Generate Video (from 3 Images)

DeepGen_I2V3 is the video node for when two bookends aren't enough. Give it three images and it generates a clip that has to pass through all of them in sequence - not just start and end, but a defined middle beat. That's the difference between a transition and a story: a beginning, a turning point, and a resolution, all shown rather than described.

Why you'd reach for it

Three-frame i2v is for moments that have a narrative arc. A character walks toward a door, opens it, and the third frame is the new room. A product rotates, assembles, then glows. An establishing shot pushes in to a detail, then widens out. With two frames the model invents the middle; with three you get to own the middle. The model list is the subset that supports it - Kling 3.0 Omni Pro and Omni Standard, Grok Imagine Video, Veo 3.1, and Seedance 1.0 Lite. All closed, all cloud, all metered; this is the API-wrapper deal doing what local video can't.

How it works

The three images become base64 PNG attachments submitted alongside the prompt; the job queues and the node polls every 15 seconds until the clip renders - expect a multi-minute wait with your workflow blocked. On completion the video downloads to ComfyUI's temp dir and comes back as ComfyVideoMock, the pack's VIDEO-type stand-in, so the standard preview and save-video nodes just work.

The inputs that matter

• image_1, image_2, image_3 (IMAGE) - the three beats, in order. All optional, but the node's whole point is all three connected. Order matters: it's 1 → 2 → 3.
• prompt - describe the action threading the frames together, not each frame individually: "she enters, pauses at the desk, then turns to the window."
• model - the three-frame-capable dropdown.
• duration (1–15, default 5), generate_audio (default on), aspect_ratio / resolution (free strings), plus nb_results, seed_value, output_prefix, config_json - the standard video set; duration and resolution are your cost dials.

Outputs

VIDEO (the clip), output_prefix_and_model (STRING), and total_credits_used (FLOAT - meter with the pack's DeepGen_F2T0).

Install

Same pack, same routine:

cd ComfyUI/custom_nodes
git clone https://github.com/deepiksdev/ComfyUI-DeepGen-API.git
cd ComfyUI-DeepGen-API
pip install -r requirements.txt

Restart, then paste your DeepGen API key (deepgen.app) into Settings → "DeepGen API Key", saved to ComfyUI/user/deepgen/config.json. Only requests and opencv-python install.

Gotchas

Three frames is more context for the model to reconcile, and the more different they are, the more likely it is to compromise on all three rather than honor them - keep the subject consistent across frames and let the prompt carry the motion. The wait and the credits are the same story as every video node here: block for minutes, spend real money. Keep durations short for tests.