FLUX-3 video generation

Black Forest Labs' FLUX-3 video model, with synchronized audio generated by default. The operation is selected by an explicit operation discriminator — fields invalid for that operation are rejected with a 400.

`operation`	Required inputs	What it does
`textToVideo`	`prompt`	Generate a clip from a text prompt.
`imageToVideo`	`prompt`, `image`	Animate a single source image as the first frame.
`firstLastFrameToVideo`	`prompt`, `startImage`, `endImage`	Generate a clip that starts on one image and ends on another.
`keyframesToVideo`	`prompt`, `keyframes` (1–10)	Pin up to 10 images to exact frame positions and generate the motion between them.
`extendVideo`	`prompt`, `sourceVideo`	Continue an existing clip from its final frames.
`draftEnhance`	`prompt`, `draftCache`, `duration`	Re-render a previously generated draft at full-quality 1080p.

Default choice: resolution: "720p", duration: 5, generateAudio: true. FLUX-3 jobs exceed the 100-second request timeout — always submit with wait=0.

Flux image vs FLUX-3 video

This page covers the videoGen engine (engine: "flux"). For Flux image generation see Flux 2 and Flux 1.

Drafts

Every generation operation accepts draft: true, which renders a fast 720p preview at roughly a third of the standard rate. A draft response carries an extra draftCache blob alongside the video; passing that blob's url to draftEnhance re-renders the same clip at full-quality 1080p.

This makes iteration cheap: explore with drafts, then pay the full rate once on the take you want.

textToVideo (draft: true)  →  output.video          720p preview,  390 Buzz at 5s
                           →  output.draftCache     opaque bundle
                                     ↓
draftEnhance (draftCache)  →  output.video          1080p final,  1885 Buzz at 5s

Two constraints follow from how drafts work:

Drafts always render at 720p. Sending draft: true with resolution: "1080p" is rejected with a 400 — take the draft at 720p and enhance it instead.
draftEnhance needs the draft's duration restated on the request, because the cache bundle is opaque and the duration cannot be recovered from it. Pass the same value you used for the draft.

The request shape

Every FLUX-3 request is a single videoGen step on SubmitWorkflow. Three keys select which leaf schema the rest of the body is validated against:

json

{
  "$type": "videoGen",
  "input": {
    "engine":    "flux",
    "version":   "v3.0",
    "operation": "textToVideo"
  }
}

"v3.0" is currently the only release. Send it explicitly — it also pins you against future versions changing the defaults under you.

textToVideo

http

POST https://orchestration.civitai.com/v2/consumer/workflows?wait=0
Authorization: Bearer <your-token>
Content-Type: application/json

{
  "steps": [{
    "$type": "videoGen",
    "input": {
      "engine": "flux",
      "version": "v3.0",
      "operation": "textToVideo",
      "prompt": "A little girl walking on a road at sunset, cinematic lighting, smooth camera movement",
      "aspectRatio": "16:9",
      "resolution": "1080p",
      "duration": 5
    }
  }]
}

Set your Civitai API token via the Token button in the navbar to enable Try It.

Request body — edit to customize (e.g. swap the image URL or prompt)

Valid JSON

As a draft

Same request with draft: true and no resolution:

json

{
  "engine": "flux",
  "version": "v3.0",
  "operation": "textToVideo",
  "prompt": "A little girl walking on a road at sunset, cinematic lighting, smooth camera movement",
  "aspectRatio": "16:9",
  "duration": 5,
  "draft": true
}

Set your Civitai API token via the Token button in the navbar to enable Try It.

Request body — edit to customize (e.g. swap the image URL or prompt)

Valid JSON

Feed the resulting output.draftCache.url to draftEnhance when you want the full-quality render.

imageToVideo

Pass a single image used as the first frame.

json

{
  "engine": "flux",
  "version": "v3.0",
  "operation": "imageToVideo",
  "prompt": "Camera slowly pushes in",
  "image": "https://image.civitai.com/.../first-frame.jpeg",
  "resolution": "1080p",
  "duration": 5
}

Set your Civitai API token via the Token button in the navbar to enable Try It.

Request body — edit to customize (e.g. swap the image URL or prompt)

Valid JSON

firstLastFrameToVideo

Generate the motion between two fixed frames.

json

{
  "engine": "flux",
  "version": "v3.0",
  "operation": "firstLastFrameToVideo",
  "prompt": "A smooth transition between the two shots",
  "startImage": "https://image.civitai.com/.../first.jpeg",
  "endImage": "https://image.civitai.com/.../last.jpeg",
  "resolution": "720p",
  "duration": 5
}

Set your Civitai API token via the Token button in the navbar to enable Try It.

