FLUX 3 Image: 10 references for one AI avatar

You finally have an AI avatar whose face you love. Then you ask for a new outfit, or a new street, or the same person holding your product, and something shifts: the jaw softens, the freckles move, the hair goes a shade darker. Every new image feels like starting over. Black Forest Labs just shipped a model built around exactly that problem: FLUX 3 Image, released on October 1, 2026.

• Up to 10 reference images in one request, each from 256 × 256 pixels up to 16 megapixels, so a face, an outfit, a product and a location can each come from their own photo.
• Native output up to 4K (about 16 megapixels) in 15 fixed aspect ratios plus auto, generated at that size rather than upscaled.
• Local edits that recolor, replace, move, resize or remove one element while the rest of the image is meant to stay untouched, including several edits in one request.
• Bounding-box layouts on a 0–1000 grid, plus base API prices from $0.048 per 1K image to $0.607 per 4K image, with 50% off the API through October 8.

FLUX 3 is a multimodal family that Black Forest Labs announced in July. We covered its video half in August. The image model is the piece that matters most for anyone building a recurring character, because every reel, carousel and talking-head clip starts as a still. If the still drifts, everything downstream drifts with it.

Until now, keeping a face stable usually meant one of two things: training an identity on a set of photos, or re-prompting with a long description and hoping. Multi-reference models changed that, and Nano Banana 2 already accepts up to 14 references. What FLUX 3 Image adds is control on top of the references: you can say where each element goes, and you can fix one detail without regenerating the whole frame. An open-weight version is expected in the coming weeks.

How to build a consistent AI avatar now that FLUX 3 Image takes 10 references

The following images were generated using Nano Banana 2:

1. Give every reference one job. Don't upload ten photos and ask the model to combine them. Write "face from image 1, outfit from image 2, location from image 3". Black Forest Labs' own guidance is to assign roles, and more references only help when each adds something the others don't.
2. Put your clearest face shot first, and set the aspect ratio yourself. With the ratio on auto, the first reference decides the output shape. If your face shot is square and you need a 9:16 reel frame, choose 9:16 explicitly.
3. Draft small, then commit. A 1K image costs $0.048 at base API price. Rendering again at 2K or 4K is a new generation, so the composition can change. Lock the shot at the size you'll publish, or upscale the draft you already like.
4. Edit one thing at a time and check the face every time. Pixel preservation outside an edit is the headline feature, but even the documentation says "usually". Faces, hands and fine textures are where small drift shows first.

Worked example — Mira at the café. This is the identity anchor: one clean, well-lit portrait with the features written out plainly (hair, freckles, eye colour). It becomes reference image 1 in every later request.

Prompt (Nano Banana 2, 1:1):
Photorealistic portrait of a woman in her late twenties with short curly auburn hair, light freckles and green eyes, wearing a rust-colored linen shirt, laughing while holding a ceramic coffee cup at a sunny outdoor cafe table in the morning, soft golden backlight, shallow depth of field, natural skin texture with visible pores, 50mm lens, candid lifestyle photography, no text, no logos

Worked example — Mira on the crosswalk. Same written identity, new outfit, new light, new motion. This is the kind of image you'd build next by feeding the café portrait as the face reference and a sweater and tote as separate references.

Prompt (Nano Banana 2, 1:1):
Photorealistic fashion photograph of a woman in her late twenties with short curly auburn hair, light freckles and green eyes, wearing a cream knit sweater and carrying a tan leather tote bag, walking across a bright city crosswalk at midday, motion in her stride, crisp daylight, clean urban background softly blurred, natural skin texture, 85mm lens, editorial street style, no text, no logos, no signs

The honest take: this is a control upgrade, not a magic consistency switch. In Astria's first matched test, published October 4, FLUX 3 Image produced the most natural, least staged photo of the four models tested and kept the reference face recognisable. Nano Banana 2 followed the requested pose more closely and finished in 32.3 seconds against FLUX 3's 87.3. Astria's identity reference was also below its own resolution threshold, so treat the face result as directional. Test it on your own character before you switch your whole pipeline.

Receipts: The Decoder on the launch, the capability and pricing breakdown and Astria's four-model comparison.

The bottom line for AI avatar creators

The hard part of an AI avatar was never the first good image. It was the fortieth, where the face, the clothes and the setting all have to agree with the last thirty-nine. FLUX 3 Image gives you more ways to hold those pieces still: separate references for separate jobs, boxes for where things go, and edits that leave the rest alone. Pair it with a locked identity and a written character description, and your recurring character gets much harder to break.

If you don't have that anchor yet, start with your avatar.

FAQs about FLUX 3 Image

When was FLUX 3 Image released?

Black Forest Labs released it on October 1, 2026, through its Playground and API. Companies can also license commercial weights, and an open-weight version is expected in the coming weeks.

How many reference images can FLUX 3 Image use?

Up to 10 per request. Each reference can be from 256 × 256 pixels up to 16 megapixels. Give each one a clear role in the prompt.

How much does FLUX 3 Image cost?

Base API prices in Black Forest Labs' release notes are $0.041 for a 768 × 768 preview, $0.048 at 1K, $0.100 at 2K and $0.607 at 4K. The API is 50% off through October 8, and third-party apps set their own prices.

Is FLUX 3 Image better than Nano Banana 2 for consistent characters?

Not across the board. In Astria's first test, FLUX 3 looked the most natural while Nano Banana 2 followed the pose better and ran faster. FLUX 3's real advantage is control: layout boxes and local edits. Nano Banana 2 accepts more references (up to 14).

Can FLUX 3 Image make video?

No. FLUX 3 Video is a separate model in the same family. Use FLUX 3 Image to make the stills, then animate them with the video model of your choice.