Three modes, one endpoint: text-to-video, image-to-video, video extension

All three modes actually use the same endpoint. Every request goes to POST /v1/flux-3-video with an identical request body. The only difference is the mode field, which simply tells the API what you're working with:

The docs pull a neat trick: they walk one scenario through all three modes. First, use t2v to generate "she takes his hand and pulls him laughing through the lantern-lit alley." Then grab that video's first frame and feed it to i2v with a new camera instruction, getting a completely different video with the same opening. Finally, feed the whole original clip to v2v and have the couple burst out of the alley into a night market. Each request feeds on the previous one's output, so the relationship between modes clicks immediately.

curl -X POST https://api.bfl.ai/v1/flux-3-video \
  -H "x-key: $BFL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "mode": "t2v",
    "prompt": "she takes his hand and pulls him laughing
               through the lantern-lit alley, the camera
               chasing them, paper lanterns swaying overhead",
    "duration": 8
  }'

Let's start with the most basic text-to-video. The official demo doesn't ease you in: a whole house is mounted on a crash-test sled and slammed into a concrete wall.

Now video extension. The rules: you provide a clip up to 4 seconds long, tell it what happens next, and it continues the motion, camera, dialogue, and audio. The official demo marks the seam directly on screen: the first 4 seconds are the original footage, everything from second 4 onward is generated. A dashcam drive down a coastal boulevard, and then a moose runs out:

Keyframes: pin up to 10 images, the model fills the gaps

Image-to-video works differently here than elsewhere — there's no separate "first frame" field. All images go into keyframes — how many you provide and where you pin them determines what you're doing.

• Provide 1 image: it becomes the opening frame.
• Provide 2 images: one at the start, one at the end, and the model invents the journey between them.
• Provide up to 10 images: you're handing over a storyboard.
• Use [seconds, image] pairs: pin a specific image to second 4, for example.

The official demo has a built-in timeline, showing the mechanism directly: three keyframes pinned at 0:00, 0:04, and 0:10, with the caption "FLUX 3 interpolates the keyframes" — FLUX 3 fills the frames between the keyframes. In the third thumbnail, the little red sailboat is already capsized. You just set three points; the how — the water coming in, the tipping over — is its story to invent.

The start-and-end-only version shows how much it makes up on its own: START FRAME on the left is an empty room, END FRAME on the right is a furnished living room, and in between the caption reads FLUX 3 fills the in-between. Furniture unfolds from the floor like origami — none of that morphing was specified by anyone.

Native audio: because sound and video are generated together, lip-sync works

AI videos have always fallen apart the moment someone opens their mouth, and it's the approach that's to blame. The old way is a three-stage pipeline, each step waiting on the last:

• Stage 1: generate the visuals. The model doesn't know the dialogue's rhythm, word stress, or pauses — it hasn't heard anything yet.
• Stage 2: run the dialogue through a separate TTS model. This model can't see the video; it just makes the words sound human.
• Stage 3: try to sync the lips to the audio afterward. By now you're patching — good enough for a rough match, but not syllable-accurate.

Two seams sit between the three stages, and either can drift. That's why the old approach dreaded face close-ups: lips half a beat off, and the audience is instantly pulled out of the scene.

The API is also designed with "audio by default:" leave generate_audio unset and it's true — you get a video with sound, no extra flag needed.

For languages, thirteen are named: English (multiple accents), Chinese, Spanish, French, German, Japanese, Portuguese, Russian, Italian, Indonesian, Turkish, Hindi, and Punjabi, followed by an "etc." Korean isn't on the named list, and how far that "etc." extends isn't specified in the materials.

The dialogue demo shows two men in suits arguing into vintage microphones across a table — a 20th-century TV debate, running a full 20 seconds without a break:

The multilingual clip is more interesting for another reason: the same woman in a red raincoat appears in three scenes, speaking in each — a restaurant booth, a night-market food stall, and a laundromat. The same face carries across the cuts without drifting.

Multi-shot transitions, and the real places it knows

Across a 20-second clip, a black cat jumps from a bedroom bed to a sofa, then the shot cuts to outside, looking back at the house's window from the street — the cat is still there — then cuts again to the garden and shrubs. The camera genuinely changes position and cuts that still connect — not just pans and zooms within one shot. Getting this effect before would have meant generating several clips and editing them yourself:

Another demo shows off what BFL calls "world knowledge + real-world grounding." The idea: a few words of prompt can produce a documentary-style or educational clip. The footage looks like it was shot in Freiburg's old town: the narrow street gutters locals call Bächle, a red-sandstone Gothic cathedral, a pair of tourists with cameras. The whole thing feels like an 80s or 90s travel video. The clip isn't labeled with a location, but Freiburg happens to be BFL's own headquarters — the "Black Forest" in the company name is the region right there.

Draft Mode: a third of the price, then re-render your pick exactly as-is

Anyone who's tried video generation knows its maddening quality: you write a prompt, get a decent result, and want another one that's "similar but better." Re-submitting is a completely new generation. Subject, composition, and motion can all change — the one you liked might never come back. And every roll of the dice costs money.

