Flux Kontext keeps a character consistent when each edit names the person by description, changes one thing and states what stays. Kontext still runs, but Black Forest Labs now calls it a previous-generation model and recommends FLUX.2 for editing; FLUX 3 Image launched on 1 October 2026. The twelve outfit, scene and pose prompts here work on all three.
Flux Kontext character consistency depends on how you word the edit: name the person by description instead of a pronoun, change one thing per request, and state what must stay the same. Kontext still runs in October 2026, but Black Forest Labs now calls it a previous-generation model and recommends FLUX.2 for new editing work, so this page gives you twelve edit prompts that work on Kontext and carry over unchanged to FLUX.2 and FLUX 3 Image.
Checked October 2026 against Black Forest Labs' Kontext overview, Kontext editing docs, pricing, release notes, character consistency guide, single-reference editing guide and pose guidance guide, plus the Kontext launch post, ComfyUI's Kontext dev tutorial and the Kontext dev model card. The prompts are written to the patterns those pages document; they are starting points to adapt, not benchmark results.
This is the edit-recipe page. For how instruction editing compares with LoRAs, reference adapters and face swap, start at our consistent character AI guide, the hub for every method. For the wider Flux route (identity block, reference set, klein LoRA), read the Flux consistent character workflow.
Is Flux Kontext still the model to use? What changed
Kontext works, and it is no longer the default. Black Forest Labs' Kontext pages now open with a notice recommending FLUX.2 and describe Kontext as "a previous-generation model"; the pricing page lists it under "FLUX.1 Models (Previous Generation)". Nothing has been switched off: the pro and max endpoints, the playground and the open-weight dev model are all still there.
- 29 May 2025FLUX.1 Kontext announced
Pro and max on the API; dev in private beta. Editing by text instruction, one input image.
- 25 Nov 2025FLUX.2 [pro] and [flex] launch
Multi-reference editing: up to 8 input images through the API, 10 in the playground.
- 16 Dec 2025FLUX.2 [max] launches
Multi-reference editing of up to 10 images.
- 15 Jan 2026FLUX.2 [klein] launches
Sub-second generation with open weights.
- 23 Jul 2026FLUX 3 announced
One model trained across image, video and audio, released in phases.
- 1 Oct 2026FLUX 3 Image launches
One endpoint that generates, edits regions and combines up to ten references, up to 4K.
Source: Black Forest Labs launch post and release notes, checked October 2026
| Model | Status | Reference images | Output | Price (checked October 2026) |
|---|---|---|---|---|
| FLUX.1 Kontext [pro] and [max] | Previous generation; still on the API and playground | One input image (pro editing docs) | About 1MP (pro editing docs) | $0.04 (pro) or $0.08 (max) per image |
| FLUX.1 Kontext [dev] | Open weights for local use; editing only; non-commercial licence | One input image in the ComfyUI tutorial | Depends on your setup | Free to download |
| FLUX.2 [pro] | What Black Forest Labs recommends for editing | Up to 8 (API) or 10 (playground) | Up to 4MP | Editing from $0.045 per megapixel |
| FLUX 3 Image | Newest; launched 1 October 2026 | 1 to 10 | Up to 4K | $0.048 at 1k, $0.10 at 2k |
The practical reading: if you already have a Kontext workflow, keep using it for single-image edits. If you are starting today, or you keep hitting the limits in the failure section below, write the same prompts against FLUX.2 or FLUX 3 Image and add reference images. Illustrative arithmetic on the list prices: running all twelve recipes with three candidates each is 36 images, which is $1.44 on Kontext [pro] and about $1.73 on FLUX 3 Image at 1k.
What you need before the first edit
- One approved hero image of a fictional character: face fully visible, even light, neutral expression
- The hero saved as a file you never overwrite; every edit branches from it
- A one-line description of the person that only matches them ("the woman with shoulder-length dark brown hair")
- A Black Forest Labs account with credits (1 credit is $0.01), or Kontext dev installed in ComfyUI
- A shot list: which outfits, scenes and poses you need this week, one edit per line
- A place to log the prompt behind each keeper, so a good edit can be repeated
Do not use a photo of a real person as the hero unless you have their written consent. Kontext edits whatever face it is given, which is exactly why the starting image has to be yours to use.
