Protoface FLUX.3 Video · Image to Video

FLUX 3 is the model you saw everywhere in July and can't run. Black Forest Labs announced it on 23 July 2026 as one multimodal model across image, video, audio and robot action - and shipped it as gated early access. FLUX 3 Dev, the open-weight backbone, is listed as "coming soon" with no date. So the showcase clips everyone was reacting to came from an API. This node is one of the ways in.

What it's good at, and what to expect

The interesting claim is architectural: instead of a still-image model with a video add-on, FLUX 3 was trained jointly across modalities. The samples landed hard - the reactions in the KB are the "this was one prompt" kind. The same threads carry the caveat that matters more for practical work: consistency across multiple shots and complex actions is the weak spot. Single continuous shots of something specific: excellent. A five-shot sequence with a character walking through three rooms: expect drift.

Image-to-video is where you can compensate for that. You bring the frame, so composition is settled before the model touches anything, and it spends its capacity on motion instead of inventing a scene.

How the node works

Protoface generates this node from a bundled catalog at startup - black-forest-labs/flux-3 × video.animate - and every generated node in the pack runs the same adapter: encode the input tensor to PNG, upload if new (content-hash cached), validate the payload against the schema before submission, submit, poll with a progress bar, download, decode into native ComfyUI types.

There is a local journal in front of all of that, and this is where it earns its keep. A 20-second 1080p video job is not something you want to accidentally buy twice, so the adapter records the request identity - account, scope, node id, generation_id, model, schema fingerprint, payload hash - before submitting. Requeue unchanged and it resumes; the API idempotency key is reused on ambiguous failures rather than retried as a new paid request.

The controls

prompt - the text description of the video. Because the image carries the look, spend the prompt on motion: what moves, how fast, what the camera does, what the light is doing.

image - your start frame. It's a mode field rather than a bare socket: pick Image and connect anything that emits an IMAGE (Load Image, a local generation, a composited frame), or References to pass a Protoface Asset Reference. Batches are rejected - one asset per field.

resolution - 720p or 1080p, default 720p. Just two tiers, which is refreshing.

duration - 5 to 20 seconds, default 5.

aspect_ratio - optional. Left on Default, the output follows the model's own choice; set it when the input image and the delivery format disagree.

generate_audio - default true. Native synchronized audio is the headline capability of a jointly-trained multimodal model; turn it off if you're cutting to your own music.

One thing that's missing

There's no seed on FLUX 3. Compare the Wan nodes in this same pack, which do expose one. Your only reproducibility lever here is the journal: keep generation_id fixed and requeue the identical graph to get the same job back. Change anything and you're rolling again. Plan for that - if you get a take you like, save the file, don't assume you can re-derive it.

Outputs

video - a list of native VIDEO objects, ready for preview, Video Combine, or chaining into FLUX.3 · Extend or Interpolate. files - the original download for Protoface Save File. run id - a string. Write it down for anything longer than a few seconds.

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 Manager → Protoface → protoface-nodes → restart → reload. ComfyUI 0.41.0+, Python 3.10+, dependencies httpx, jsonschema, Pillow, av. No weights - which is the entire point, since the open FLUX 3 weights don't exist yet. Paste your key via the Protoface button bottom-right, or PROTOFACE_API_KEY_FILE for a server.

Traps

Treating it like a local video model. No seed, no steps, no sampler, no VRAM tuning. If a guide mentions CFG on FLUX 3, close the tab.

Re-running to "get a better one" and losing the old one. Downloads land in the pack's state directory and results are never silently regenerated by a package update, but if you want a keeper, save it out.

Assuming the open weights are imminent. They were "coming later in 2026" at announcement with no date, and the KB's note on BFL is that the frontier stays proprietary while the open piece stays a promise.

Bumping generation_id after a timeout. Don't. Requeue unchanged or resume by run id.