EvoLink Veo 3.1 Text-to-Video (Official)

Veo 3.1 has one thing almost nothing else in video generation has: it generates the audio natively, in the same pass as the picture. Dialogue, sound effects, ambient noise, synced to what's on screen. That's the genuine capability gap this model owns - open video can chase the visuals, but native audio stayed closed through all of 2026. The EvoLink Veo 3.1 Text-to-Video node is the official wrapper that puts it in a ComfyUI graph, no Google Cloud project required.

How it works

Same EvoLink pattern as every node in the pack: submit to /videos/generations, poll, download before the 24-hour links expire, hand back a native VIDEO output for Save Video. The model picker gives you two flavors of the same model - veo-3.1-generate-preview (the full model) and veo-3.1-fast-generate-preview (faster, cheaper, trades some polish). Both generate up to 4K with audio on by default.

That last part is where people get surprised. generate_audio defaults to true, and the tooltip is clear: audio is extra charge. If you just want a silent draft to iterate on, flip it off and save the credits. Turn it on when you're showing someone the actual result - that's where Veo earns its keep.

The inputs that matter

• prompt - required, ~2000 tokens, Chinese works. Write the camera move and the scene; add dialogue in quotes if you want the model to voice it.
• model - the two Veo variants, full vs. fast.
• duration - 4, 6, or 8 seconds. Veo clips are short; plan shots accordingly.
• quality - 720p (default) / 1080p / 4k.
• generate_audio - the cost-bearing toggle described above.
• negative_prompt and seed - negative prompting is a real thing here, and a fixed seed gives you reproducible results. 0 means random.

Outputs

video → Save Video. task_id for the job log at evolink.ai, status_info for a model/time/credits summary, result_urls (raw links, 24h) and response_json (full payload) for chaining - e.g., straight into the Topaz upscale node if you want the 4K output bumped further.

Install and gotchas

Part of the comfyui-evolink pack. Manager → search EvoLink → Install → fully restart ComfyUI (close the console window; a browser refresh doesn't count). Or:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Only requests as a dependency - no models, no GPU, no Google Cloud account. API key from evolink.ai/dashboard/keys (sk-), fill once, leave blank so the key doesn't travel in shared workflow JSON.

Biggest rookie move: leaving generate_audio on while iterating and wondering where the credits went. The other usual trips - node not appearing after install (restart fully, check for nested folders), "finished but no video" (wire video into Save/Preview), 401/402/429 error codes - all behave like the rest of the pack. Failed and moderation-blocked tasks don't charge; unchanged reruns hit ComfyUI's cache and don't bill twice. If you want the image-driven flavors, there are sibling EvoLink Veo 3.1 nodes for first/last-frame and reference-to-video that share this exact DNA.