How to edit a character with Flux Kontext, step by step
Every Kontext request is one image plus one instruction. The editing docs list a single input_image (up to 20MB or 20 megapixels) and an output of about 1MP, so the whole method is about spending that one instruction well.
- 1Load the hero
Upload the approved image in the playground, or wire it to Load Image in ComfyUI. Expected: the untouched hero is the input.
- 2Pick one change
One line from the shot list: an outfit, a scene or a pose. Expected: you can say the edit in a single sentence.
- 3Write the three-part prompt
Who (by description), what changes (from, to), what stays. Expected: no pronouns and no "transform".
- 4Set the frame
Leave aspect ratio alone to match the input, or set it between 3:7 and 7:3. Expected: output near 1MP in the shape you need.
- 5Generate two or three candidates
Change only the seed between runs. Expected: the same edit with small variations to choose from.
- 6Compare against the hero at full size
Check eyes, nose, jaw and hairline side by side. Expected: one keeper, or a reworded prompt.
- 7Save and log
Download at once (API result links expire after 10 minutes) and note the prompt and seed.
- 8Branch the next edit from the hero
Start again from step 1 unless the next change truly depends on this output.
The prompt pattern behind every recipe below:
[the person, by description] + [the one change: from what, to what] + [what must stay the same]Each part comes from the documentation. ComfyUI's Kontext tutorial says to use a specific description such as "the woman with short black hair" rather than "she", and to state what to preserve with a clause like "while maintaining the same facial features, hairstyle, and expression". Black Forest Labs' editing guide adds the middle: name the element as it is now and as it should be.
Flux Kontext outfit change: four recipes
Outfit edits are the safest place to start, because the face is not part of the change. Name the garment the character is wearing now, so the model has exactly one thing to replace.
| Edit | Prompt to adapt | Check before you keep it |
|---|---|---|
| 1. Swap one garment | Change the grey hoodie worn by the woman with shoulder-length dark brown hair to a cream knit sweater, while maintaining the same facial features, hairstyle and expression. Keep the pose and background unchanged. | Neckline and hair length: a new collar often invites new hair. |
| 2. Full outfit | Change the clothes of the woman with shoulder-length dark brown hair to a black blazer over a white T-shirt and straight blue jeans, while preserving the facial features and hairstyle. Keep the pose, framing and background unchanged. | Body proportions against the hero, not just the face. |
| 3. Colour only | Change the colour of the dress to deep emerald green, keeping the fabric texture, folds and fit. Keep everything else unchanged. | That only the dress changed colour, not the lighting on the skin. |
| 4. Add a layer | Add a tan trench coat over the current outfit of the woman with shoulder-length dark brown hair, and small gold hoop earrings. Keep the face, hairstyle, pose and background unchanged. | Ears and jawline, which the new items sit next to. |
Exact brand colours are easier on the newer models: Black Forest Labs' single-reference guide documents hex codes in the prompt ("Change the man's grey T-shirt to #c0392b, keeping the print on the front") for FLUX.2 and FLUX 3 Image. On Kontext, use plain colour names and check the result.
Scene and background changes: four recipes
A scene change goes wrong when the model re-stages the whole photo. Pin the subject first, then describe the new world around it.
| Edit | Prompt to adapt | Check before you keep it |
|---|---|---|
| 5. Background swap | Change the background to a sunlit cafe terrace while keeping the woman in the exact same position, scale and pose. Maintain the same facial features, hairstyle and outfit. | Subject size in the frame: it should not have moved. |
| 6. Light and time of day | Change the lighting to golden hour with warm low sun from the left. Keep the woman, the outfit and the composition unchanged. | Skin tone and eye colour under the new light. |
| 7. Weather or season | Keep the woman's pose unchanged. It is now raining lightly, the pavement is wet and reflective and the sky is overcast. Maintain the same facial features, hairstyle and outfit. | Wet-hair effects you did not ask for. |
| 8. New location, new framing | The woman with shoulder-length dark brown hair is now standing in a hotel lobby with a marble floor, photographed from the waist up. Maintain the same facial features, hairstyle and expression. | Everything. This is the widest edit on the list; compare at full size. |
Recipe 7 follows Black Forest Labs' own season-change example, which opens with "Keep the woman's pose unchanged" before it describes the snow. Recipe 5 uses the wording ComfyUI's tutorial gives for exactly this problem: "keeping the person in the exact same position, scale, and pose".
