Flux 3 Video T2V (BFL)

Here's the pitch: one node, a prompt box, and Black Forest Labs' FLUX 3 Video renders up to 20 seconds of clip with native audio - and your GPU never gets within sniffing distance of the workload. Flux3VideoT2V_BFL is the text-to-video entry point in gelasdev's ComfyUI-FLUX-BFL-API, a pack that wraps BFL's hosted API so you can drive it from inside a normal ComfyUI graph.

This is the node you reach for when you want FLUX 3 quality but don't want to babysit a local video model. Wan, LTX, and their friends will eat your VRAM and your patience; the trade here is that you're renting BFL's hardware instead. FLUX 3 Video is still early-access as of mid-2026 - the community's been cautious about pricing transparency and reliability - so treat it as "try it and see if the bill makes you wince," not as a settled default.

How it works

The mechanism is the same for every node in this pack: POST to https://api.bfl.ai/v1/flux-3-video with your arguments, then poll get_result every five seconds until the task lands on a terminal status. Video is slow, so the polling ceiling is cranked to 240 attempts - that's about twenty minutes of waiting. The maintainers literally note in the changelog that a real fhd v2v job was still generating at eleven minutes. When it's Ready, the MP4 gets downloaded and handed back to ComfyUI as the native VIDEO type. No model files, no downloads, nothing to babysit except your patience and your API balance.

The inputs that matter

Most of them are self-explanatory, so here's the short list:

• prompt - your scene description, and honestly the whole game. FLUX 3's video language handling is good enough that a plain-English description beats keyword soup.
• duration - auto, or a fixed 5–20 seconds. auto lets the model pick, which is fine until you need a specific cut.
• resolution - hd or fhd. If you're just testing, hd + draft keeps the bill and the wait sane.
• generate_audio - on by default. FLUX 3 ships synchronized sound, which is still rare enough in video models that you should flip it off only when you're doing your own audio pass.
• safety_tolerance (0–4, default 2) - how strict BFL's moderation is. Lower is stricter. Relatedly: this is a BFL product, so assume heavy censorship baked into the pipeline.
• draft - a faster, cheaper preview mode. Start here before spending full-res generations.
• config (optional) - the BFL_CONFIG socket for overriding key/base URL/region per-node.

The output is video (VIDEO), which you wire into ComfyUI's built-in Save Video or preview nodes.

Installing it

Via ComfyUI Manager, search "ComfyUI-FLUX-BFL-API" and hit install. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/gelasdev/ComfyUI-FLUX-BFL-API.git
cd ComfyUI-FLUX-BFL-API
pip install -r requirements.txt

Dependencies are just torch (plus requests, numpy, Pillow, av - all already present in ComfyUI). Two non-negotiables: ComfyUI ≥ 0.3.30 (the node literally disables itself on older versions because it needs the native VIDEO type), and a BFL API key. Drop it in config.ini or, better, in a Flux Config (BFL) node wired into this one - the config file gets clobbered on updates, and people have been complaining about that since the pack's early days.

Common issues

The big one: a black or blank video output. The node falls back to a 1-frame black 512×512 clip whenever the API errors, gets content-moderated, or the polling times out - so a black result almost always means the job died server-side, not that ComfyUI is broken. Check the console output for the task ID and status; Content Moderated means your prompt tripped their filters. And set expectations on time: an fhd generation can sit in Generating for over ten minutes. That's not a hang - the node's waiting for the cloud.