zhenzhen-flux-3-video-lowprice

FLUX 3 is BFL's multimodal turn - one model spanning image, video, audio and even robot action - and as of mid-2026 it ships as gated early access: API plus private weights, with the open FLUX 3 Dev backbone still "coming soon." Which means, if you want to touch it, you're calling an API. This node is that door, at the domestic low-price rate: Comfly_flux3_video_lowprice wraps FLUX 3 video through the Zhenzhen Seedance2 relay and hands you back a real VIDEO tensor.

The community reaction to the FLUX 3 launch was split between awe and the "open source ≠ open weights" grumble - nobody can run these locally. That's exactly why a relay node exists. You get the capability without the waitlist friction, and the pack has already wired the early-access endpoints.

The model dropdown is the workflow

The model enum covers the FLUX 3 video lineup: t2v (text to video), i2v (image to video), v2v (video to video), draft-enhance (generate a draft then enhance it), each with a global- variant. That "draft" concept is a real BFL thing - you produce a cheap quick pass and then refine it, which the node supports natively via the draft toggle and the draft_cache output. Chain draft_cache from one run into the next and you're doing the two-stage workflow the model was designed for.

Other required knobs: seconds from 5 to 16, resolution (hd or fhd), ratio (auto up to 21:9 and 9:16), audio_mode (api_default / enabled / disabled - FLUX 3 can generate native audio, which is one of its headline features), and safety_tolerance from 0 to 4 for how strict the content filter is.

Optional inputs are the feed: image1 through image10 for i2v/global-i2v, input_video or video_url for v2v, plus seed, skip_error, and api_config.

Outputs: video (VIDEO tensor → straight to a save node), video_url, draft_cache (for the two-stage flow), task_id (find it in your Zhenzhen task history), and response (raw JSON).

Getting it running

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Restart, or use ComfyUI Manager (search "Comfyui-zhenzhen"). No model files - the pack is pure HTTP plumbing over an API key from api.seedance.nz, wired via the Seedance2 low-price settings node into api_config. The overseas site at ai.t8star.org also carries it.

Fair warning

Three things to know before you spend credits. First, video is where API costs bite hardest - a 16-second fhd clip with audio is not a 1-credit afternoon, so check the relay's model pricing before you batch. Second, FLUX 3 is genuinely new; expect the kind of rough edges early-access models have, and treat upstream 500s as "run it again" rather than "my setup is broken." Third, if you came hoping this was the open-weight FLUX 3 Dev - it isn't, and it can't be, because those weights still don't exist. This node is the paid door to the frontier model, which is the only door there is right now.