Introduction

Black Forest Labs just unlocked the stills chapter of FLUX 3. On Oct 1, 2026 (19:00 UTC / 03:00 MYT Oct 2),

introduced FLUX 3 Image — generate, layout, and edit from one multimodal stack, with the pitch boiled down to four words: control every pixel.

This is not a soft teaser. The product page, docs overview, and release notes are live the same hour as the X drop. After FLUX 3 Video GA (Aug 4) and open-weights FLUX 3 Action (Sep 23), Image is the creator-facing beat: bounding-box compositions, multi-turn edits that refuse to smudge the rest of the frame, up to 10 references, and native output through 4K.

What BFL Claimed (and What Docs Confirm)

Pitch (X)Primary confirmation

Precise multi-turn edits without changing other pixels

Product demos + docs: recolor / replace / move / remove by box; untouched regions stay locked

Layout with bounding boxes

Boxes appended to prompt as a JSON element table on a 0–1000 canvas

Generate up to 4K

resolution: 768sq, 1k, 2k, 4k (~16 MP at the top end)

Up to 10 references

images accepts 1–10 URLs or base64 refs

Commercial weights for scale

Product “Commercial weights” section — fine-tune / deploy on your infra; talk to sales

Open weights in the coming weeks

Stated on X; not shipping today

Endpoint (docs): POST https://api.bfl.ai/v1/flux-3-image with x-key. Only prompt is required. There is no mode field — the prompt (plus optional images) decides generate vs edit. Unknown fields return 422.

Bounding Boxes Are the New Brush

The fun part is composition-as-data. Pick an aspect ratio; the canvas is a 0–1000 grid on both axes. Drag (or have an agent write) a box per element, describe what lives inside, add one scene line that ties them together — then FLUX 3 renders each element inside its box.

A layout prompt has two halves:

  1. A global caption that cites each element by id
  2. A JSON array of { id, bbox, desc } rows

Docs write boxes as [top, left, bottom, right]. The same boxes power targeted edits: re-describe, replace, move, or remove one element while anchoring the rest. Playground and agent flows both matter — BFL even demos an LLM planning the element table from a one-line brief so you only nudge the boxes you dislike.

You can also skip boxes and write a normal text prompt. Boxes are the control surface when layout, type placement, collages, or crowded scenes need to stay obedient.

Specs Creators Will Actually Quote

From the docs overview and release notes:

  • References: up to 10, each 256×256 to 16 MP
  • Aspect ratios: a long menu from 21:9 through 9:21, plus auto
  • grounding: default true (web + image search before generate)
  • Safety: safety_tolerance 0–4, default 2
  • Pricing per image: 768sq $0.041 · 1k $0.048 · 2k $0.100 · 4k $0.607

Native 2K / 4K is framed as detail preservation — shop signs, lantern glyphs, chopsticks through window glass — not a post-hoc upscale bolt-on.

Weights & Where to Try It

  • API / Playground: live now via BFL (/v1/flux-3-image)
  • Commercial weights: available for companies running image gen at scale (self-host / fine-tune path)
  • Open weights: “coming weeks” per the official note — watch HF / BFL for the Dev-style drop

ArtRealmAI Gen already speaks Flux (including Klein). FLUX 3 Image itself may land on Gen later; until then, use BFL’s Playground/API for the new boxes-and-pixel-edit tricks, and keep spinning Flux-family stills on Gen.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=flux-3-image-boxes-4k-pixel-edits

How This Differs From Prior FLUX 3 Coverage

  • July 23 multimodal announce — Image was “early access in the following weeks”
  • Aug 4 — FLUX 3 Video GA (generation + native audio)
  • Sep 23 — FLUX 3 Action open weights (robotics WAM)
  • Today — FLUX 3 Image for stills: boxes, pixel-locked edits, 10 refs, 4K, commercial weights

Same family, different chapter. If you only care about prettier stills and surgical edits, this is your drop.

Original Source

Also: FLUX 3 Image product page · docs overview · release notes / pricing · API reference

Conclusion

FLUX 3 Image is BFL putting a ruler on the canvas: box the elements, lock what you love, rewrite what you don’t, scale to 4K, and stack up to ten refs — with commercial weights for the infra crowd and open weights on the near calendar. The stills era of FLUX 3 just stopped being a roadmap bullet.

—Aurelia ♡