ComfyUI FLUX.1 Kontext: Multi-Turn Editing and Character Consistency

You already have a character portrait you like: short silver hair, amber eyes, and the right clothing details. You only ask for a city street instead of a stage, but the second image changes the face, hair color, and outfit at once. After another pose and expression edit, the third image starts to blur. FLUX.1 Kontext addresses this problem by treating the input image as context and letting the prompt separate what should change from what should stay fixed.

It is not a permanent identity lock. Black Forest Labs demonstrates multi-turn consistency, but it also shows a failure case with visible artifacts after six edits. The safer workflow is short iterative sessions, saved intermediate images, and one main variable per turn.

What FLUX.1 Kontext is and is not for

FLUX.1 Kontext Dev is a 12B image-editing model that accepts an image and a text instruction. It can change backgrounds or clothing, replace local details, transfer style, edit text, and continue editing a previous output.

Goal	How Kontext handles it	Prompt invariants	Main risk
Change the background, keep the person	Edit the background alone	face, pose, scale, position	The face or framing changes with the scene
Change clothing, keep the face	Describe only material and color	facial features, hairstyle, expression	Clothing and body shape drift together
Move a product to another scene	Rewrite only the environment	shape, logo, material, proportions	The logo or package geometry is redrawn
Build a series through several turns	Change one dimension at a time	Repeat the required invariants	Small errors accumulate
Maintain a long-lived character asset	Use Kontext to explore directions	Keep a clean baseline	Training is more stable for repeated production

Kontext solves a different problem from other controls. ControlNet is better for pose, depth, and outline constraints. LoRA is better for a character or style reused over time. IPAdapter injects reference-image conditions, while FaceID-style tools focus on facial identity.

Install Kontext Dev models in the right folders

The current official ComfyUI tutorial uses an FP8 scaled diffusion model, a VAE, and two text encoders. The model page may require accepting BFL terms before downloading files.

ComfyUI/models/
├── diffusion_models/
│   └── flux1-dev-kontext_fp8_scaled.safetensors
├── text_encoders/
│   ├── clip_l.safetensors
│   └── t5xxl_fp16.safetensors
└── vae/
    └── ae.safetensors

The official workflow also supports t5xxl_fp8_e4m3fn_scaled.safetensors instead of the FP16 T5 file. Refresh the model list after installation. If a loader still cannot find a file, verify the folder name and download integrity, then restart ComfyUI. For general import and missing-node problems, see the ComfyUI workflow reuse guide.

Open the native template

Update ComfyUI or ComfyUI Desktop and look for Flux.1 Kontext Dev in Workflow Templates. Menu names can change, so update the application before installing an unknown node pack to replace a missing template.

Run the first edit with the native node chain

The official Dev workflow follows this sequence:

1. Select flux1-dev-kontext_fp8_scaled.safetensors in Load Diffusion Model.
2. Select clip_l.safetensors and a T5 encoder in DualCLIP Load.
3. Select ae.safetensors in Load VAE.
4. Load the source image in Load Image(from output).
5. Enter an English instruction in CLIP Text Encode.
6. Click Queue or press Ctrl/Cmd + Enter.

The current ComfyUI page still states that this Kontext Dev workflow supports English prompts only. Do not rely on occasional success with another language. A short, explicit English instruction makes it easier to determine whether a failure comes from the model, prompt, or graph.

Change one object in the first turn

Test a background replacement before changing the scene, outfit, pose, age, and style together. A single-variable edit creates a clean baseline and reveals which instruction introduced drift.

Change the background to a city street at night.
Keep the person in the exact same position, scale, and pose.
Preserve the same facial features, hairstyle, expression, and clothing.

Write both the edit and the invariants

A Kontext prompt does not need a stack of generic quality terms. It needs an object, an action, and the parts that must remain fixed. Words such as beautiful, better, or make it different leave the edit boundary unclear. Verbs such as change, replace, keep, and preserve separate the editable region from the invariant one.

Scenario	Too vague	More controllable instruction
Background replacement	Put her on a beach	Change the background to a beach while keeping the same face, pose, scale, and position
Clothing edit	Transform her into a Viking	Change only the clothing to Viking armor while preserving facial features and hairstyle
Product recolor	Create a new bottle design	Change only the bottle color to blue; preserve shape, logo placement, material, and proportions
Style transfer	Make it a sketch	Convert to a pencil sketch while maintaining the original composition and subject identity

Preservation terms for people, products, and framing

Target	Invariants to state	Failure signal
Character identity	facial features, hairstyle, expression, body shape	Eyes, face shape, hair, or apparent age changes
Product geometry	shape, logo placement, material, proportions	Logo spelling, package curve, or material changes
Composition	composition, camera angle, position, scale	Subject moves, crops, or changes perspective
Pose	pose, hand position, body orientation	Limbs and silhouette deviate

If pose must be exact, adding more prose to the Kontext prompt is usually the wrong move. That requirement points toward ControlNet.

