flux / v3.0 / imageToVideo

The name isn't a promise about your GPU - it's a door to Black Forest Labs' FLUX 3 Video, the closed multimodal model that generates video (and audio, and eventually "actions") from a single set of weights. As of now there are no open FLUX 3 weights to download; the video runs on BFL's gated API, and Civitai resells it through the orchestration service this pack talks to. So this node is the honest version of the thing: the model isn't on your machine, it costs Buzz, and that's the whole point.

It's part of civitai-comfy-nodes, Civitai's official pack of ~160 nodes generated straight from its Orchestration API spec. Every node in the Civitai/Video menu works the same way: the node bundles your inputs, submits a videoGen job to Civitai's fleet, long-polls until it finishes, and drops a native VIDEO tensor back into your graph. No local model, no VRAM, no download.

What you actually set

prompt and image are the only required inputs. The image is an IMAGE socket, so wire it from a Load Image or anywhere else in your graph - it gets converted to base64 and shipped up. Then the dials:

• duration - 5 to 20 seconds, default 5. Longer costs more; it's a linear meter.
• resolution - 720p or 1080p.
• aspect_ratio - auto (let the model choose), 16:9, 9:16, 1:1, 21:9 and the rest.
• generate_audio - on by default. FLUX 3's native audio is the headline feature: dialogue, SFX, ambience that stay synced to the video. Turn it off only if you're adding sound yourself.
• draft - this one's worth finding. It renders a fast, cheap 720p draft instead of the full-quality render, and the result carries a draft cache that a later draftEnhance operation re-renders at full quality. Preview your composition for a few Buzz, enhance only the take you like.

Outputs are video (VIDEO), draft_cache_audio (AUDIO), draft_cache_video (VIDEO), plus workflow_id and raw_json (STRING) for debugging and cost inspection. The node also shows the workflow id and cost on itself after each run.

Install and auth

The pack is official, published to the Comfy Registry, and installs in one step:

# ComfyUI Manager: Custom Nodes Manager → search "Civitai Comfy Nodes"
# or from the command line:
cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt   # requests + python-socketio

Then restart ComfyUI. Auth resolves in order: a connected Civitai Auth node, the CIVITAI_API_TOKEN env var (headless installs, easiest), or a stored OAuth login from the Civitai sidebar's connect panel.

Where people get burned

Three things, in order of how often they hurt. Cost. Video is where a Buzz session goes sideways fast - every second is billed, and a 20s 1080p render is not pocket change. Use draft mode before committing. Your inputs leave the machine. That's the mechanism, not a bug: prompt and image go to a server you don't control, subject to its moderation. It's early preview. The pack README says so plainly - nodes, APIs and behavior "may change without notice." If a node suddenly looks different after an update, that's the lifecycle, not you breaking something.

The FLUX 3 video tier is genuinely strong - consistent across shots is its known weak spot per early evals - but you're renting that quality per job, not owning it. Plan your prompts before you hit Run.