Review Video

What it does

DeepGen_V2T0 is the "have a model watch your video" node. It ships the pack's own video (or up to five still images) up to a Gemini 3 vision model along with your instructions, and gets back a block of text - a critique, a description, a shot list, whatever you asked for. The display name says "Review Video," and that's the honest job: it's a reviewer you drop into the graph to look at what you just rendered.

Three models, all from Google's Gemini 3 line: Pro Preview, Flash, and Flash Lite. That's a deliberate pick by the pack author - Gemini is the strongest video understanding you can reach through an API, and it's the same "call the cloud, nothing local" pattern as every other node in the pack, just text-shaped instead of video-shaped.

How it works

Same client as the rest of the pack. The video file gets base64-encoded and attached to a POST to {base_url}/{model}/api (it's the text/"Chat" path, with streaming turned off), and the response is parsed back into a text output. No queue, no polling - text comes back synchronously, which makes this the fastest node in the pack.

The inputs and outputs

• model - the three Gemini options. Flash is the sensible default; reach for Pro Preview when the job is genuinely hard (dense scene, lots of motion).
• prompt - this is where you do the work. "Describe the motion and list the artifacts" gets you a QA report; "turn this into an LTX prompt" gets you a rewritten one. Be specific.
• seed_value - required, and mostly a formality for an LLM; leave it at 1000.
• config_json - optional extra API parameters.
• video - the pack's VIDEO type (see below).
• image_1 through image_5 - IMAGE inputs, so you can throw in key frames alongside, or instead of, the video.

Outputs: output (STRING - the review text; wire it to a text display or straight into your next prompt), model (STRING - which agent actually served it), and total_credits_used (FLOAT).

The natural loop: generate with DeepGen_T2V0 → feed the VIDEO back into this node → take the critique → refine the prompt and regenerate. That review loop is the whole reason to put this in a graph rather than pasting clips into a chat window.

The VIDEO gotcha

"VIDEO" here is the pack's own type, not anything native to ComfyUI. It comes out of DeepGen's generation nodes and the pack's DeepGen_LVID Load Video node. A VideoHelperSuite clip won't drop into this socket - if the input won't connect, load the file with DeepGen_LVID instead of fighting the type system.

Install and key

Same as every node in this pack: Manager → search "ComfyUI-DeepGen-API", or clone https://github.com/deepiksdev/ComfyUI-DeepGen-API.git into custom_nodes and pip install -r requirements.txt (just requests and opencv-python - no models to download). Restart, then set the DeepGen API key via the Settings gear; it's stored in ComfyUI/user/deepgen/config.json.

Where it falls short

A Gemini review costs a little on every call, and it has the weakness every vision model shares: with multiple subjects in frame it mixes up who's doing what. Use it for a fast pass, not as an oracle. And the security note applies here as it does everywhere in this pack - it's a small, lightly-known pack that ships your key to a network API by design, so a quick skim of the source before you trust it with a paid key is the sensible move.