FLUX 3: Image, Video and Robot Actions

The FLUX 3 products share a multimodal foundation. Use the hosted Image and Video endpoints for creative work, and the Action checkpoints for robot or game controls.

Start from the route below. The broader FLUX 3 Dev backbone has a separate planned rollout; the currently downloadable Action checkpoints provide task-specific controls.

BFL released Image Klein on October 15 with downloadable weights and a separate /v1/flux-3-image-klein API. The 5.6B image transformer uses additional encoder and VAE components. Choose the distilled 16-step checkpoint for generation and editing, or Klein Base for adaptation and research.

The direct BFL API charges $0.009 for a 1K image, with the same rate for generation and editing. This is a separate product from the full Image endpoint above. The Klein guide covers resolution prices, complete local workflows and its FLUX Kommunity License conditions.

Vannarot Roeung’s October 1 Creative AI News analysis inspected BFL’s published edit samples rather than generating a new test set. The nine showcase masks varied from 67.8% to 94.1% identical pixels, with some changes away from the requested subject. Separate examples on the editing page preserved far more of the area outside the edit box.

The practical distinction is between an edit that looks localized and an edit that preserves every surrounding pixel. When a product packshot or brand layout must remain exact, compare the result with the original and composite the accepted edit back into a controlled region. The amount of preserved background varied across the sample groups.

Use bounding boxes to describe placement, then verify the output against the actual brief. For brand assets, pay particular attention to logo edges, lettering and background texture.

OpenArt community user ra13it generated the same prompt at four output sizes, then compared unscaled 512-pixel crops. The author saw finer detail at higher resolutions, with a 5456×3072 top-tier file. The crops show the extra detail delivered in that scene.

The same author changed a garment, a cover word and a brand name on a can. Reported unchanged-pixel shares were 75% for the garment and 99% for each text edit, measured per channel with a tolerance of 8/255. The tolerance allows small pixel changes, and the edit’s area affects the percentage. The author found naming individual background elements more useful than saying only “keep the background.”

A ten-reference composition kept the intended face, but a wool texture meant for the knit appeared on the jacket. Assign each reference a role and check where its details actually land. A reference can influence the wrong nearby object even when the overall composition looks convincing.

The Image endpoint accepts a prompt and up to ten reference images. Describe which reference supplies the subject, style or background. For layouts, append bounding boxes on a 0–1000 coordinate grid; each box uses top, left, bottom and right coordinates. Output choices range from 768sq to 4k.

This is useful when a poster needs an object in a specific region or a product photo needs one detail changed. Keep the original, inspect areas outside the requested edit, and read generated text at the delivery size. Composite the accepted region onto the original when the surrounding pixels must stay exact.

For a first trial, keep the default 1k resolution and ask for one composition. The grounding option defaults to web and image search; set it to false when you want the result based only on your supplied prompt and references. The request returns a polling URL; download the image once it is ready, before its one-hour result link expires.

Video supports text-to-video, image-to-video, keyframes and continuation, with optional generated audio. Text and image generation run from 5 to 20 seconds; continuation adds 5 to 15 seconds. The API documentation says HD is the default, with FHD, QHD and UHD finished through a video upsampler.

For exploratory work, generate an HD draft and keep its draft_cache bundle. Enhance the draft you approve to preserve that interpretation of the scene. Submitting the prompt again as a fresh full render can produce a different take, so the cached enhancement path matters when you have already approved the motion.

Specify duration, dialogue and framing, then inspect the generated file’s actual length, audio and continuity. Save the accepted output and generation settings promptly; the Video overview says signed result links expire about two hours after completion.

BFL’s launch evaluation used ten-second, 720p text-to-video clips with audio. It reported preferences for FLUX 3 over Kling v3 Pro in 60% of comparisons and Seedance 2.0 in 52%, and called the results preliminary.

These publisher-run comparisons concern older competitor versions and one generation mode. The 52% result is close to an even split; a current buying comparison needs Kling 4.0 and Seedance 2.5 at matched settings. Use your own dialogue, keyframes or animated-text brief, and include audio fixes and rerolls in the cost.

As checked October 3, BFL’s documented text/image-to-video rates are $0.06 per output second for HD draft, $0.17 for standard HD, $0.29 for FHD, $0.40 for QHD and $0.80 for UHD. Continuation costs $0.12 per second in draft and $0.41, $0.53, $0.65 or $0.95 at those standard resolution tiers. These are BFL rates; a partner platform may bill differently.

A ten-second text-to-video attempt costs $1.70 at standard HD or $2.90 at FHD. Three HD drafts cost $1.80; enhancing one ten-second draft at the listed FHD rate adds $2.90, for $4.70 total. That example includes exploration and one selected render, not further retries or editing.

FLUX 3 Max is fal Research’s video model, post-trained in collaboration with Black Forest Labs. Use blackforestlabs/flux-3-max/text-to-video for a new scene or blackforestlabs/flux-3-max/image-to-video for a starting image and motion prompt. This hosted video route is separate from Image Klein and its downloadable weights.

Checked October 23, 2026, both fal routes quote $0.0425 per output second at 720p and $0.0725 at 1080p during the October 15–November 7 promotion. After the promotion, the listed regular rates are $0.17 and $0.29 per second. A 10-second attempt therefore costs $0.425 or $0.725 during the promotion, compared with $1.70 or $2.90 at regular rates. Use the regular rates when planning work after November 7.

Choose a 5–20-second duration and inspect the finished clip, including any generated audio. The 1080p option uses upscaling. For image-to-video, an optional ending image guides the final frame. Each generated attempt contributes to the project bill, including clips you reject.

FLUX 3 Action is a 7B world action model. It uses camera images, state and an instruction to predict chunks of controls together with future video latents. BFL documents 32-action chunks for DROID and games and 42 for SO-101. Start with the checkpoint and action format for the target robot; the current inference interface returns actions.

The release connects to LeRobot and supplies code and fine-tuning recipes. BFL’s SO-101 example uses about 200 demonstrations; its displayed footage runs at four times speed with pauses removed. Those details matter when judging responsiveness from the demo.

Action’s weights use the FLUX Kommunity License. Review that checkpoint’s license for your organization and intended use. Budget robot hardware, demonstrations, adaptation and inference separately from creative API charges.

Choose the hosted creative routes when your project needs an API workflow with explicit generation modes. Image’s references and boxes help specify an edit; Video’s cached draft enhancement lets you approve the shot before buying a higher-quality render. For local still-image work, compare Qwen and Ming’s complete pipelines and licenses.

Keep Action in a robotics evaluation. Compare it with another policy on the same robot and success criteria, including pauses and recovery. Use control-task results to choose a robot policy and creative outputs to choose an image or video workflow.

Community examples linked on BFL’s Video page include dialogue, multi-shot tutorials and image-to-video motion. That publisher-curated collection is useful for finding original creator work. Use the linked dialogue and motion clips to choose a representative brief for your own production-cost comparison.