Most "how to use FLUX 3" tutorials published this month have a problem: their author has never used FLUX 3. Access is request-gated, so a step-by-step with confident screenshots of a settings panel is, at best, a screenshot of something else.
Here is what an honest version looks like — the access path that actually exists, the inputs the model accepts according to Black Forest Labs' own capability list, and the prompt discipline that transfers directly from FLUX.2 to FLUX 3 so you are not starting cold when your access clears.
Sourcing note: access mechanics and capabilities come from BFL's FLUX 3 announcement (July 23, 2026) and model page. Prompt technique comes from BFL's official FLUX.2 prompting guide — the current documented rules for the FLUX family, which is the closest verifiable ground for FLUX 3 technique until BFL ships FLUX 3-specific docs. Nothing here is reconstructed from a demo video. Last verified July 28, 2026.
What this guide solves
The pain point: you cannot follow a tutorial for a product you cannot open. People searching "how to use FLUX 3" hit one of two dead ends — either a fake walkthrough, or a page that says "coming soon" and stops there. Both waste the time you could be spending getting your prompt library ready.
The differentiator: this guide separates the three things clearly — (1) how to get access, (2) what the model will accept when you have it, and (3) the transferable workflow you can practise today on tools you already have. That third part is the one nobody writes, and it is the only part that pays off this week.
Step 1 — Get access the only way that exists
FLUX 3 access runs through Black Forest Labs. There is no reseller, no key marketplace, no bypass.
- Open the FLUX 3 model page or the announcement post.
- Click Request early access.
- Understand which stage is open. As of late July 2026, FLUX 3 Video is the capability in Early Access. FLUX 3 Image early access was described as opening "in the following weeks." Action prediction is going to selected research and commercial partners, beginning with mimic robotics.
- Expect review, not instant provisioning. BFL describes the early access phase as being for smooth rollout, feedback collection, and rigorous safety testing.
- If you are a company with volume or partnership needs, BFL routes that through enterprise contact rather than the standard request.
Red flag check: if a site offers you FLUX 3 without a BFL request, it is running a different model. There is no public FLUX 3 API and no published FLUX 3 pricing.
Step 2 — Know what FLUX 3 accepts as input
This is the part worth memorising before access clears, because it determines how you prepare assets.
FLUX 3 generates video with native audio — audio is not a second pass — and BFL states it can produce up to 20 seconds in a single generation. The documented modes:
| Mode | What you provide | What you get |
|---|---|---|
| Text-to-video | Prompt only | Video with audio |
| Image-to-video (animation) | A starting frame | Video continuing from that frame |
| Image-to-video (reference) | Images as visual references | Video guided by that identity or style |
| Video-to-video | A reference clip | New scene carrying elements — e.g. the same character |
| Video-audio continuation | Input video and audio | Generated continuation |
| Keyframe-to-video | Defined moments | Controlled transitions between them |
Two capabilities to plan around: multilingual dialogue, and agentic chaining of individual clips into longer, multi-shot sequences. BFL notes that with visual references maintaining character consistency, chained sequences can run several minutes.
For images, BFL says FLUX 3 synthesises and edits across a wide range of styles, aspect ratios, and resolutions, with notably improved complex-prompt handling and multilingual text rendering — reported from mid-training evaluations, with the caveat that results are preliminary.
Step 3 — Practise the prompt discipline that transfers
FLUX 3-specific prompting docs do not exist yet. FLUX.2 docs do, and the family's prompting behaviour is documented and consistent. Practising these now means your library carries over.
The structure BFL actually recommends
Subject + Action + Style + Context.
- Subject — the main focus: person, object, character
- Action — what they are doing, or their pose
- Style — artistic approach, medium, aesthetic
- Context — setting, lighting, time, mood, atmosphere
The four rules that matter most
- No negative prompts. FLUX.2 does not support them. Describe what you want, never what you do not want. If you are coming from a Stable Diffusion workflow, this is the habit that will hurt you most.
- Word order is weight. The docs are explicit: FLUX pays more attention to what comes first. Priority order is main subject → key action → critical style → essential context → secondary details.
- Length has a sweet spot. Official guidance: 10–30 words for quick concepts, 30–80 words for most projects, 80+ only for complex scenes needing detailed specification.