Flux Kontext pose change: four recipes
Kontext can change a pose from words alone, and the smaller the move, the better the face survives. State the new pose as a fact ("is now sitting"), then list what stays.
| Edit | Prompt to adapt | Check before you keep it |
|---|---|---|
| 9. Gaze | The woman with shoulder-length dark brown hair is now looking directly at the camera. Keep the head angle, hairstyle, outfit and background unchanged. | Eye shape and spacing. |
| 10. Expression | Change the expression of the woman with shoulder-length dark brown hair to a relaxed closed-mouth smile, while maintaining the same facial features and hairstyle. Keep everything else unchanged. | Cheeks and mouth width; smiles reshape a face the most. |
| 11. Sit or stand | The woman with shoulder-length dark brown hair is now sitting on the steps with elbows resting on knees. Maintain the same facial features, hairstyle and outfit. Keep the location and lighting unchanged. | Hands, and face size if the frame got wider. |
| 12. Head turn | Turn the head of the woman with shoulder-length dark brown hair to a three-quarter view facing left. Maintain the same facial features, hairstyle and expression. Keep the outfit and background unchanged. | Nose and jaw from the new angle; Kontext is inventing the side it never saw. |
- Transform her into a businesswoman
- Put her on a beach
- Make her look happier and change the outfit and the background
- She turns around
- Make it better
- Change the clothes to a navy suit while preserving facial features
- Change the background to a beach, same position, scale and pose
- Three requests: expression, then outfit, then background
- Turn the head to a three-quarter view facing left; keep the features
- Warm the light to late-afternoon sun; keep the background unchanged
When Flux Kontext loses the face
Kontext drifts for a short list of reasons, and most of them are documented. Work down this list before you blame the model.
- The verb was too big. ComfyUI's tutorial flags "Transform the person into a Viking" as the prompt that changes a character too much, and gives "Change the clothes to be a viking warrior while preserving facial features" as the fix. Use change and replace.
- The subject was vague. A pronoun, or no subject at all, leaves the model free to redraw the person. Use the one-line description every time.
- The chain was too long. Black Forest Labs' own launch post lists it as a limitation: "Excessive multi-turn editing sessions can introduce visual artifacts that degrade image quality", with an example that breaks down after six edits.
- The angle was never shown. Kontext has one input image. A profile or a back view asks it to invent the parts of the head the hero never showed, so expect a relative rather than a twin.
- The face got small. Output is about 1MP in total. In a full-body frame the face is a small share of those pixels, and there is less detail left to match.
- The instruction was ignored. The launch post also says the model "occasionally fails to follow instructions accurately". Rerun with a new seed before you rewrite a good prompt.
When to move the same prompts to FLUX.2 or FLUX 3 Image
Move when one input image is the thing holding you back. The newer models read the same instructions and add reference images, which answers the two hardest Kontext cases: unseen angles and exact poses.
- 01Image 1: hero
The identity. Put it first; FLUX 3 Image takes the output shape from image 1.
- 02Image 2: the extra
A second approved view, a garment, or a pose reference.
- 03Prompt by position
Refer to "image 1" and "image 2" and give each a role.
- 04Check against the hero
Same test as before: eyes, nose, jaw, hairline.
- Exact poses. The pose guidance guide gives the wording for FLUX.2: "Match the exact pose from image 2: same arm position, same body angle, same gaze direction. Use the person and clothing from image 1." It also warns that structure is interpreted semantically, so do not expect a pixel-perfect match.
- Outfits from a product photo. The multi-reference guide's pattern is "The person from image 1 wears the jacket from image 2 and stands in the kitchen from image 3."