Draft Mode splits this into two steps:

Step one: add draft: true to the request and get a quick preview. Drafts run at 1x time and 1/3 cost; full quality takes 1.8x time. Those three numbers are literally printed on the comparison video's frame. At that price, you can afford to roll a few.

Step two: once you've picked one, the draft result includes a draft_cache package. Download it, set mode to draft_enhance, and send it back. The model re-renders that same generation at full quality.

"The same shots, the same seed, nothing re-interpreted."

Black Forest Labs developer docs (translated by this site)

So Draft Mode isn't just a discount. It splits exploration from delivery into two separate costs: exploring costs a third per roll, and only the one you approve gets full price — and what you pay full price for is exactly what you approved.

Price sheet: video extension at $0.43/sec — 2.5x the base rate

Full pricing for all three modes is in the docs:

Three things jump right off the table:

• Video extension is significantly pricier. At HD, extension costs $0.43/sec — more than 2.5x the base $0.17/sec. A 10-second video costs $1.70 from scratch, but $4.30 for extension. Why? The docs don't say.
• Video extension is also shorter. The other two modes go up to 20 seconds. Extension maxes out at 15 — a full 5 seconds less.
• Drafts are always rendered in hd — there's no fhd draft option.

Let's translate that into a concrete example. One 10-second HD text-to-video:

• Go straight to full quality: $1.70, but you get one result. Don't like it? Roll again.
• Blind-roll 3 full-quality versions and pick one: $5.10.
• Roll 3 drafts, pick one, re-render: 3 × 0.6 + 1.7 = $3.50, and the final version is the one you picked.

One more trap to watch for: generated result links are signed and expire after about 2 hours. Download your videos promptly — don't wait until the next day.

Scores: first in text-to-video, ties Seedance 2.0 in image-to-video

BFL published a human-preference ELO chart, computed from pairwise comparisons. Both full rankings are in the image:

Text-to-video (left): FLUX 3 T2V 1135, Gemini Omni Flash 1090, Minimax H3 1082, Seedance 2.0 1069, Kling v3 Pro 1054… Image-to-video (right): FLUX 3 I2V 1051, Seedance 2.0 1049, Minimax H3 1039, Grok Imagine v1.5 1023, Gemini Omni Flash 1022… Source: Black Forest Labs

The leads aren't in the same league:

The 45-point lead in text-to-video is meaningful. In image-to-video, it's a mere 2 points ahead — and the official text itself uses the word "ties" Seedance 2.0, not "wins."

Two more things worth noticing on that chart:

• Veo 3.1 isn't on the text-to-video leaderboard. It appears at #8 (975) on the image-to-video side, but isn't among the ten entries on the text-to-video chart. The materials don't explain why.
• The comparison used Seedance 2.0. ByteDance released Seedance 2.5 on July 31st, and it's not included here.

One important caveat: this benchmark was organized by BFL itself, and it puts BFL's own model in first place. It's not a third-party evaluation. The demo videos above are embedded as-is, so you can judge for yourself.

Not-yet-released features, and a few restrictions

The docs' first screen states that FLUX 3 is in preview status. Two features are explicitly marked "coming soon": video editing, and Omni Reference, which uses images and videos as reference material to control generation.

Further down the roadmap are FLUX 3 Image (image generation and editing) and FLUX 3 Dev (open-weight version). Open weights remain dead last, consistent with the July announcement. The launch tweet also mentions 2K and 4K are on the way.

On safety: BFL engaged a third-party firm, Cinder, for pre-release evaluation covering risks like Non-Consensual Intimate Imagery (NCII) and Child Sexual Abuse Material (CSAM). The API has a safety_tolerance parameter from 0 to 4, defaulting at 2, with 0 being strictest. Three hard caps are written into the docs regardless of what you set:

• Sexual content is capped at level 3;
• Hate content is capped at level 2;
• Any request containing a reference image or video is capped at 2.

The launch page is titled "FLUX 3 Video, Part 1: Generation." There's no architecture, training methodology, or technical report anywhere — only the capabilities and the API. Exactly how the model achieves synchronous audio-video generation isn't something you can verify from published materials yet.

AI videos always fell apart the moment someone spoke. The old approach was a three-stage relay: generate video, then TTS, then sync — with two seams that could each drift half a beat. That's why close-ups were its nemesis. FLUX 3 folds all three into a single generation, and the API is audio-on-by-default: leave generate_audio unset and it's on.

The most annoying thing about video generation: you roll something decent, want "similar but better," re-submit, and get a completely new generation — subject, composition, and motion can all change, and the one you liked may never come back. Draft Mode splits this into two steps.

All pricing is public. One detail isn't on the launch page and only lives in the developer docs: if you have an existing clip and want to extend it, the rate jumps to $0.43/sec.

The launch page includes a human-preference ELO chart (relative scores from pairwise comparisons; it only shows which model was preferred more often, not absolute quality). The two leads aren't in the same league.