- Be concrete about optics. "Shot on Fujifilm X-T5, 35mm f/1.4" outperforms "professional photo." Name cameras, lenses, film stocks, and eras.
Text and colour, where FLUX is unusually strong
For readable text in an image, the documented technique: put the string in quotation marks ("The text 'OPEN' appears in red neon letters above the door"), specify placement relative to other elements, describe the style, and set size. For brand work, hex codes work — associate them explicitly with an object ("the logo text 'ACME' in color #FF5733") rather than dropping a code loose in the prompt.
Practise on something you can actually run. Flux 3 AI is an independent browser workspace built for this waiting period — write prompts using the structure above, generate, iterate, and keep the ones that work. Open the image generator, or upscale your best frames for delivery. No waitlist, no request form.
Step 4 — Build assets FLUX 3 will actually use
Because FLUX 3's strongest modes are reference-driven, the highest-value preparation is not prompt text — it is a reference set:
- Character sheets. A consistent subject rendered in 4–6 angles and lighting conditions. Video-to-video and reference-to-video both carry identity from these.
- Style plates. One image per look you intend to reuse: lighting, palette, grain, lens character.
- Keyframes. If you know the beats of a shot, generate the first and last frames now. Keyframe-to-video exists specifically to interpolate defined moments.
- Product frames. Clean, well-lit, single-subject product shots that will survive being placed into a new scene.
Every one of these is producible today with an image model. None of them requires FLUX 3 access. All of them shorten your first FLUX 3 session from "experimenting" to "producing."
Common mistakes to avoid
- Porting negative prompts. They do nothing in the FLUX family. Rewrite them as positive descriptions.
- Burying the subject. A prompt that opens with three clauses of mood before naming the subject dilutes the thing you care about.
- Over-long prompts by default. Past ~80 words, you are usually adding noise, not control, unless the scene genuinely requires specification.
- Assuming FLUX.1 tier names apply. "Pro," "Ultra," "schnell," "dev" are FLUX.1/FLUX.2 vocabulary. FLUX 3's announced names are FLUX 3 Video, FLUX 3 Image, FLUX 3 Action, and FLUX 3 Dev.
- Waiting to start. The people who will get the most out of FLUX 3 in month one are the ones who arrive with a tested prompt library and a reference set.
FAQ
Can I use FLUX 3 right now? Only via BFL's early access request, and currently only the video capability.
Is there a FLUX 3 API? Not publicly. BFL's launch plan lists API access for video and image, each after an early access phase.
Do FLUX.2 prompts work on FLUX 3? The structural rules — subject-first ordering, no negatives, concrete optical detail — are family-level behaviour and should carry. FLUX 3-specific guidance has not been published yet.
Can FLUX 3 generate sound effects and speech? BFL lists native audio generation on all video outputs, plus multilingual dialogue, and cites associating sounds with physical events as an early strength.
How long can a FLUX 3 video be? Up to 20 seconds in a single generation. Longer sequences come from agentic chaining of clips.
Is there a free FLUX 3 tier? No pricing of any kind has been published.
Bottom line
You cannot "use FLUX 3" today in the way a tutorial normally means. What you can do is request access, learn the documented FLUX prompt rules, and build the reference assets the model's strongest modes depend on.
That is not a consolation prize — it is the actual work. Start building in the Flux 3 AI workspace, and by the time your access request clears, you will have a prompt library that is already tested.
Sources
- FLUX 3 — Real World Models (BFL, July 23, 2026) — access status, input modes, 20-second limit, launch plan
- FLUX 3 model page — early access request flow
- FLUX 3 x mimic (BFL, July 23, 2026) — partner access route for action prediction
- Official FLUX.2 prompting guide (BFL docs) — prompt structure, no-negatives rule, word order, length guidance, text and hex technique
- BFL documentation overview — current model family and recommended usage
- BFL API pricing — confirms no FLUX 3 pricing exists
- FLUX.2 model page — multi-reference behaviour that FLUX 3 builds on
- FLUX.2: Frontier Visual Intelligence (BFL, November 25, 2025) — reference-image capabilities and resolution limits
- black-forest-labs/flux on GitHub — reference inference code for the open models
- FLUX open weights licensing — access tiers for weight-level use
Scope note: Flux 3 AI is an independent creator workspace and is not affiliated with Black Forest Labs, nor does it provide FLUX 3 access. Official access, pricing, and terms are at bfl.ai.
