Flux 3 on Genfire: Black Forest Labs' Video Model With Audio, Keyframes and a Draft Mode

What Flux 3 Is

Flux 3 is Black Forest Labs' video generation model, now live on Genfire. It generates clips of 5 to 20 seconds at 720p or 1080p, and every render ships with synchronized audio at no extra cost — you don't pay a separate audio surcharge and you don't need a second pass to add sound.

Two things set it apart from the other video models in the catalog. First, it's not one endpoint but five, and Genfire routes between them automatically based on what you attach. Second, it has a Draft tier roughly 3.5x cheaper than full quality that hands back a cache you can later promote to a 1080p render of the exact same shot. That second point changes how you should work with it, and we'll get to the workflow below.

The full aspect-ratio list is unusually wide: auto, 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, and 9:16. The 21:9 and 2:1 options are rare in AI video and useful for cinematic widescreen and banner-shaped placements.

The Five Routing Modes

You pick one model — Flux 3 — and the inputs decide which endpoint runs. On the API, the routing order is fixed: a draft cache outranks everything, then pinned keyframes, then a first+last frame pair, then a source video, then a start image, then a bare prompt.

Text-to-Video

Prompt only. Set duration (whole seconds, 5–20, or auto and let the model pick a length), resolution, and aspect ratio. Audio is generated alongside the picture.

Image-to-Video

Attach a start image and the clip animates from it. Same duration and resolution controls. auto duration is still available here.

First-Last-Frame

Attach a first frame and a last frame, and Flux 3 interpolates the motion between them. This is the two-point special case of keyframes, and it needs an explicit duration — there's no auto here because the model has to know how far apart the two frames sit in time.

Keyframes-to-Video

This is the mode worth learning. You pin up to 10 images to exact positions on a 24 fps timeline. Each keyframe is an image plus a frame_index; frame 0 is the first frame, and the largest usable index is duration × 24 (a 10-second clip runs to frame 240). Indices must be unique — Genfire de-duplicates and clamps out-of-range indices before submitting, so a slip in the numbers won't burn a paid run.

Where it pays off:

• Storyboards. If you've already built a shot sequence in the AI storyboard generator, you have the frames. Pin frame 1 at index 0, frame 2 at 72, frame 3 at 144, frame 4 at 216 on a 10-second clip, and Flux 3 renders the transitions between your boards — with sound — rather than re-imagining the shot from a prompt. The boards become the spine of the video instead of a reference the model might drift from.
• Product turntables. Render or photograph the product at four or five angles, pin them at even intervals, and Flux 3 produces the rotation between angles. You control exactly which angles land on screen and when; the model only has to fill the arcs.
• Beat-locked edits. If a cut needs a specific image visible at a specific second — a logo at the 8-second mark, a reveal at 15 — the frame index puts it there. No prompting for timing and hoping.

In the Workflow editor, the Flux 3 Keyframes node has no timeline scrubber; it consumes the images wired into its reference port in order and distributes them across the clip according to a keyframe_spacing parameter (even by default). For hand-placed indices, use the API or MCP, where you pass the keyframes[] array directly.

Extend-Video

Attach a source clip and Flux 3 continues it. The source must be an MP4 of at most 15 seconds and 50 MB; fal enforces both. Extend is how you get past a single generation's length when a shot needs more runway, and because audio is native, the continuation carries sound too.

Draft Mode: Iterate Cheap, Then Finish

Every Flux 3 mode has a /draft twin: Flux 3 Draft, Flux 3 Keyframes Draft, and Flux 3 Extend Draft in the studio picker, or video.flux_3_draft on the API. Drafts render at a fixed 720p (the draft schemas expose no resolution control at all) and cost roughly 3.5x less than a full-quality render.

The important part is what a draft returns. Alongside the video you get a draft_cache_url — an encrypted bundle that encodes the shot's prompt, duration, aspect ratio, and the model's internal state. Pass that URL back to video.flux_3_draft and Genfire routes to the draft-enhance endpoint, which renders the same shot at full 1080p. Not a re-roll with the same prompt — the same shot. The cache is terminal: it carries its own settings, so you don't resend a prompt or duration, and if you pass a stray image alongside it the cache still wins.

That makes a very different workflow possible:

1. 1Generate several drafts of the shot. Change the prompt, move a keyframe, try a different aspect ratio.
2. 2Pick the take that works.
3. 3Enhance only that take to 1080p.

If you would normally burn four full-quality renders finding the shot, drafting first and enhancing once costs a fraction of that. It also lets you sanity-check the audio before committing, since drafts carry sound too.

