abyz22_AddPrompt

Lighting is the difference between "same prompt, boring" and "same prompt, mood." Most people handle it by hand-editing the prompt between runs. abyz22_AddPrompt is the lazier version: it picks a random lighting phrase from a list you give it, bakes in a strength weight, and appends it to your existing conditioning. Feed it a seed, hit queue, and the same scene comes back under "sunshine from window" this time and "neon light, cyberpunk" the next.

It ships in abyz22/image_control, a single-developer fork of Impact Pack. Most of that pack is body-editing and outpainting tooling, but this node is the one you'll actually reuse in everyday workflows.

How it works

You type a bunch of candidate phrases into the text box, separated by /. The node picks one at random with random.choice(text.split("/")), wraps it as (phrase:1.2), tokenizes it with your CLIP, and concats the result onto the incoming conditioning. The selected_prompt output is a plain STRING of whatever got chosen, so wire it into a ShowText node and you can see which mood you rolled without reading pixel noise.

A couple of inputs actually matter:

• text - your candidate list. The default is a 30ish-entry lighting list (sunset, black and white, warm atmosphere...), which is a fine starting point. The / separator is the whole contract, so don't put literal slashes inside a phrase.
• weight_factor (0.5–1.5, default 1.2) - the emphasis weight wrapped around the chosen phrase. 1.0 is neutral; the default nudges lighting to actually show up.
• seed - makes the pick reproducible. Same seed, same lighting. You'll want this when you find a winner and want to re-roll the rest of the workflow around it.

There's also an optional opt input that accepts anything. If you wire in a string that ends with /<number>, it uses that number as the seed instead of the widget - useful when another node (like a queue/batch controller) owns the seed and you want the lighting to stay locked to it.

Install

ComfyUI Manager → search "image_control", or:

cd ComfyUI/custom_nodes
git clone https://github.com/abyz22/image_control

Restart ComfyUI after installing. The pack pulls in openpyxl, pytorch-lightning, and kornia from its requirements.txt - pytorch-lightning is a chunky install, so the first start after adding this pack can take a while.

Where it bites

The biggest trap is the weight_factor range topping out at 1.5. Lighting at 1.2–1.5 on an SDXL or Illustrious checkpoint is usually fine, but slam it too high on a model with a fragile CFG balance and you'll get banding or a blown-out frame. If you're also using a wildcard encoder elsewhere in the graph, don't double-dip - a lighting wildcard plus AddPrompt both appending can stack emphasis past what the model tolerates. And remember this node only adds text; if your negative conditioning needs the same treatment, run a second one on the negative side and paste your own "avoid" list into it.