- More pixels. FLUX.2 outputs up to 4MP and FLUX 3 Image up to 4K, which keeps a face detailed in wider frames.
- FLUX 3 Image specifics. Requests carry an
imageslist of 1 to 10 references, each at least 256 by 256 pixels and at most 16 megapixels, andaspect_ratiodefaults toauto, which follows image 1.
If you want an instruction editor with open weights instead, our Qwen Image Edit guide covers the main alternative, and the same three-part prompt applies. When edits stop being enough, because you need the same face across hundreds of images and every angle, that is the point to train a LoRA: the AI Influencers program takes it from creating the persona through LoRA training for consistent faces and ControlNet for body consistency.
Where to run Flux Kontext in 2026
- Black Forest Labs playground. The fastest way to try the recipes. The Kontext overview still points to the playground for testing pro and max, and pricing is the same there as on the API.
- The API.
flux-kontext-protakes a prompt and an input image and returns a link to the result. Useful once your shot list is a spreadsheet. - ComfyUI, hosted models. ComfyUI's tutorial notes that Kontext pro and max run through Partner Nodes, so you can keep the edit inside a larger graph.
- ComfyUI, local. Kontext dev runs from open weights and its template loads the previous output back in for multi-round edits. The repository is gated and the licence is non-commercial for the model, so read the terms before client work. Our ComfyUI consistent character workflow shows where an edit step sits in a full persona graph.
Whichever route you pick, the review step is the same. Our AI face consistency guide covers how to compare a candidate against the hero and which features to judge first.
Flux Kontext character editing: FAQ
Is Flux Kontext good for character consistency?
Yes, within limits. Black Forest Labs says FLUX.1 Kontext keeps a character consistent across multiple edits, and its open-weight model card describes successive edits with minimal visual drift. It works from one input image, so it holds a face best when the change is small and the face is large in the frame. Long edit chains and big angle changes are where identity slips.
Is Flux Kontext still worth using in 2026?
It still runs on the Black Forest Labs API and playground, and the open-weight dev version still runs locally in ComfyUI. Black Forest Labs now calls Kontext a previous-generation model and recommends FLUX.2 for new editing projects because it accepts several reference images and outputs up to 4MP. FLUX 3 Image, launched on 1 October 2026, takes up to ten references.
How do I change an outfit in Flux Kontext without changing the face?
Name the person by description, name the garment as it is now and what it becomes, then say what stays. For example: change the grey hoodie worn by the woman with shoulder-length dark brown hair to a cream knit sweater, while maintaining the same facial features, hairstyle and expression. Use the verb change or replace, and edit one garment per request.
Can Flux Kontext change a pose and keep the same person?
Small pose changes work from a prompt: a new gaze, an expression, sitting instead of standing. State the new pose and list what stays, such as facial features, hairstyle and outfit. Kontext has no pose reference input, so for an exact pose use FLUX.2, where Black Forest Labs documents passing a pose image as a second reference and naming the arm position, body angle and gaze to match.
Why does Flux Kontext change my character's face?
Usually the prompt. ComfyUI's Kontext documentation shows that the verb transform causes broader edits than change, that pronouns are weaker than a specific description, and that an unpinned scene change moves the subject. Black Forest Labs also states that excessive multi-turn editing can introduce artifacts, so a long chain of edits on top of edits degrades the face.
Can I use Flux Kontext dev for commercial work?
Read the licence before you do. FLUX.1 Kontext dev is released under the FLUX.1 dev Non-Commercial License and the Hugging Face repository is gated. The model card says generated outputs can be used for personal, scientific and commercial purposes as described in that licence, and Black Forest Labs sells commercial licences for the model itself. The hosted pro and max versions are paid per image.
One hero image, twelve edits. Then build the whole persona.
AI Influencers, included in All Access, covers Flux and SDXL setup, creating the persona, LoRA training for consistent faces, ControlNet for body consistency and video, with the other three programs, weekly coaching and the community in one subscription.
Write the one-line description first
The free prompt generator drafts a persona description you can reuse in every edit. Then share your hero image and your first edits in the free Telegram for feedback.