"Up to 10 reference images" is the headline number on FLUX.2's product page. It is also not the number you will actually get.
Buried in Black Forest Labs' multi-reference documentation is the constraint that governs real work: the [pro] API has a 9MP total limit covering input and output together. At 1MP output you can use up to 8 references. At 2MP output, 7. Push the output higher and the budget for references shrinks accordingly. FLUX.2 [klein] tops out at 4 references regardless.
Character consistency is where this bites first, because consistency is exactly what you spend references on.
Sourcing note: the reference limits, budget rule, and prompting guidance below come from Black Forest Labs' multi-reference editing and character consistency documentation. Last verified July 30, 2026.
What this guide solves
The pain point: teams design a campaign workflow around "10 references," then hit validation errors or quality drops at production resolution and cannot work out why. The reference count is not a fixed allowance — it is what is left after your output resolution takes its share.
The differentiator: the actual budget arithmetic, plus the prompt structure that makes multi-reference work, rather than another gallery of consistent-looking characters with no method attached.
The budget rule
Per BFL's documentation:
| Output resolution | Max reference images ([pro]) |
|---|---|
| 1 MP | 8 |
| 2 MP | 7 |
| higher | fewer, on the same 9MP total |
FLUX.2 [klein] supports up to 4 references regardless of output size.
The mental model: you have 9 megapixels of total budget. Your output claims its share first, and references split the rest — each one counting at least 1 MP, rounded up.
The practical consequence is a trade you should make deliberately: a 4MP hero shot leaves room for very few references. If identity consistency matters more than resolution, generate at 1MP with a full reference set and upscale afterwards. If resolution is non-negotiable, accept fewer references and lean harder on prompt description.
This also explains a cost pattern people notice but rarely diagnose: a multi-reference edit is several times the price of a plain generation, because every reference is billed as at least 1 MP on top of the output.
How consistency actually works
BFL describes character consistency as relying on multi-reference editing: you provide one or more references of your character, then describe the new scene, pose or context. The model preserves identity — face, clothing, proportions, style — while adapting to the new setting.
Their documentation highlights four things it holds across generations: face, clothing, proportions, and style. That list is worth reading as a checklist of what you can expect to survive an edit, and by omission, what you should re-specify each time.
Documented workflows where this matters:
- Iterative editing — sequential edits where the character stays recognisable through the whole chain
- Fashion editorial — a coherent model across every shot in a series
- Character placed in a new scene — same person, completely different setting
- Season and outfit change — environment and clothing change, character does not
The prompt rule that makes it work
BFL's guidance for multi-reference is short and it is the whole game:
Describe the role of each image so the model knows what to pull from where.
Without role assignment, several references become an averaging problem. With it, each image has a job:
- "The person of image 1" — identity source
- "in the setting of image 2" — environment source
- "wearing the jacket from image 3" — garment source
Their broader instruction is to be specific about what changes and clear about what stays. Note the asymmetry with virtual try-on, where you deliberately stay silent about what stays — in try-on the person image is the thing being edited, so naming its contents confuses the target. In scene compositing, the references are separate sources and naming their roles is what disambiguates them.
Building the reference set is the real work. Flux 3 AI is a browser workspace for generating character sheets, style plates and scene references — the assets a consistency workflow runs on. Open the image generator or see the plans.
A working method for campaign sets
- Build a character sheet first. Four to six images of the same subject across angles and lighting. This is your reusable identity source.
- Decide output resolution before reference count. Do the 9MP arithmetic, then pick how many references fit.
- Assign a role to every reference in the prompt. No unassigned images.
- Re-specify what should stay in scene compositing, since references are separate sources.
- Generate at 1MP for consistency-critical work, then upscale. Resolution is recoverable; identity drift is not.
- Keep the same reference set across the series. Swapping references mid-campaign reintroduces the drift you were avoiding.
Where it still drifts
Honest limits, since no vendor documents these:
- Long sequential chains accumulate small changes; BFL's own examples show sequential editing working, but re-anchoring to the original reference periodically is safer than editing the edit of the edit.
- Faces at small scale have less signal to preserve — a character 200 px tall in frame is a weaker identity constraint than a portrait.
- Style plus identity plus composition all at once is more asymmetric than it sounds; if something has to give, it is usually the least-described one.
FAQ
How many reference images can FLUX.2 use? Up to 8 at 1MP output, 7 at 2MP, fewer as output grows — a 9MP total budget on [pro]. Klein supports up to 4.
Why does the product page say 10? That is the model's maximum; the API's 9MP input+output budget is what constrains a real request.
What does FLUX preserve across edits? BFL names face, clothing, proportions and style.
Why is my multi-reference edit so expensive? Every reference counts as at least 1 MP on top of the output, at your tier's per-MP rate.
How do I stop references from blending together? Assign each one a role in the prompt — identity from image 1, setting from image 2, and so on.
Should I generate at high resolution for consistency work? Usually no. Generate at 1MP with a fuller reference set, then upscale.
Does klein work for character consistency? Yes, with up to 4 references — enough for simpler continuity, tight for full campaign work.
Bottom line
Character consistency in FLUX.2 is a budgeting exercise as much as a prompting one. Nine megapixels, split between what you make and what you show the model. Decide that split deliberately, give every reference an explicit role, and favour 1MP-plus-upscale when identity matters more than pixels.
Build the character sheet before you need it — generate it in the Flux 3 AI workspace and reuse it across the whole campaign.
Sources
- Multi-Reference Editing (BFL documentation) — 9MP budget, per-resolution reference counts, klein limit, role-assignment guidance
- Character & Style Consistency (BFL docs) — what identity preservation covers, documented workflows
- FLUX.2 model page — multi-reference positioning and the headline count
- FLUX.2 image editing (BFL docs) — editing workflow detail
- BFL API pricing — per-megapixel billing including reference images
- Product Consistency (BFL docs) — the same technique applied to products
- Single-Reference Editing (BFL docs) — the simpler case, for contrast
- Official FLUX.2 prompting guide — general prompt structure that still applies
- FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — multi-reference as a headline FLUX.2 capability
- FLUX.2 klein model page — the cheaper tier and its reference ceiling
Scope note: limits as of July 30, 2026. Flux 3 AI is an independent creator workspace, not affiliated with Black Forest Labs.
