Veo2 Image To Video

The tensor-based Veo 2 image-to-video node is the one you'll actually use most: it takes a plain ComfyUI IMAGE - from a Load Image node, a Nano Banana call, an img2img pass, whatever - and animates it with Google's Veo 2. No GCS buckets, no file paths, just wire a tensor in and get a video out. It's the most direct way to bring a still you're happy with to life.

How it works

The node encodes the image tensor, attaches it to a Veo 2 generation request on Vertex AI as the first frame, and polls the long-running operation until the video is ready. The optional last_frame input (another IMAGE tensor) anchors the end of the clip, so the animation interpolates between two images you control instead of wandering. Output comes back as video_paths - a VEO_VIDEO path list pointing at MP4s in ComfyUI's temp folder.

Inputs that matter

• image - the required starting frame tensor.
• last_frame - optional ending frame for interpolation.
• prompt - what happens in the video. An empty prompt is tolerated (Veo interprets the image), but a real description gets you intentional motion.
• image_format - PNG, JPEG, or MP4, a format hint for the API.
• aspect_ratio - 16:9 or 9:16.
• compression_quality - optimized or lossless. Lossless requires output_gcs_uri or the node throws an interactive error.
• person_generation - allow_adult / dont_allow.
• duration_seconds - 5 to 8 seconds (Veo 2's range differs from Veo 3.1's 4–8).
• enhance_prompt - on by default; the model rewrites your prompt, and this disables seed.
• sample_count - up to 4 variations, billed per video.

Optional: negative_prompt, seed, output_gcs_uri, and the pack-standard gcp_project_id / gcp_region.

Gotchas

The seed vs. enhance_prompt interaction is the one that bites: with enhance_prompt on (the default), your seed is ignored. Unlike Veo 3.1 - where you can't turn enhancement off at all - this node lets you flip it, and that's the only way seed reproducibility actually works here. So if you're chasing a repeatable look, disable enhancement and set a seed; if you want the model's prompt polish, accept the randomness. And remember Veo 2 has no native audio and no resolution selector - you get the model's fixed 720p-class output.

Installing and wiring it

Shared pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/GoogleCloudPlatform/comfyui-google-genmedia-custom-nodes
pip install -r comfyui-google-genmedia-custom-nodes/requirements.txt

Or ComfyUI Manager → search comfyui-google-genmedia-custom-nodes. Restart, look under Google AI/Veo2. Set project/region, authenticate, and each video bills to Vertex AI. The video_paths output feeds VeoVideoSaveAndPreview to watch it, or VeoVideoToVHSNode for frames. If you're starting a fresh pipeline today, Veo 3.1's tensor node is the better model - but this one is simpler, cheaper per run, and genuinely toggleable in ways the newer node isn't.