Edit Video (with Elements)

What it is

DeepGen_V2VR is the most interesting node in the pack, and also the one with the most intimidating name. It's a video-to-video editor like DeepGen_V2V0, but it's built around Kling 3.0 Omni (Pro or Standard), and "with Elements" means you can hand it reference images of subjects you want preserved - a character, a prop, an object - while it re-renders your video. You edit the scene; the element inputs try to keep the thing you care about consistent.

That matters because subject drift is the classic V2V failure mode. Without a reference, "the character changes identity every few frames" is the complaint in just about every video-editing thread. The element inputs are the pack's attempt to pin that down.

How it works

Same cloud pipeline as the other video nodes - base64 attachments, a queue, polling every 15 seconds, and an mp4 downloaded into ComfyUI's temp folder as the pack's VIDEO type. The difference is what you attach. Alongside the source video and your prompt, you can attach up to two "elements," each built from two images:

• element_1__frontal_image / element_1__reference_image_1
• element_2__frontal_image / element_2__reference_image_1

The frontal image is the subject's straight-on shot - the identity anchor. The reference image is a second angle or pose that tells the model more about the subject's look. Two elements max, so pick your battles: usually the main character, and maybe a key prop or a second character. There are also image_1 and image_2 as general context frames, in case you want to show the model more of the source scene.

Inputs and outputs

• model - kling-o3-pro (Omni Pro) or kling-o3 (Omni Standard). Pro is the flagship and bills like one; Standard is the sensible first try.
• prompt - the edit instruction.
• seed_value, nb_results (1–10 variations, only the first returns - keep it at 1), output_prefix, config_json.
• Optional: video, the element images above, aspect_ratio, resolution, duration (Kling supports 3–15s), generate_audio.

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

Wiring and install

The video input wants the pack's VIDEO type (a DeepGen generation result or DeepGen_LVID), and the element inputs are plain IMAGEs from any ordinary image node - a Load Image, a previous frame, whatever. Install is the pack standard: Manager → "ComfyUI-DeepGen-API", or clone https://github.com/deepiksdev/ComfyUI-DeepGen-API.git + pip install -r requirements.txt, restart, and set the DeepGen API key via the Settings gear (stored in ComfyUI/user/deepgen/config.json).

Where it bites

This is the most expensive node in the pack - Kling Omni Pro is a flagship closed video model and it charges like one. Test on Standard first, short duration, low resolution, before you commit a long edit. Elements reduce drift, they don't eliminate it, and a weak frontal image (facing away, bad lighting) defeats the whole feature. And the usual closed-model package deal applies: your source video goes to a server you don't control, moderation is applied at the source and can't be patched, and every run costs credits.