FLUX 3 Video

FLUX 3 is Black Forest Labs' multimodal turn - one model family trained across image, video and audio - and FLUX 3 Video is the part that's actually live. It's also gated: early-access API only, with the open-weight Dev backbone promised but never shipped as of writing. So the honest statement of what this node does is "the only way most of you will ever run FLUX 3 Video," because it drops that gated API into ComfyUI as a native VIDEO output. It's a closed-frontier model that doesn't require a closed-minded workflow.

The launch reception tells you what to expect. The showcase threads were genuinely excited - "Flux 3 looks insane. This was 1 prompt" - and the consistent community caveat was that FLUX 3 Video struggles with consistency across multiple shots and complex actions. That's the real job this node has: testing whether it holds up on your shots before you commit budget to it, because this is a pay-per-run model and video runs are where cloud sessions get expensive fast.

Inputs follow the current multimodal-video shape. positivePrompt is your text or instruction, in full sentences. frameImages conditions on a starting frame; video takes a source clip as UUID or URL for video extension. settings.audio is on by default for synchronized sound. The feature that actually sets this node apart is draft mode: settings.draft gives you a fast, cheap preview to approve the idea, and the returned draftCache (a URL or Base64 bundle you feed back into the same input on a full-quality run) is how you go from draft to final without paying full price for your rejects. If you're iterating on video, that two-step loop is the difference between a sane budget and a painful one. size defaults to 720p ~16:9; settings.safetyTolerance (default 2) trades moderation strictness for permissiveness; numberResults, outputFormat, outputQuality and the TTL settings round it out. Output is VIDEO.

Install

Standard Runware pack setup - search Runware in ComfyUI Manager and install, or cd ComfyUI/custom_nodes && git clone https://github.com/Runware/ComfyUI-Runware && pip install -r ComfyUI-Runware/requirements.txt, restart. Then a Runware API key from the dashboard (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login). No weights download; that's the deal with this model.

The practical take

Use draft mode aggressively - that's the node's escape valve against the real danger of cloud video, which is paying full rate for bad ideas. And keep the community caution in your head while judging results: if a multi-shot sequence falls apart, that's the model's known weak spot, not your prompting. Your prompts and frames go through Runware's moderation, and the model's own filtering is baked in at BFL's end, so anything the vendor refuses, nobody's node settings are going to un-refuse. For the one thing FLUX 3 Video offers that no open model does - BFL's current-generation multimodal video inside a Comfy graph - that's the price.