Black Forest Labs connects FLUX image models with a growing video API

Black Forest Labs develops FLUX generative models and offers them through a playground, a hosted API and selected open-weight licensing routes. The current offer spans FLUX.2 image generation and editing alongside FLUX 3’s video preview. Its practical appeal is to teams that want model capabilities inside a larger creative or software workflow. They still need to supply the application, asset handling and acceptance decisions around each result.

01 / ProductImage generation and video now share the FLUX umbrella

The company overview presents an API, playground, open weights and enterprise arrangements. It currently leads with FLUX 3 while retaining FLUX.2 and other tools. A buyer should therefore avoid treating “FLUX” as a single interchangeable endpoint. The model family, version, media type and access route all influence what can be requested and how the work is charged.

FLUX.2’s overview describes image generation and editing with multiple reference images, including ways to combine visual elements and guide a composition. It lists several variants with different intended tradeoffs. Those are vendor descriptions rather than independent performance findings. For a product-image workflow, evaluate whether the required subject and composition survive the requested edit, rather than assuming that a more expensive variant will automatically reduce finishing work.

The FLUX 3 documentation describes a preview model for video with synchronized audio. It accepts text, pinned image keyframes or a source clip to continue; video editing is a separate tool endpoint. The guide distinguishes available modes from forthcoming Omni Reference functionality. That distinction matters: an announced direction should not be included as a working dependency in a production specification.

02 / AudienceUseful for builders and creative teams with a defined handoff

A software team may want customers to request image variants from inside a product. A creative studio may want a controlled way to generate concept frames and then develop a short sequence. Black Forest Labs is relevant when the team wants direct access to generation capabilities and is prepared to design the surrounding process. A model endpoint does not provide the same working environment as an entire editing application.

For a team that values access to multiple model providers through hosted infrastructure, Replicate’s deployment approach offers a useful comparison. Check the actual model variant, request schema, output retention and unit price on each route. A common model family name does not prove identical features, current versions or commercial treatment across hosting services.

A creator whose main need is an integrated video working surface should also examine Runway’s workspace. Compare the path from initial concept through revision to an approved shot. The API route offers room to fit generation into custom software, while the workspace route may supply more of the everyday creative interaction. The best choice depends on which parts of that process the team already owns.

03 / WorkflowA proposed image-to-video campaign pipeline

Consider a fictional outdoor-equipment campaign that needs a still concept and a short atmospheric clip. Use an owned reference image and an invented brand rather than a real unreleased product. The proposed evaluation should answer two separate questions: whether the still meets the art direction, and whether the moving sequence retains the elements that matter. Neither result should be assumed from the other.

1. 01

   Approve the still-image brief

   Define the product silhouette, setting, palette and required empty space. Generate a small set of image concepts and retain the selected source with its model identifier and references. Ask a reviewer to identify the exact details that must remain stable when motion is added.

2. 02

   Specify a modest motion task

   Start with one understandable camera or environmental movement. Keep the scene short and avoid combining several difficult actions in the first evaluation. Judge whether the movement improves the message and whether the product remains legible at the intended display size.

3. 03

   Separate exploration from final rendering

   Where the chosen endpoint supports a draft path, review the draft before spending on a final render. Record the selected draft’s data with the job. A new request with similar words should not be treated as the same creative decision.

4. 04

   Store and review the delivered media

   Bring the completed file into controlled storage and review the entire clip, including its opening and closing frames. Inspect the audio separately if it is included. Preserve the approved still, clip and final edit so a subsequent campaign revision does not depend on recreating an earlier output.

The API quickstart describes project-scoped keys, organization-level billing and asynchronous requests: submit a job, retain the returned polling URL and retrieve the result when ready. It explicitly warns against exposing the API key in client-side code. In the proposed application, the backend should own credentials and associate each provider request with the customer’s job before the interface reports that generation has started.

Build the review screen around what the user needs to know: the requested asset, its current state and the next useful action. A request still waiting for completion should not look like an invitation to submit another paid job. If the browser closes, the application should be able to reconnect the person to the existing job rather than lose the relationship between the request and its result.

For the video preview, BFL documents a draft-enhancement path that reuses a returned draft bundle and warns that result URLs expire roughly two hours after completion. Treat the final download as part of completing the job. The storage task deserves its own failure state: “generation succeeded” is not the same as “the application safely retained the deliverable.”

04 / PricingCost depends on the endpoint, resolution and media unit

The official pricing documentation uses credits worth US$0.01 and lists the same pricing for API and playground usage. Image variants can use resolution-dependent prices, while video charges depend on duration and mode. The table gives selected entry points to show the units, not a claim that every request will cost the displayed minimum.

Illustrative arithmetic: a five-second hd text-to-video request at $0.17 per second is $0.85 before any additional operations. Four such candidates would be $3.40 even if the reviewer approves only one. This is arithmetic from the listed rate, not a tested acceptance rate or full production quote. Editing, higher-resolution delivery and the team’s review time can change the effective cost substantially.

The open-weight licensing page separates Builder, Platform, Professional and Enterprise arrangements. Its distinctions include internal or marketing use, features delivered to end users, agency client work, and broader negotiated requirements. The route is not simply “download a model and do anything.” Match the intended use and model coverage to the actual agreement before comparing self-hosted economics with API pricing.

A useful cost sheet should contain one row per production step and one column for discarded attempts. Do not assume every endpoint uses the same unit because it shares the FLUX name. For self-hosting, include model licensing, hardware utilization and operating effort. For the hosted route, include all generations and transformations that lead to the final accepted file.

05 / DistinctionsDirect model access makes the application design consequential

BFL’s offer puts important creative decisions within reach of software. An application can maintain a consistent brief, choose an allowed endpoint and connect the result to an approval process. This is valuable when the team already has an asset library or publishing workflow and wants generation to fit inside it. The model call should serve that process rather than force users to manage a separate collection of untracked outputs.

Reference-led image work also suggests a useful evaluation method: keep the reference assets fixed while changing one requested attribute. Ask whether identity, material and composition remain acceptable. Then reverse the test by holding the art direction stable while changing the subject. This reveals what kind of control the team can rely on without turning a few examples into a claim of universal consistency.

For video, the draft and final-render distinction can support a clearer approval checkpoint. A reviewer can choose the intended shot before the system commits to more expensive processing. That benefit depends on retaining the correct draft data and making the transition understandable. If the application loses the relationship, the existence of a draft feature alone does little for the production team.

06 / QuestionsPreview status and licensing leave specific questions to resolve

FLUX 3’s preview label should influence the rollout. Start with work where a rejected or delayed clip can be replaced, and preserve a fallback for time-sensitive delivery. Ask the provider about the support and change policy required for the application’s use case. Public documentation describes available functionality; it does not prove a particular customer’s service-level agreement.

The same discipline applies to claims about image detail, text or identity preservation. Inspect the actual product label or character across the whole workflow and judge it at delivery size. A visually plausible result can still fail a factual product requirement. If exact packaging must survive, consider preserving approved source artwork in the final composition rather than asking the model to reconstruct it.

Finally, confirm commercial rights for the selected access route. An API subscription, a hosted partner’s terms and an open-weight license are separate agreements. Client-facing generation, internal marketing and training-data creation can raise different licensing questions. Resolve the relevant one before a successful prototype spreads to a second department with a different use case.

07 / DecisionChoose the narrowest FLUX route that proves useful