FLUX.3 Video on fal: Text and Image to Video at $0.085 per Second, With No Face Filter (October 2026)
Black Forest Labs' FLUX.3 video family covers text-to-video, image-to-video, keyframes, extension and a cheaper draft tier. What it costs, how long clips can be, and why it is the model we use when no real person is in frame.
FLUX.3 is Black Forest Labs' video family, launched on 4 August 2026 after a 23 July announcement. It is the model we use when a scene has no real person in it. It lists on fal at $0.085 per second for text-to-video and image-to-video at 720p, with a draft tier below that, and it has two properties that matter for our work: it has no real-face filter, and it supports frame-anchored generation (first and last frame, keyframes) that makes it controllable.
What is in the family
| Endpoint | Job |
|---|---|
| text-to-video | a clip from a prompt |
| image-to-video | animate a still |
| first-last-frame | start and end images, the model fills the movement between |
| keyframes | up to ten anchor frames |
| extend-video | continue an existing clip |
| draft variants | cheaper, lower-fidelity versions for testing a composition |
| edit-video | restyle an existing clip (covered in its own guide) |
Clips run from 5 to 20 seconds, or the model chooses the length; eight aspect ratios are available, and there is an option to generate audio. A safety_tolerance setting from 0 to 4 controls the content filter.
What it costs
| Length | 720p ($0.085 per second) |
|---|---|
| 5 seconds | $0.43 |
| 10 seconds | $0.85 |
| 20 seconds | $1.70 |
Resolutions of 1080p, 1440p and 2160p are produced through an upsampler and cost more; a reseller lists up to $0.53 per second at the top end, and fal's own page gives the 720p base. As with several newer models, check the price at the resolution you will publish. The draft tier is cheaper again: we read $0.03 per second in September, and other listings differ, so confirm before you budget.
The reason we use it: no real faces, no filter
Models from ByteDance refuse real human likenesses upstream of every reseller (why Seedance blocks real faces), and some others restrict named people. FLUX.3 video did not block real faces in our approved use. That cuts two ways. For a scene with nobody recognisable in frame, such as food, a street, a product, a landscape, it is a clean fit. For a scene with a real person, the absence of a filter is not a feature: the likeness and rights questions are entirely yours.
In our pipeline, the shots of a Reel that contain no people are not generated by a face-replace model at all. When a Reel needs a generated face-less scene, text-to-video from FLUX.3 is the model we route to, and where a reference frame exists, image-to-video continues from it. We approved it for that purpose on 16 August 2026.
Where it earns its place
- B-roll and inserts: a coffee pour, a skyline, a close-up of a product with soft light.
- Controlled transitions: first-last-frame lets you design the start and end images and let the model animate the middle.
- Longer ambient clips: up to 20 seconds in one call.
- Draft passes before a more expensive final.
Where it does not fit
A specific person's face. Use a model with a track record for faces (Kling 3.0 Pro, HappyHorse reference-to-video). Copying a specific choreography needs motion transfer (Motion Control). Replacing a person in an existing clip needs a video-edit model (replace a person).
The competition on price
MiniMax H3 Max lists text-to-video at a $0.0325 base for 768P (guide), about 2.6 times cheaper than FLUX.3 on paper. We have approved FLUX.3 for face-less scenes after one successful run, and H3 Max as a candidate that has not yet been run. If you are price-bound and have no face in frame, test H3 Max at the resolution you will publish; our own switch will depend on a result, not on a price list.
A workflow that uses its strengths
A good pattern for a face-less piece is to design the first and last frames as images (cheap, and you can iterate in cents), then use the first-last-frame endpoint to let the model animate between them. Run the draft tier first to check the movement, and only then run the full tier on the keeper. For longer pieces, extend the approved clip rather than generating a long single take, because each extension can be reviewed before the next is paid for. This keeps the expensive calls few and the decisions early.
Common questions
Does FLUX.3 video block real faces?
Not in our approved use. That is useful for scenes with nobody recognisable and a responsibility for consent when there is a person, since no filter protects anyone.
How long can a clip be?
From 5 to 20 seconds, or the model chooses a length.
What does the draft tier do?
It is a cheaper, lower-fidelity version for proving a composition before the full render. We read $0.03 per second in September; confirm before you budget.
Why is 1080p more expensive?
Resolutions above 720p run through an upsampler, and a reseller lists up to $0.53 per second at the top end. Price a clip at your target resolution first.
Can it extend a clip?
Yes, there is an extend-video endpoint for continuing an existing clip.
How to do this without code: Scenes with no person in them are the ones that need a generator like this, and the autopilot handles them. In AutoHustle this is the autopilot on a Reel with no person in it, where face-less scenes are kept from the original footage rather than regenerated (quoted in credits before you confirm the plan). See Faceless content and Recreate a viral video.
Related reading: FLUX.3 Image and Klein on fal: Per-Megapixel Pricing, Ten References and Where Each Tier Fits (October 2026) · LTX 2.5 on fal: Open-Weights 22B Video With Multishot and 4K, and a Price Unit Nobody Explains (October 2026)
For developers: the fal endpoints
| Model | fal endpoint |
|---|---|
| FLUX.3 text-to-video | blackforestlabs/flux-3/text-to-video |
| FLUX.3 image-to-video | blackforestlabs/flux-3/image-to-video |
| FLUX.3 first-last-frame | blackforestlabs/flux-3/first-last-frame-to-video |
| FLUX.3 keyframes | blackforestlabs/flux-3/keyframes-to-video |
| FLUX.3 extend-video | blackforestlabs/flux-3/extend-video |
| FLUX.3 draft (text and image) | blackforestlabs/flux-3/*/draft |
720p price $0.085 per second, read from fal on 18 October 2026. Clip length 5–20 s or auto; safety_tolerance 0–4; optional audio.