📝 S42 CutFlow Scene Describer

Writing a good LTX prompt means describing what's actually in your footage - the lighting, the color palette, how much motion there is - and doing that by eye is slow and subjective. S42CF_SceneDescriber takes a video (or a single image) and analyzes the actual pixels to produce a scene description plus a ready-to-steal list of prompt elements. It's from S42 CutFlow's LTX Bridge group, and its whole purpose is to sit at the front of a regeneration workflow: feed it a reference clip, read what it says, paste the suggestions into your LTX prompt.

Let me be upfront about what this isn't, because the name invites the wrong expectation: it is not a vision LLM, it does not "understand" the scene, and it calls no API and needs no model download. It's deterministic frame statistics dressed up as a description. That's not a knock - it means it's instant, free, and reproducible - but set your expectations accordingly. It'll tell you your footage is warm-toned with high-key lighting and fast motion. It won't tell you there's a red car driving past a gas station.

What it computes

Feed it video (an IMAGE batch) and pick detail_level - basic or detailed. Under the hood it does four cheap measurements on the first frame (plus the last, if there's more than one):

• Lighting - luminance from the standard 0.299/0.587/0.114 weights, bucketed into high-key, balanced, or low-key.
• Color - which RGB channel dominates, giving you "warm tones," "cool tones," "natural tones," or neutral.
• Motion - mean absolute difference between consecutive frames, bucketed from "static" through "high motion/action."
• Composition (always computed, and detailed adds high-contrast and cinematic hints) - aspect ratio tells it whether the framing is widescreen, vertical, or standard.

Outputs

Two STRING outputs. scene_description is the human-readable summary - resolution, framing, lighting with the actual measured brightness and contrast numbers, dominant color, and the motion verdict. prompt_suggestions is the useful one: it assembles prompt-ready phrases from the measurements ("bright lighting," "dramatic low-key lighting," "dynamic movement, action shot") and always tacks on 4k, highly detailed. In detailed mode it adds contrast and "cinematic" when the footage warrants it.

So the honest workflow is: use the description to confirm what you think you're seeing, and use the suggestions as a starting vocabulary rather than a finished prompt. LTX punishes short, vague prompts - the KB on this is blunt that LTX 2.3's prompt adherence still lags - so starting from these phrases and then adding the specifics a human can see (the car, the location, the lens) beats starting from a blank text box. Real users typically feed these suggestions into their prompt and then edit.

Installing it

Standard S42 CutFlow install - it's one of 50+ nodes in the pack, so there's nothing extra to fetch:

cd ComfyUI/custom_nodes/
git clone https://github.com/GeekyGhost/S42-CutFlow.git
pip install -r S42-CutFlow/requirements.txt   # Windows portable: python_embeded\python.exe -m pip

Or use ComfyUI Manager and search S42 CutFlow. Restart, and it's under S42 CutFlow/LTX Bridge. No model files, no HuggingFace downloads, no VRAM cost - that's the whole selling point.

Gotchas

Because it measures the first frame and the average of all frames for motion, a clip that starts dark and ends bright will read as dark. And if you feed it a single image, motion will correctly say "static image" - don't be surprised. The biggest trap is over-trusting the output: the color verdict is "red dominates" or "blue dominates," which is a useful shorthand but collapses any warm scene into the same bucket. Use it as the opening draft of a prompt, not the final word, and you'll get real mileage out of it.