Flux 3 Text to Video

Describe a scene, get a video. That's the whole pitch, and FLUX 3 is good enough at it that the sentence stops being a demo and becomes a workflow. This node is Black Forest Labs' hosted text-to-video model in ComfyUI: type a prompt, pick a length and resolution, and a finished clip with synchronized audio comes back. It shipped in ComfyUI core in August 2026.

The thing that sets it apart from most text-to-video nodes is generate_audio, on by default. FLUX 3 doesn't just animate - it layers in sound, and the prompt is where you steer it. Say what happens visually, then describe the audio track separately: the tooltip literally suggests separating ambient sound, music and speech so each layer lands. That's a genuinely useful trick most people discover late.

How it works

This is a partner node, which in plain terms means the generation happens on BFL's servers, not your machine. The node POSTs your prompt to Comfy's flux-3-video proxy endpoint, polls while BFL works, and downloads the clip when it's done. You need a Comfy account with credits - billing is prepaid and per call, shown as a price badge on the node before you run it. Your local GPU does nothing but render the graph, which is the point: this is the "I want a clip and I want it now" escape hatch when no local video model fits in VRAM.

The prompt itself is interpreted and expanded before generation, so you don't need to nail every detail - write it like you'd brief a director, not like you're stuffing a classifier.

The inputs that matter

• prompt - the whole creative brief. Visuals, then sound, and it's happy being plain English.
• duration - "auto" or 5 to 20 seconds. Auto fits the length to the content.
• resolution - 720p or 1080p. Bigger costs more per second; the price badge does the math for you.
• generate_audio - on by default. Off gives you a silent video, which you may actually want if you're adding a soundtrack later.
• safety_tolerance - moderation strictness from 0 (strictest) to 4. Text-only requests can use the full range.
• seed - cosmetic, really. FLUX 3 chooses its own seed, so results are nondeterministic; change it only to force a re-run.

The output is a single VIDEO - save it from the result player or route it into a video node downstream.

Gotchas

The seed being ignored catches everyone - don't hunt for a "good" seed, it won't reproduce anyway. And mind the moderation: BFL's hosted tier is heavily filtered, so if your prompt trips the safety net you'll get an error status rather than a clip, and the strictness slider is your only dial. Finally, remember this runs in the cloud: a clip costs credits whether or not you love it, so it pays to iterate on short, cheap takes before committing to a 20-second 1080p render.