FLUX 3 Image API
Black Forest Labs' FLUX 3 Image pushes generation control to the pixel level: place every element with bounding boxes, edit per box while untouched areas stay pixel-identical, and render natively from 768sq up to 4K. Roko API serves it as an async task from $0.041 per 768sq request.
- Public model ID
flux-3-image- Output tiers
768sq / 1K / 1.5K / 2K / 4K (1K default)- Billing
Per request by resolution- Endpoint
POST /v1/images/tasks- Catalog snapshot
What is FLUX 3 Image?
FLUX 3 Image is the image generation and editing model Black Forest Labs released on October 1, 2026, under the slogan “control every pixel”. Its differentiator is bounding-box layout: the canvas is abstracted into a 0–1000 normalized coordinate grid, and you close each element in a box with its own description at the end of the prompt — the output satisfies those placement constraints.
Editing is also per box: changes to a finished image run box by box across multiple rounds, and everything outside the edited boxes stays pixel-identical, keeping repaint error local. It also takes up to 10 reference images (named by position in the prompt) plus grounding — live web lookup to enrich the prompt.
Roko currently exposes five resolution tiers (768sq, 1K, 1.5K, 2K, 4K) and 16 aspect ratios, billed per request at the chosen resolution ($0.041 at 768sq up to $0.607 at 4K). The values matched BFL's list price when reviewed, but Roko billing always follows the current project catalog. Calls run as async tasks: submit, get a task_id, then poll for the result.
Core capabilities
Bounding-box layout control
Place each element on a normalized coordinate grid and describe it per box — turning generation from a prompting-accuracy problem into a layout problem.
Per-box local editing
Modify finished images box by box across rounds; untouched areas stay pixel-identical, avoiding whole-image drift.
10-reference fusion
Combine up to 10 reference images, name each by position in the prompt, and let the model place and render them together.
Native 4K and 16 aspect ratios
Five native tiers from 768sq to 4K (~16.8 megapixels) across 16 aspect ratios — 4K keeps detail when enlarged.
FLUX 3 Image API pricing
Pricing comes from the current Roko catalog and is billed per request at the chosen resolution. The numbers matched BFL's list price when reviewed, but upstream pricing and promotions do not determine Roko billing; confirm the Roko estimate and final settlement.
| Resolution | Price | Billing unit |
|---|---|---|
| 768sq | $0.041 | per output |
| 1k | $0.048 | per output |
| 1.5k | $0.07 | per output |
| 2k | $0.1 | per output |
| 4k | $0.607 | per output |
Cost examples
Default tier — standard cost for web and social assets.
Native top tier for print and large-format output; billed per request, failed requests are not billed.
Parameters and supported scope
These fields come from the current public generation_config. Bounding-box syntax goes at the end of the prompt (0–1000 normalized coordinates); capabilities outside the form follow request validation.
| Parameter | Values | Notes |
|---|---|---|
prompt | textarea | Describe the image to create or the changes to apply to references. · Required |
images | url / base64 · 1–10 | Reference images may use URL or Base64; the current model field shows the applicable limit. · Optional |
aspect_ratio | auto / 21:9 / 2:1 / 16:9 / 3:2 / 7:5 / 4:3 / 5:4 / 1:1 / 4:5 / 3:4 / 5:7 / 2:3 / 9:16 / 1:2 / 9:21 | Selects the output aspect ratio; auto lets the model decide from the content. · Default: auto · Optional |
resolution | 768sq / 1k / 1.5k / 2k / 4k | Selects an output resolution tier configured for this model; the value may directly affect price. · Default: 1k · Optional |
safety_tolerance | 0–4 | Roko currently exposes 0–4: 0 is strictest and 4 is loosest; defaults to 2. · Default: 2 · Optional |
grounding | boolean | When enabled, the model may consult live web sources to enrich the prompt; on by default. · Default: true · Optional |
Call FLUX 3 Image with Roko API
Run the async task flow: submit a task to get a task_id, then poll for the result. Never expose an API key in client-side code.
