Generate Video (from Images with Elements)

DeepGen_I2VR is the pack's most ambitious video node and the one most people overlook. "VR" here means video with recurring elements - the node takes your start and end frames, plus definitional images and videos for specific characters or objects, and generates a clip where those elements are guaranteed to appear, and appear right. This is the closest the DeepGen pack gets to character-controlled video.

Why you'd reach for it

Ordinary image-to-video gives you a scene; this one gives you a scene with named cast. The element slots are the hook: for each element you can provide a frontal image (the clean, face-on reference), a reference image, and - in the first element slot - a reference video showing how that element moves. Two element slots exist in the schema (element_1__frontal_image, element_1__reference_image_1, element_1__video, and the same trio minus the video for element_2). Feed the model a frontal shot of your main character plus their reference video, and it can render that character doing things across the clip while holding identity - which is the recurring-elements feature Kling's Omni models are known for. It's only available for the two Kling 3.0 Omni models (Pro and Standard), which is what you'd expect: this is a flagship capability on a flagship model.

How it works

Start and end images plus all the element inputs become base64 attachments on the request (images as PNGs, the element video read from its file path and attached as video bytes). The job queues, the node polls every 15 seconds until it renders - the longest waits in the pack - and the finished clip downloads to ComfyUI's temp dir as a ComfyVideoMock for the usual preview and save nodes.

The inputs that matter

• start_image, end_image (IMAGE) - the scene's bookends.
• element_1__frontal_image (IMAGE) - the clean, face-on reference for your main element. This is the one that anchors identity.
• element_1__reference_image_1 (IMAGE) - a second angle or context for the same element.
• element_1__video (VIDEO) - a reference clip showing how the element moves. Feed it a few seconds of your character walking and the output moves like them.
• element_2__frontal_image, element_2__reference_image_1 (IMAGE) - the second element's pair (no video slot).
• prompt - the scene and action.
• model - Kling 3.0 Omni Pro or Omni Standard.
• duration, generate_audio, aspect_ratio, resolution, nb_results, seed_value, output_prefix, config_json - the standard video set.

Outputs

VIDEO, output_prefix_and_model (STRING), total_credits_used (FLOAT).

Install

Same pack:

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 set your DeepGen API key (deepgen.app) in Settings → "DeepGen API Key" (ComfyUI/user/deepgen/config.json).

Gotchas

This is the most expensive and slowest node in the pack - full render, audio on, Omni Pro is not a budget move. The frontal image does the heavy lifting for identity, so a good face-on reference beats a dramatic one. And everything you feed it - including the reference video - leaves your machine for DeepGen's servers, which is worth remembering before you feed it footage you'd rather not share. It's a niche node, but if recurring characters in video is your use case, there's nothing else like it in this pack.