Request body — edit to customize (e.g. swap the image URL or prompt)

Valid JSON

keyframesToVideo

Pin 1–10 images to exact positions in the generated video. The output runs at 24 fps, so frameIndex must be between 0 and duration × 24 — at duration: 5 the last valid index is 120. Every frameIndex must be unique.

json

{
  "engine": "flux",
  "version": "v3.0",
  "operation": "keyframesToVideo",
  "prompt": "A slow pan across the scene",
  "keyframes": [
    { "image": "https://image.civitai.com/.../a.jpeg", "frameIndex": 0 },
    { "image": "https://image.civitai.com/.../b.jpeg", "frameIndex": 120 }
  ],
  "resolution": "720p",
  "duration": 5
}

Set your Civitai API token via the Token button in the navbar to enable Try It.

Request body — edit to customize (e.g. swap the image URL or prompt)

Valid JSON

extendVideo

Continue an existing clip from its final frames. The source must be an MP4 under 15 seconds and under 50 MB; duration is the length of the new footage.

json

{
  "engine": "flux",
  "version": "v3.0",
  "operation": "extendVideo",
  "prompt": "The camera keeps drifting forward as the sun sets",
  "sourceVideo": "https://image.civitai.com/.../clip.mp4",
  "resolution": "720p",
  "duration": 5
}

extendVideo re-renders the source alongside the continuation, so it bills at roughly 2.4× the standard rate — see Cost.

Supply your own MP4

There's no live widget here — the operation needs an MP4 under 15 seconds, and the shared sample clips are longer. Point sourceVideo at your own short MP4 to try it.

draftEnhance

Take output.draftCache.url from a completed draft and submit it back. prompt is required for moderation but does not steer the render — the enhance reproduces exactly what the draft previewed, at 1080p.

json

{
  "engine": "flux",
  "version": "v3.0",
  "operation": "draftEnhance",
  "prompt": "A little girl walking on a road at sunset, cinematic lighting, smooth camera movement",
  "draftCache": "https://orchestration.civitai.com/v2/consumer/blobs/abc123.draftcache?...",
  "duration": 5
}

Restate the draft's duration — the cache is opaque, so the value can't be recovered from it, and it determines the price.

Swap in your own draft-cache url

draftCache above is a placeholder. There's no live widget for this operation because the url has to come from your own completed draft run. Draft-cache urls are signed and expire like any other blob url — refetch the draft workflow for a fresh one rather than storing it long-term.

Parameters

Shared across operations unless noted. The per-operation schema in the API reference is authoritative.

Field	Default	Used by	Notes
`engine`	— ✅	All	`"flux"`
`version`	— ✅	All	`"v3.0"` is the only release.
`operation`	— ✅	All	See the table above.
`prompt`	— ✅	All	Required on every operation, including `draftEnhance` (moderation only there).
`duration`	`5`	All	Integer seconds, 5–20. Required on `draftEnhance` — restate the draft's value.
`resolution`	`"720p"`	All generation ops	`"720p"` or `"1080p"`. Rejected alongside `draft: true`; `draftEnhance` always outputs 1080p.
`aspectRatio`	`"auto"`	All generation ops	`"auto"`, `"21:9"`, `"2:1"`, `"16:9"`, `"4:3"`, `"1:1"`, `"3:4"`, `"9:16"`. `"auto"` lets the model choose.
`generateAudio`	`true`	All generation ops	Set `false` for a silent clip. No cost difference.
`draft`	`false`	All generation ops	Fast 720p preview that also returns `draftCache`. See Drafts.
`image`	— ✅	`imageToVideo`	Single image used as the first frame.
`startImage` / `endImage`	— ✅	`firstLastFrameToVideo`	First and last frame.
`keyframes[]`	— ✅	`keyframesToVideo`	1–10 `{ image, frameIndex }` entries, unique `frameIndex` ≤ `duration × 24`.
`sourceVideo`	— ✅	`extendVideo`	MP4 url, under 15 s and 50 MB.
`draftCache`	— ✅	`draftEnhance`	`output.draftCache.url` from a prior draft.

FLUX-3 takes no seed — the model does not accept one. Identical requests therefore resolve to the same cached result; vary the prompt if you want a different take.