Submit generation
curl https://api.rokoapi.com/v1/images/tasks \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: YOUR_UNIQUE_REQUEST_ID" \
-d '{
"model": "flux-3-image",
"prompt": "Modern city park poster: bounding boxes — fountain center (400,500)-(600,800), skyscrapers background left (0,0)-(350,700), oak trees background right (650,100)-(1000,750)",
"aspect_ratio": "16:9",
"resolution": "1k"
}'Query recovery task
curl https://api.rokoapi.com/v1/tasks/$TASK_ID \
-H "Authorization: Bearer $API_KEY"POST /v1/images/tasks returns 202 with a task_id; once generation completes, query GET /v1/tasks/{task_id} for the result or failure reason. Idempotency-Key must be unique; replays with the same key return the original task without re-billing.
Best-fit scenarios
Precisely composed layouts and posters
Headlines, subjects, and background elements all obey box constraints — ad layouts and poster composition.
Poster: product bottle center, headline text box top-left, lifestyle scene bottom — each element in its own bounding box.
Multi-round local refinement
Edit box by box with pixel-stable surroundings — fit for finals that go through repeated sign-off.
Edit only the bounding box around the logo: replace with the new wordmark, keep everything else pixel-identical.
Multi-reference scene synthesis
Fuse people, products, and settings from several references into one box-laid-out image.
Combine the model from image 1, the outfit from image 2, and the studio background from image 3 into one composition.
4K print and large format
Native 4K (~16.8 MP) holds detail when enlarged — trade-show panels and print collateral.
Trade-show backdrop at 4K: brand hero center, product lineup along the bottom edge, clean gradient background.
Important limitations
- Billing is per request and the 4K tier ($0.607) costs far more than lower tiers — use 768sq or 1K for volume drafts.
- Bounding boxes use 0–1000 normalized coordinates written at the end of the prompt; malformed boxes are treated as plain text or rejected by request validation.
- Roko currently limits safety_tolerance to 0–4 (0 strictest, 4 loosest), defaulting to 2; confirm compliance before raising it.
- A submitted task is not a successful generation — trust the task query result, and review outputs before use.
FLUX 3 Image or another image model?
Pick FLUX 3 for precise layout and per-box editing; the Seedream family fits Chinese-language scenes and image sets; GPT Image 2.5 suits transparent backgrounds and everyday content. Weigh control granularity, resolution, and price.
All models →FLUX 3 Image FAQ
What model ID should I use for FLUX 3 Image?
Use the public Roko API model ID flux-3-image exactly as shown in the model field.
How much does the FLUX 3 Image API cost?
Billing is per request at the chosen resolution: $0.041 at 768sq, $0.048 at 1K, $0.07 at 1.5K, $0.10 at 2K, $0.607 at 4K. The values matched BFL's list price when reviewed, but Roko billing follows the project estimate and final settlement at submission.
How do bounding-box layouts work?
The official scheme normalizes the canvas to 0–1000 and puts one box plus description per element at the end of the prompt, e.g. “fountain (400,500)-(600,800)”. The output satisfies those placement constraints.
What does per-box editing mean?
You edit a finished image box by box: only the area inside a box is repainted, everything outside stays pixel-identical, and you can iterate across rounds without whole-image drift.
Which resolutions and aspect ratios are supported?
Roko currently exposes 768sq, 1K, 1.5K, 2K, and 4K tiers plus 16 aspect ratios (auto by default). 4K is about 16.8 megapixels — a native render, not an upscale.
Which endpoint does FLUX 3 Image use?
Submit async tasks via POST /v1/images/tasks to get a task_id, then poll GET /v1/tasks/{task_id} for the result or failure reason.
What does the grounding parameter do?
With grounding enabled, the model may consult live web sources to enrich the prompt (real-time facts or brand-material descriptions). It is on by default; disable it explicitly when not needed.
Can I use the open weights instead?
As of the latest review (2026-10-06), BFL says open weights will follow in the coming weeks and commercial weights require direct licensing; on Roko, requests run through BFL's official API channel.