Save, inspect, and roll back multi-turn edits

Load Image(from output) uses the previous result as the next input. A practical order is background first, then clothing or accessories, then slight pose or expression changes, with lighting and style left until the end.

1. Save the untouched baseline.
2. Change only the background and save round-01.
3. Change only clothing or accessories and save round-02.
4. Make a small pose or expression adjustment.
5. Apply style, color, or lighting last.

When to roll back

Return to the last clean intermediate image as soon as the face, eyes, logo, proportions, or sharpness clearly worsens. Do not keep adding instructions that ask a degraded image to restore the original face. Each new turn can redraw more areas and carry the mistake forward.

There is no fixed safe number of turns. BFL shows a six-edit failure with visible artifacts, but that does not mean five turns are always safe or the sixth always fails. The source image, edit scope, and instruction quality all matter.

Practical consistency recipes for characters and products

Describe visible character features instead of relying only on she or him:

Change only the background.
Keep the same woman with short silver hair and amber eyes.
Preserve her facial features, hairstyle, expression, body shape, pose, and scale.

For a product, list brand-sensitive details as invariants:

Place the same bottle on a marble counter.
Preserve the exact bottle shape, logo placement, label text, material, proportions, and camera angle.
Change only the environment and surrounding lighting.

These prompts reduce drift; they do not replace review. Check every commercial product image for logos, text, proportions, packaging claims, and materials. Use real-person images only when you own them or have permission.

Choose between Dev, Pro, Max, FP8, and GGUF

Option	Access	Best fit	Important boundary
Kontext Dev	Local open weights	Research, customization, privacy, frequent tests	Weights use the Non-Commercial License
Kontext Pro	API or Partner Nodes	Commercial editing and quick iteration	Pricing and terms depend on the current service
Kontext Max	API or Partner Nodes	Stronger instruction following and text editing	It is not the local Dev weight
FP8 scaled	Official ComfyUI tutorial path	Fastest route to the native workflow	Memory needs still depend on resolution and environment
GGUF or Nunchaku	Community or specialized implementation	Lower memory use or acceleration	Nodes, quantization quality, and compatibility need testing

Do not treat “12GB runs it” or “16GB is always smooth” as guarantees. Memory also depends on T5, resolution, VAE, previews, offload, and additional nodes. See ComfyUI low-VRAM optimization for that separate problem.

Start guidance testing from the example

The Hugging Face Diffusers example uses guidance_scale=2.5. It is a reasonable test value, not proof that a fixed range is optimal for every ComfyUI graph. Keep the input and seed fixed and change one parameter at a time.

Troubleshoot common failures in order

Symptom	Check first	Next action
The model is missing from a loader	Folder, filename, download integrity	Refresh models and restart
The prompt has little effect	English wording and a specific target	State the edit with change or replace
The face changes with the background	Too many dimensions changed at once	Add face, hair, and pose invariants
Each turn looks blurrier	A degraded output is still being reused	Return to the last clean intermediate image
The product logo changes	Logo and label were not declared invariant	Narrow the edit and review manually
Composition or pose drifts	The task needs structural control	Switch to or combine ControlNet
VRAM is insufficient	FP8, T5, resolution, extra nodes	Evaluate GGUF, offload, or an API

Keep the source image and seed fixed during troubleshooting. Replacing the model, quantization, resolution, and prompt together makes the result impossible to diagnose.

Licensing, responsible use, and the final tool choice

FLUX.1 Kontext Dev weights are currently governed by the FLUX.1 dev Non-Commercial License. The model card says generated outputs may be used for personal, scientific, and commercial purposes, but that does not grant permission to place the Dev weights in a paid service. Check current BFL licensing for commercial deployment. Pro and Max are API products whose terms depend on the provider.

Choose by the result you need:

• Use Kontext for short-term background, clothing, local, and concept edits.
• Use LoRA for a character or style reused in long-term production.
• Evaluate FaceID-style tools when the main requirement is the same face, with proper portrait consent.
• Use ControlNet when pose, outline, or composition must be constrained precisely.
• Evaluate IPAdapter when the task is to inject style or subject conditions from a reference image.

After the first successful run, keep the baseline and proceed through background, clothing, slight motion, then lighting or style. Kontext’s advantage is not perfect permanence. It gives each edit explicit context and lets you return to the most recent clean version when drift appears.