Cost

Billed per output second in Buzz on the workflow's transactions. Use whatif=true for an exact preview; see Payments (Buzz) for currency selection.

total = buzzPerSecond × duration

Operation	720p	1080p	Draft (720p)
`textToVideo`, `imageToVideo`, `firstLastFrameToVideo`, `keyframesToVideo`	221 Buzz/s	377 Buzz/s	78 Buzz/s
`extendVideo`	533 Buzz/s	689 Buzz/s	156 Buzz/s
`draftEnhance`	—	377 Buzz/s	—

Totals at duration: 5: 1 105 / 1 885 / 390 for the standard operations, 2 665 / 3 445 / 780 for extendVideo, and 1 885 for draftEnhance.

A draft-then-enhance round trip at 5 s costs 390 + 1 885 = 2 275 against 1 885 for going straight to 1080p — so drafts pay off when you expect to discard takes, not when you already know what you want.

Reading the result

A standard videoGen step output, plus draftCache when the step ran a draft:

json

{
  "status": "succeeded",
  "steps": [{
    "name": "0",
    "$type": "videoGen",
    "status": "succeeded",
    "output": {
      "video": { "id": "blob_...", "url": "https://.../signed.mp4" },
      "draftCache": { "id": "abc123.draftcache", "url": "https://.../signed.bin" }
    }
  }]
}

draftCache is absent on non-draft runs. Its contents are opaque and encrypted — the only thing to do with it is pass its url to draftEnhance.

Blob urls are signed and expire — refetch the workflow or call GetBlob for a fresh url.

Long-running jobs

FLUX-3 jobs typically take a few minutes, longer at 1080p and at higher durations; drafts are substantially faster. All exceed the 100-second request timeout — submit with wait=0 and:

Webhooks (recommended): register a callback with type: ["workflow:succeeded", "workflow:failed"] — see Results & webhooks.
Polling: GET /v2/consumer/workflows/{workflowId} on a 10 s → 30 s → 60 s cadence.

Troubleshooting

Symptom	Likely cause	Fix
`400` with unknown field	Field isn't valid for this `operation`	Each operation maps to its own typed schema; check it via `SubmitWorkflow`.
`400` "FLUX-3 drafts always render at 720p"	Sent `draft: true` with `resolution: "1080p"`	Drop `resolution` (or use `"720p"`), then use `draftEnhance` for the 1080p render.
`400` "keyframesToVideo requires between 1 and 10 keyframes"	`keyframes` was empty or had >10 entries	Provide 1–10.
`400` "Keyframe frameIndex values must be unique"	Two keyframes shared a position	Give each a distinct `frameIndex`.
`400` "Keyframe frameIndex must be between 0 and duration × 24"	Index past the end of the clip	The video is 24 fps — cap indexes at `duration × 24`, or raise `duration`.
`400` "requires a source video shorter than 15 seconds"	`sourceVideo` too long	Trim the clip before submitting.
Step `failed` after a `draftEnhance`	Draft-cache url expired	Refetch the draft workflow for a fresh `draftCache.url` and resubmit.
Step `failed`, `reason = "no_provider_available"`	Provider queue busy	Retry shortly.
Step `failed`, `reason = "blocked"`	Moderation rejected the input or output	Re-prompt with different content. FLUX-3 jobs run with the provider's safety filter at its most permissive setting and Civitai moderates instead, so there is no knob to loosen — though the model still applies a floor of its own.

SubmitWorkflow — operation used by every example here
GetWorkflow — for polling
Results & webhooks — production result handling
Full parameter catalog: the Flux3V3<Operation>Input schemas in the API reference — Flux3V3TextToVideoInput, Flux3V3ImageToVideoInput, Flux3V3FirstLastFrameToVideoInput, Flux3V3KeyframesToVideoInput, Flux3V3ExtendVideoInput, Flux3V3DraftEnhanceInput
Happy-Horse video generation — comparable multi-operation commercial video model
MiniMax H3 video generation — 2K native video with reference image / video / audio
Veo 3 video generation — Google's multi-mode video model
Flux 2 image generation — the Flux family's image models

FLUX-3 video generation ​

Drafts ​

The request shape ​

textToVideo ​

As a draft ​

imageToVideo ​

firstLastFrameToVideo ​

keyframesToVideo ​

extendVideo ​

draftEnhance ​

Parameters ​

Cost ​

Reading the result ​

Long-running jobs ​

Troubleshooting ​

Related ​

FLUX-3 video generation

Drafts

The request shape

textToVideo

As a draft

imageToVideo

firstLastFrameToVideo

keyframesToVideo

extendVideo

draftEnhance

Parameters

Cost

Reading the result

Long-running jobs

Troubleshooting

Related