Flux Consistent Character: Workflows for Same-Face Generations

To get a Flux consistent character, give FLUX.2 or FLUX 3 Image one or more reference images of the character, refer to each by position ("the woman in image 1"), and start every prompt with the same fixed identity description. For pose and scene changes, edit in a chain with one change per step and say what must stay; for strict likeness at volume, train a LoRA on FLUX.2 [klein] Base.

Models, limits and prices checked October 2026 against Black Forest Labs' FLUX.2 overview, pricing, FLUX.2 prompting guide, character consistency guide, multi-reference editing, single-reference editing, Kontext editing and klein training docs.

This page is the Flux-specific route: which Flux model to use, where in the prompt the face actually lives, the edit chain for poses, and when to train a LoRA. For how Flux compares with Nano Banana, Midjourney, PhotoMaker and face swap, read our consistent character AI guide first; it is the hub for every method.

What you will have at the end

Which Flux model to use for a consistent character

Black Forest Labs now ships three generations side by side. The Kontext docs call FLUX.1 Kontext a previous-generation model and recommend FLUX.2 for new editing projects. FLUX 3 Image, the newest, takes 1 to 10 reference images per request. Prices below are from the official pricing page (checked October 2026); FLUX.2 prices scale with output megapixels, so treat them as starting points.

Illustrative arithmetic: a month of 30 feed images at FLUX.2 [pro]'s starting generation price is about $0.90 before retries, and at FLUX 3 Image's 1k price about $1.44. Real spend runs higher because you generate several candidates per post; budget for the candidates, not the keepers.

Prompt-side anatomy: which words carry the face

The FLUX.2 prompting guide states it plainly: word order matters, the model pays more attention to what comes first, and the recommended order is subject, then action, style and context. For a persona that means the face lives in the first clause. Everything that changes per post goes after it.

• Stable features first, in fixed wording. Age range, face shape, eye colour and shape, hair colour, length and texture, and one or two signature details. Copy the block exactly; a synonym is a different instruction.
• Repeat it every time. Black Forest Labs' own comic-strip example keeps a character consistent by repeating the same detailed description in every panel prompt.
• Name references by position. With input images, the multi-reference guide says to call them "image 1", "image 2" and so on, use the same words every time, and state each image's role: subject, product, setting or style.
• Describe, never negate. FLUX.2 does not support negative prompts and FLUX 3 Image has no negative prompt field. Write "bare skin with natural texture", not "no makeup".
• Keep style out of the identity block. Words like glamorous or editorial pull facial features as well as the look. Put them in the style clause.

The woman in image 1, late twenties, oval face, wide-set hazel eyes, straight dark brown hair to the collarbone, small mole under the left eye, sits at a cafe window reading, cream knit sweater, overcast daylight, 50mm photo, shallow depth of field

For automation, the same split fits the JSON structured prompts the overview documents: the identity block becomes a fixed subject value and the scene fills background, lighting and camera_angle. Our free AI influencer prompt generator writes a starting identity block from a persona brief.

Flux consistent character workflow, step by step

Step 1 expected result: one hero image. Step 3: three views where the landmarks (eyes, nose, jaw, hairline) match the hero. The single-reference guide's pattern for these edits is name the target, say what changes, then name what stays, for example "turn her head to a three-quarter view facing left; keep her face, hair and expression unchanged".

Step 5: watch the aspect ratio. On FLUX 3 Image, aspect_ratio defaults to auto, which follows the first reference image, so a square hero gives you square feed posts unless you set it. The multi-reference guide also says to put the base image first; for a persona that is the hero. FLUX 3 Image's grounding field, on by default, runs a web and image search before generating; a fictional persona has nothing to look up, so consider switching it off.

Flux Kontext consistent character: the edit chain for poses

The Kontext docs say FLUX.1 Kontext keeps characters consistent across multiple edits, and the FLUX.2 character consistency guide shows the same thing: one reference photo, then sequential edits that change the scene and weather while the character stays recognisable. Use that as an edit chain, with one change per step:

Pose prompts that work well state the new pose and the things that stay: "she now sits on the steps with her elbows on her knees, looking at the camera; keep her face, hair and outfit unchanged". Black Forest Labs' season-change example opens with "Keep the woman's pose unchanged" before describing the new weather and coat, which is the same pattern in reverse.

The chain runs the same way on FLUX.2 editing and FLUX 3 Image, which add multi-reference input, so you can include the hero as image 2 while editing image 1. Our Flux Kontext guide covers Kontext itself, including running it in ComfyUI.

Flux consistent character LoRA: when references are not enough

Reference images carry a face well on angles close to the references and less well far from them. When you need the exact persona across hundreds of images, many angles and expressions, train a LoRA. Black Forest Labs' klein training guide recommends the undistilled FLUX.2 [klein] Base models for LoRA training and lists character consistency as a use case:

• License. klein 4B Base is Apache 2.0; klein 9B Base is under the FLUX Non-Commercial License. For a commercial persona, 4B is the clean choice.
• Hardware. Minimum listed: an NVIDIA GPU with 12GB VRAM and 32GB of RAM. The guide puts LoRA training at 1 to 3 hours on consumer GPUs.
• Dataset. High-resolution images (1024px or more), consistent quality, artifacts removed, detailed captions and a consistent trigger word.

Your approved reference views and the best scene outputs from the steps above become the start of that dataset. Our character LoRA dataset guide covers curation and captions, and the LoRA training guide covers the run itself. The AI Influencers program walks through the whole stack, from identity block to trained LoRA to video.

Run it locally or through the API?

Both routes use the same prompts and the same reference discipline; the difference is cost, control and licensing.

• API or playground. FLUX 3 Image, FLUX.2 [pro], [max] and [flex] run in Black Forest Labs' playground and API with no setup, at the same price in both according to the pricing page. Best when you want the newest model and no GPU to manage.
• Local with klein. The overview says FLUX.2 [klein] 4B runs on consumer GPUs with about 13GB of VRAM under Apache 2.0, and takes up to four references. Best for high volume once your reference set is fixed.
• Local with dev. FLUX.2 [dev] is free for non-commercial use with about six recommended references. Good for testing a workflow before you pay for API runs, not for a commercial persona.

If you go local, ComfyUI is the usual host because it lets you save the whole chain (references, prompts, edit steps) as one reusable graph. Our ComfyUI consistent character workflow has the node templates, including the built-in FLUX.2 klein edit route, and the guide to the best ComfyUI models covers where Flux checkpoints fit.

Troubleshooting Flux face drift

• The face softens into a generic one. Too many competing references, or style words in the identity block. Cut the set back to the hero plus the closest view.
• Small details wander. Moles and freckles are the hardest features to hold. Pick signature details that survive generation, such as hair colour and cut, and judge identity on the landmarks.
• Two characters merge. Give each its own reference image and name each one by position every time it is mentioned.

For the general causes of drift (seed, denoise, restoration, crop scale), see our AI face consistency guide.

Flux consistent character: FAQ