FLUX 3 Image Explained: Edits That Leave the Rest of Your Photo Alone

October 6, 2026 · 7 min read

FLUX 3 Image is the image-generation side of Black Forest Labs' FLUX 3 model family, released on October 1, 2026 — and its headline trick is multi-step editing that changes only what you ask for, leaving every other pixel in the photo untouched. If you've ever asked an AI to change the jacket in a portrait and got back a different face, different lighting and a different dog in the background, you understand why this matters. BFL — the lab founded by the researchers who built Stable Diffusion's core architecture — has been teasing the FLUX 3 family since July, and the image model is the piece photographers and casual editors have been waiting for.

What FLUX 3 Image actually is

FLUX 3 is BFL's first multimodal foundation model — one backbone that generates images, video with native audio, and even robot actions. The family was announced on July 23, 2026, and the video side went generally available in early August (we covered the landscape in our Wan 3.0 guide, which launched the same week). FLUX 3 Image, opened on October 1, is the still-image half of that system, and it arrives with a very specific engineering goal: controlled editing instead of lucky regeneration.

Three capabilities define the release, per BFL's launch material and early coverage:

• Multi-step instruction editing. You can chain edits — "change the jacket to leather", then "make it evening light", then "remove the car" — and each step builds on the last instead of re-imagining the whole image from scratch.
• Bit-identical pixel preservation. Regions you didn't ask to change come back numerically identical to the source — not "close", but the same pixels. That's the difference between an editing tool and a slot machine.
• JSON bounding-box layouts. You can hand the model explicit regions — this box is the subject, this box is the background — and it locks the geometry of what stays and what changes, which is how the preservation above is achieved.

The model is live on BFL's own playground and rolling out through API partners, with per-image pricing in the few-cents range depending on the host.

Why the "leave the rest alone" problem was the real bottleneck

AI photo editing has had a dirty secret since the beginning: most models don't edit your photo, they re-generate it with your request mixed in. Ask for a different shirt and the model quietly re-draws the hands, softens the skin, moves the shadows — because it has no concept of "everything else stays". For playful generation that's fine. For anything real — a product photo where the label must stay legible, a portrait where the face must stay yours, a brand image where the logo can't drift — it's been the dealbreaker that kept professionals doing final edits in Photoshop.

Pixel-preserving, region-locked editing is the fix, and FLUX 3 Image is the highest-profile model to ship it as a first-class feature rather than a mask-painting side mode. Expect every major image lab to answer within a quarter — the same way native audio became table stakes in video after Veo 3.

What you can do with this on Hermosso today

An honest note first: FLUX 3 Image is brand new and is not yet in Hermosso's studio picker as of writing — the Create studio currently runs FLUX 2 Pro among 20 image models across 21 engines (including GPT Image 2, Nano Banana 2 and Seedream 5), and FLUX 3 already powers one of the 18 video engines behind Photo to Video. But the jobs FLUX 3 Image demos are jobs Hermosso's editing tools do today, so here's how to act on the news without waiting:

• Change one thing, keep the rest. Change Outfit swaps clothing on your photo while preserving face, pose and background — the exact use case BFL is showing off. Remove Object erases a person, car or blemish and rebuilds only that region.
• Relight without re-shooting. Relight re-does the lighting of a portrait — evening warmth, studio softbox, golden hour — on the same pixels, which is the second most popular multi-step edit after clothing.
• Try the style before the model. If what excites you about FLUX 3 Image is the aesthetic, you don't need to wait for any specific engine: browse the photoshoot packs, pin the looks you like to a moodboard, and generate yourself in that style with the studio's current lineup.
• Finish the job. Whatever you edit, take the keeper to print resolution with the 22K Upscaler — five tiers from faithful 4K up to 22K — or animate it into a short clip with Photo to Video.

If you specifically want FLUX-class image generation inside the studio, FLUX 2 Pro is in the picker today — select it in Create and route the same prompt through it and a rival engine to compare, which is the whole point of a multi-model studio.

What to watch next

Two things follow from this release. First, the editing race: OpenAI added character-consistency reference editing to ChatGPT Images 2.5 the same week (covered in our GPT Image 2.5 guide), and Midjourney opened its first instruction-edit model in August — "edit, don't re-roll" is now the frontier everyone's competing on. Second, the open-weights question: BFL built its reputation on open FLUX variants, and whether a downloadable FLUX 3 Image checkpoint follows will decide how fast the capability spreads into local tools like ComfyUI. As of writing, access is hosted-only.