FLUX 3 Max: ten seconds of video generated in twelve seconds, without sacrificing quality

Black Forest Labs and fal introduce FLUX 3 Max, a video model capable of generating ten seconds in 12.45 seconds, complete with audio and keyframe control.

Generating a ten-second sequence in just over twelve seconds: with FLUX 3 Max, Black Forest Labs and fal aim to reduce the wait between an idea and its video result. The model, developed from FLUX 3 and then adapted by fal’s research teams, shows a generation time of 12.45 seconds in the comparison highlighted by both companies, compared to 277.8 seconds for Seedance 2.5. This approximately 22-fold difference is accompanied by a ranking among the top image-to-video generation models in Artificial Analysis’s evaluation. Beyond speed, FLUX 3 Max offers three creation methods, from a simple prompt to the use of multiple keyframes, with integrated sound generation. Speed becomes a creative argument

In AI video production, generation latency directly influences the way people work. A scene that takes several minutes to render limits experimentation opportunities, especially when comparing different directorial intentions, correcting a camera movement, or modifying a character's behavior. Reducing this time to a few seconds changes the iteration pace, without necessarily guaranteeing that every generation will be usable.

The positioning of FLUX 3 Max relies on this trade-off between speed and quality. The model is built on FLUX 3, developed by Black Forest Labs, to which fal Research applied additional training and inference optimizations. The partners thus distinguish the original model from its Max variant, which is designed to produce videos faster while improving prompt adherence and visual quality.

In its communication, Black Forest Labs announces the generation of a ten-second video in 12.45 seconds, compared to 277.8 seconds for Seedance 2.5. The ratio is approximately 22.3 times. Technical specs from fal also mention a time close to eleven seconds, corresponding to an acceleration greater than 25 times. These values stem from measurement configurations presented separately and should not be confused with a guaranteed time for all requests. Resolution, duration, reasoning level, and service conditions can significantly alter delays. A top-ranking position in image-to-video

Visual performance does not rely solely on the two companies' demonstrations. In Artificial Analysis's AA-Video-I2V v1.1 leaderboard, dedicated to animating an image into video, FLUX 3 Max achieves an Elo score of 1,138, compared to 1,140 for MiniMax H3 and 1,138 for Seedance 2.5. Confidence intervals overlap, placing these models in a statistically close group without establishing a clear superiority of FLUX 3 Max over each of its competitors.

The model also appears in Artificial Analysis's text-to-video ranking, with an Elo of 1,159, identical to that of Seedance 2.5 in the consulted results. It sits behind Wan 3.0 and Utopai X in this general ranking. The claim of a tied first place must therefore be placed in the context of the image-to-video evaluation, and not interpreted as dominance across all video generation categories.

Black Forest Labs and fal notably highlight the handling of voices, sound effects, physical movement, and unusual prompts. These capabilities are important for narrative sequences and less conventional visual concepts, but their quality remains dependent on the requested scene. Elo scores measure preferences gathered during head-to-head comparisons between generations; they guarantee neither physical accuracy nor the consistency of every result. Three methods to build a sequence

FLUX 3 Max is available on fal through three generation interfaces that meet different creative needs. The first, Text-to-Video, produces a scene directly from a written description. The prompt can specify characters, their actions, camera movements, lighting, and the sound environment, without requiring an initial image.

The second, Image-to-Video, uses an image as a starting point. This approach allows animating a photograph, illustration, or pre-defined composition while describing the expected movements. An optional end image can also serve to guide the outcome of the sequence. For studios with a precise artistic direction, this workflow avoids leaving the framing and initial appearance of elements entirely up to the model.

The third interface, Keyframes-to-Video, extends this principle to multiple moments within the same video. It accepts up to ten keyframes, each associated with a position on a 24-frames-per-second timeline. A prompt describes the actions and transitions that should connect these references. This method notably allows preparing a scene entrance, a visual transformation, or a progression between several compositions, without having to independently generate each intermediate shot.

Keyframes remain generation guides, rather than images that must be reproduced identically. The model interprets movements between reference points, along with the variations and imperfections that this may entail. Five to twenty-second videos with audio

All three interfaces support sequences from five to twenty seconds, with duration adjustable in one-second increments. The standard definition is 720p, while a 1080p option relies on upscaling the produced video. Several aspect ratios are available, including 16:9, 9:16, 1:1, 4:3, and 21:9, to adapt generations to editorial, advertising, or social platform uses.

Audio generation is enabled by default, but can be turned off. Instructions can describe dialogue, ambient sound, or sound effects associated with visible actions. The model also offers two levels of reasoning effort: low and high. The latter aims for better scene interpretation at the expense of longer generation times, which adds nuance to comparisons based solely on speed.

For creators, the main advantage lies in the ability to generate multiple variations of a single concept. A reference image can be kept while movement instructions change, or several keyframes can be reorganized to alter the sequence's flow. The announced speed reduces the time cost of these trials, without removing the work of selection, editing, and