Protoface FLUX.3 Video · Text to Video

Text-to-video on FLUX 3, the gated multimodal model BFL shipped in July 2026. Prompt in, up to twenty seconds of video with native audio out. It's the purest test of the model's joint training - no start frame to lean on, no reference to anchor to, just your description.

The context, because it changes what to expect

FLUX 3 was announced as one model trained across image, video and audio rather than an image model with a video adapter bolted on. The samples were striking enough that the reactions in the KB read like a launch event, and the caveats in the same threads are the useful part: multiple shots and complex actions are where consistency goes. A single sustained shot of something specific is the strong case.

Text-to-video is the operation that exposes that most directly. If you want control, generate a still and use image-to-video - you'll spend the same money and get a composition you chose. Come here for exploration, for B-roll, or when the motion you want is the point and the frame doesn't matter.

Also worth internalising: the open FLUX 3 Dev backbone is "coming soon" and not on any HuggingFace repo, so there's no "just run it locally later" fallback. This node is the access.

How it works

Protoface isn't a model host - it's a catalog plus an adapter. At startup it reads bundled request schemas for each supported model/task pair and generates a native ComfyUI V3 node, black-forest-labs/flux-3 × video.generate in this case. On queue: schema validation first, then submission, polling with a progress bar, download into the pack's state directory, and decoding into a native VIDEO object.

Before any of that, the request identity gets journaled locally - account, request_scope, node id, generation_id, model id, schema fingerprint, payload hash. Requeue the exact same graph and it resumes the recorded run instead of submitting another paid job; the idempotency key gets reused on ambiguous network failures rather than being treated as a fresh request.

Controls

prompt is your only creative input, so be specific about the shot. Subject and setting, then action, then camera. One motion idea per clip. Twenty seconds is long enough that two competing actions will collide.

duration - 5 to 20 seconds, default 5. Since there's no per-second resolution trade-off to worry about here beyond tiers, start at five and extend the winner rather than generating twenty seconds of a bad idea.

resolution - 720p or 1080p, default 720p. Only two choices, and 1080p on a 20-second job is the expensive combination.

aspect_ratio - a required combo defaulting to 9:16 with 16:9, 4:3, 1:1, 3:4, 2:1 and 21:9 around it. Note that default, because it's the opposite of Wan's 16:9 and it's the first thing you'll change for anything not vertical. It's a strong hint about who BFL expects to use this.

generate_audio - default true. Native synchronized audio is the reason this model is interesting rather than just another video endpoint; leave it on unless you're dubbing.

And no seed. That's a real limitation, not an oversight in the node - the FLUX 3 family exposes none. Reproducibility is generation_id plus an unchanged graph, so save what you like.

Outputs

video - a list of native VIDEO results, so downstream nodes map over them and a Video Combine or preview just works. files - the original media as a PROTOFACE_FILE for Protoface Save File. run id - a string; it's your recovery handle and your support ticket reference.

Install

cd ComfyUI/custom_nodes
git clone --branch v0.1.0 --depth 1 https://github.com/protoface-ai/protoface-comfyui.git protoface
python -m pip install -r protoface/requirements.txt

Or install protoface-nodes through ComfyUI-Manager by searching Protoface, then restart ComfyUI and reload the browser. ComfyUI 0.41.0+ and Python 3.10+. The pip step pulls httpx, jsonschema, Pillow and av - no checkpoints, no CUDA, no VRAM. Then enter your Protoface key via the Protoface panel in the lower-right; use PROTOFACE_API_KEY_FILE if it needs to survive restarts.

Notes from the trenches

Check the ratio before you queue. 9:16 default catches everyone once. A batch of "cinematic 16:9" B-roll that comes back vertical is an expensive lesson.

Silence means auth or credit. A failed key or an empty balance doesn't produce a dramatic error, it produces nothing. Open the Protoface settings dialog and confirm the endpoint reads https://api.protoface.com.

Long jobs outlive browsers. Fifteen- and twenty-second 1080p generations take a while. If the tab disconnects, don't touch the graph - requeue unchanged, or use Protoface Resume Run with the run id.

Don't treat cancellation as a refund. The README is explicit that interrupting requests cancellation on the API side as a best-effort operation.

Comfy Cloud can't install this. It doesn't allow arbitrary custom nodes, and the billing is Protoface's own account rather than Comfy credits.