Comfyroll_CustomNodes/
│   ├── ComfyUI-AutoCropFaces/
│   ├── was-node-suite-comfyui/
│   └── ComfyLiterals/
└── 📂 models/
    ├── 📂 diffusion_models/
    │   └── flux-2-klein-9b-fp8.safetensors
    ├── 📂 text_encoders/
    │   ├── qwen_3_8b_fp8mixed.safetensors
    │   └── (optional) your-uncensored-LLM.safetensors
    ├── 📂 vae/
    │   └── flux2-vae.safetensors
    ├── 📂 sam3/
    │   └── sam3-fp16.safetensors
    └── 📂 loras/
        └── your_flux2_klein_lora.safetensors

🎬 How to use

1️⃣ Text → Image

Leave every toggle off, pick an aspect ratio in ResolutionSelector, write a prompt, run. Denoise is forced to 1.0 automatically.

2️⃣ Edit / character swap / clothing / pose

Load a photo into LOAD IMAGE 1, enable ENABLE EDIT MODE, then take a preset from the PROMTPS FOR HELP panel:

• CHARACTER SWAP — new person, same pose, framing, light and background

• CLOTHING — new outfit, same person and scene

• POSE — same person and outfit, new body position

Paste the preset into User prompt and finish the sentence after Specific request: with what you actually want.

3️⃣ Inpaint

Enable ENABLE EDIT MODE + ENABLE INPAINT MODE, then give it a mask in one of three ways:

• paint it on the mask loader node and connect MASK → mask of IMAGE 1 PROCESS

• type an object name into the SAM3 prompt inside IMAGE 1 PROCESS (jacket, hair, sign…)

• both at once — the masks are summed

Denoise is the only mode where the slider matters: 0.5–0.7 for a soft retouch, 0.9–1.0 for a full repaint.

4️⃣ Outpaint

Enable ENABLE EDIT MODE + ENABLE OUTPAINTING MODE and set Expand top / bottom / left / right in pixels. White borders are added, and the model paints them in with matching perspective, light and grain.

⚙️ Recommended settings

Parameter Value Note Steps 4 Klein 9B is distilled — more steps give a waxy, over-cooked look CFG 1.0 distilled model Guidance 4.0 drop to 2.5–3.5 if edits drift away from the original Sampler / Scheduler euler / simple avoid ancestral samplers Denoise 1.0 everywhere except inpaint forced automatically Aspect ratio match the source photo the input is center-cropped to the selected ratio

💡 Tips

• Reference images 2 and 3 are for style, an outfit or an object you want carried into the main photo. Turn their groups off in the bypass panel when unused — they cost VRAM.

• LLM enhancer is worth it for short requests. A long, precise prompt is still better written by hand — bypass the group and your text goes to the model untouched.

• Mask stuck on the image? ComfyUI saves painted masks into input/clipspace/ and re-points the loader at that file. Pick the original file in the node dropdown to reset it — that is why masking lives on a separate loader here, so the source photo never gets overwritten.

• Preprocessors (pose / depth / canny / lineart) are optional guides — enable one at a time, they are heavy.

• Low VRAM: bypass the LLM group and reference slots 2–3, and keep the resolution at 1MP.

🐛 Report Issues

Before reporting, please:

1. Update ComfyUI to the latest version — Update Guide

2. Update all custom nodes via ComfyUI Manager

3. Check that every model sits in the folder listed above

4. Tell me your GPU, VRAM and the exact error text from the console

Made by VeloPilot — AI Content Creator

This workflow is provided as-is for educational and creative purposes.

The author assumes no responsibility for any content generated with it.

Users are solely responsible for ensuring their use of all included

models, LoRAs, and custom nodes complies with their respective licenses

and terms of distribution. Third-party models retain their original

licenses - refer to each model's page on Hugging Face or GitHub for

full terms.