ComfyUI Flux Kontext Guide: AI Image Editing with Text Instructions

What is Flux Kontext?

FLUX.1 Kontext is a generative model designed for text-driven image editing. Unlike text-to-image models, Kontext understands both the input image and your text instructions, allowing precise edits like changing object colors, swapping backgrounds, transferring styles, and editing text in images — all while preserving the parts you don't mention.

Model Versions

The Dev version is non-commercial by default. Commercial licenses are available from Black Forest Labs.

Hardware Requirements

Setup

Kontext Dev shares CLIP and VAE files with the Flux model family. If you already have those, you only need the Kontext diffusion model.

Model Downloads

Diffusion Model (choose one):

Shared Flux components:

Basic Workflow

1. Load Diffusion Model → Kontext model
2. DualCLIPLoader → clip_l + t5xxl
3. Load VAE → ae.safetensors
4. Load Image → the image you want to edit
5. CLIP Text Encode → your editing instruction (English only)
6. KSampler → generates the edited image
7. VAE Decode → Save Image

What Can Kontext Edit?

Object Modification

Simple, direct changes to specific elements:

Change the car color to red

Replace the flowers with sunflowers

Add a small cat on the table

Style Transfer

Convert the image to a different artistic style:

Convert to oil painting with visible brushstrokes, thick paint
texture, and rich color depth while maintaining the original
composition

Background Changes

Swap the environment while preserving the subject:

Change the background to a beach while keeping the person in the
exact same position, scale, and pose

Character Consistency Editing

Modify character attributes while preserving identity:

Change her clothing to a red dress while preserving her exact
facial features, hairstyle, and expression

Text Editing in Images

Modify text that appears in the image:

Replace 'OPEN' with 'CLOSED' while maintaining the same font
style and color

Prompt Writing Tips

Be Specific, Not Vague

Explicitly Preserve What Matters

Always state what should stay the same:

Change the background to a forest while maintaining the same
camera angle, subject position, and lighting

Name Subjects Directly

Use Quotes for Text Edits

Replace 'joy' with 'BFL'

Choose Verbs Carefully

Multi-Round Editing

Kontext supports iterative editing — edit the output of one generation as input for the next. Two approaches:

1. Load Image (from output) — load the previous result and re-run with a new prompt
2. Group Nodes — ComfyUI added a quick "Edit" button for chaining Kontext edits in one workflow

For complex transformations, break them into steps:

1. First edit: change the environment
2. Second edit: modify clothing
3. Third edit: adjust lighting

Multiple Image Input

You can reference multiple images by:

1. Image Stitch — combine images into one and input as a single reference (better results)
2. ReferenceLatent chaining — encode images separately and chain conditions (may mix features)

For stitched input, make the main reference image larger in proportion.

Accelerated Inference

TeaCache

Install ComfyUI-TeaCache — add the TeaCache node to your existing Kontext workflow for faster generation with minimal quality loss.

Nunchaku (4-bit)

Install ComfyUI-nunchaku for 4-bit quantized inference. Requires a specific Nunchaku model:

• svdq-int4_r32-flux.1-kontext-dev.safetensors

Common Issues and Fixes

Image doesn't change at all

• The Dev version needs more specific prompts than Pro/Max
• Be more explicit about what to change: "Change the dress color from blue to red" instead of "Change the color"
• Try a different seed

Identity changes too much during character edits

• Add preservation instructions: "while preserving exact facial features, eye color, and expression"
• Use "Change the clothes to..." instead of "Transform the person into..."

Composition shifts unexpectedly

• Add: "keeping the person in the exact same position, scale, and pose"
• Specify: "Maintain identical camera angle, framing, and perspective"

Style transfer loses important details

• Be more specific about style characteristics
• Add: "while preserving all scene details and object positions"

Related Guides

• Flux Guide — Flux text-to-image generation
• Inpainting Guide — Alternative approach for targeted edits
• Image to Image — Transform entire images