flux / v3.0 / extendVideo

The frustrating thing about text-to-video is that you get a nice 5-second loop and then it just ends. FLUX 3's video extension node is the answer to that specific frustration: it takes a clip you already have and continues it, coherently, with audio, without the fade-to-black hop you get from stitching two separate generations. It's a closed model running on Civitai's fleet - no weights, no local GPU - which for FLUX 3 video is the only door that exists.

What it is

flux / v3.0 / extendVideo from civitai/civitai-comfy-nodes, under Civitai/Video/flux. It takes your source clip, a prompt describing what should happen next, and extends the video by the requested duration. BFL's FLUX 3 is a single multimodal model - image, video, audio - which is why this node can carry native audio through an extension instead of dropping it.

The inputs

• prompt - required. What happens in the extension. The model reads the end of your clip and continues from it, so the prompt should describe the next beat, not recap what already happened.
• source_video - required. A VIDEO socket. The tooltip gives the limits: MP4, under 50 MB, under 15 seconds. If you hit a size error, that's the 50 MB ceiling; if a long clip gets rejected, that's the 15-second one.
• duration - 5–20 seconds, default 5. How long the extension is, not the total.
• resolution - 720p or 1080p, default 720p. Start at 720p; 1080p is where the bill grows.
• aspect_ratio - auto plus 21:9, 2:1, 16:9, 4:3, 1:1, 3:4, 9:16. Auto lets the model match your source, which is usually right; set it manually if you want the extension to force a crop.
• generate_audio - default true. Native synchronized audio, because FLUX 3 does audio natively. Turn it off for a silent extension (cheaper, and sometimes you're scoring it separately anyway).
• draft - default false. The cheap-first path: render a fast, cheap 720p draft that carries a draft cache, then feed that cache to the flux / v3.0 / draftEnhance node for the full-quality render. Draft-first is the smart default for anything you're not sure about.

The outputs

• video - the extended clip, as a VIDEO.
• draft_cache_audio (AUDIO) / draft_cache_video (VIDEO) - present when draft is on; the input to a later draftEnhance.
• workflow_id / raw_json - the pack standard.

Installing it

ComfyUI Manager → Civitai Comfy Nodes → Install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Gotchas

• 50 MB / 15 seconds are hard ceilings on the source. Compress or trim before this node, not after a rejection.
• Metered, per second, at resolution. 1080p with audio on a 20-second extension is a real spend. Draft mode exists precisely to de-risk this.
• Your source video is uploaded to Civitai's servers. It's a closed-model API call; nothing stays local.
• Credentials - Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth.
• The extension quality depends on what you feed it. Clean, stable endings extend better than clips that end mid-motion.