Flux 3 vs Seedance 2.5 vs Wan 3.0

Three of the current audio-native video models overlap in the mid-length range. Here's how they differ on the dimensions that decide the choice:

The read:

• Pick Flux 3 when you need control over what appears when — storyboards, turntables, timed reveals — or when you want to extend a clip, or when you're going to iterate and want the Draft path. The 21:9 and 2:1 ratios are a bonus for cinematic and banner formats.
• Pick [Seedance 2.5](/seedance-2-5) when the shot is built from references: a character, a location, a voice, a product you need to hold consistent across a scene. Its reference pools are the largest in the catalog, and it runs to 30 seconds in one pass.
• Pick [Wan 3.0](/wan-3) when you want the cheapest 1080p with sound, need clips as short as 2 seconds, or want reference-driven generation at a smaller scale. Wan 3.0 Prime is the same model faster and sharper at 1.4x the price.

None of these three has a 4K tier. If the deliverable is 4K, look at the Kling or Seedance 2.0 lines instead.

How to Use Flux 3 on Genfire

Flux 3 runs everywhere Genfire's video models run: the browser studio, the REST API, the MCP server, the CLI, and the Workflow editor.

In the Video Studio

Open the AI video generator and pick Flux 3 from the model dropdown. The family expands into six entries:

• Flux 3 — auto-detects text-to-video, image-to-video, or first-last-frame from what you attach
• Flux 3 Draft — the same, at 720p and ~3.5x cheaper
• Flux 3 Keyframes and Flux 3 Keyframes Draft — pinned frames
• Flux 3 Extend and Flux 3 Extend Draft — continue a clip

Duration is a whole-second picker from 5 to 20; auto appears on the plain and extend modes and is hidden on first-last-frame and keyframes, which need a fixed length. Resolution shows 720p/1080p on the full-quality entries and is hidden on drafts, because the draft endpoints have no resolution parameter. The studio hides what the endpoint can't take, so there's nothing to misconfigure.

On the API

Two model ids: video.flux_3 and video.flux_3_draft. Which endpoint runs depends on the fields you populate:

• prompt only → text-to-video
• prompt + image_url → image-to-video
• prompt + first_frame_url + last_frame_url → first-last-frame
• prompt + keyframes: [{ image_url, frame_index }, …] → keyframes-to-video (max 10; frame_index ≤ duration × 24)
• prompt + source_video_url → extend-video
• draft_cache_url alone, on video.flux_3_draft → draft-enhance to 1080p

A Draft run's response includes draft_cache_url. Store it with the run; it's the handle you pass back to enhance. The developer docs cover the full request shape and the model listing endpoint that returns Flux 3's limits (max_keyframes, source_video_max_seconds, source_video_max_mb) at runtime.

Through MCP

Connect the Genfire MCP server (genfire mcp setup from the CLI configures Claude Code, Claude Desktop, or Cursor) and call genfire_generate_video with model: "video.flux_3" or "video.flux_3_draft". The tool accepts the same keyframes[] array and draft_cache_url field as the API, and genfire_recommend_models will steer you to Flux 3 when you tell it you need multi-point keyframes. Local images can be attached as keyframes — the server uploads them before the run.

From the CLI

Install @genfire/cli and run the plain modes directly:

genfire generate video "a ceramic mug rotating on a walnut table, soft window light" \
  -m video.flux_3_draft -d 8 -o mug-draft.mp4

genfire generate video "a ceramic mug rotating on a walnut table, soft window light" \
  -m video.flux_3 -d 8 -r 1080p -i mug.png -o mug.mp4

-i sets the start frame and accepts a local path or an https:// URL. genfire cost video "..." -m video.flux_3 -d 8 quotes the credit cost before you run.

Practical Notes

• Duration is whole seconds. 5 through 20. The auto option exists on text-to-video, image-to-video, and extend, but not on first-last-frame or keyframes.
• Drafts are 720p, full stop. There's no resolution to set. If you want 1080p, enhance the draft or run full quality from the start.
• A draft cache encodes its settings. Enhance doesn't take a new prompt, duration, or aspect ratio. Change those at the draft stage.
• Keyframe indices are absolute frames at 24 fps. Work out seconds × 24 for each pin, keep them unique, and stay at or below duration × 24.
• Extend sources are MP4 only, ≤15s, ≤50MB. Trim or compress before submitting.
• Audio is always on. There's no toggle; it's part of the render.

Getting Started

1. 1Open the Video Studio and select Flux 3 Draft.
2. 2Write a prompt, set a duration, and generate a few takes at 720p.
3. 3When one lands, promote it: via the API or MCP, pass its draft_cache_url back to video.flux_3_draft for the 1080p render.
4. 4For multi-frame shots, switch to Flux 3 Keyframes, attach your frames, and pin them.
5. 5Need more runway? Feed the result to Flux 3 Extend.

Genfire is pay-as-you-go. Credit packs start at $19 for 1,000 credits, purchased credits last 12 months, every purchase includes a commercial license, and optional monthly plans start at $29/month. The pricing page has the full breakdown, and the studio quotes the exact credit cost of every Flux 3 run before you commit.

Read the model page at /flux-3 for specs and prompt ideas, or head straight to the AI video generator and draft your first shot.