Prompt Suggester

Prompt writer's block is a real workflow problem, and Suggester is the pack's answer: it takes one prompt and a GGUF model and hands you five variations in a single run. What makes it interesting is the randomize switch, which flips its whole personality. It's the pack's "give me options" node, and it's tuned for consistency-by-default rather than chaos.

What it's for

When you've got a prompt that's 90% right and you want five siblings to A/B, or you're brainstorming a series and want a batch of near-miss variations to shuffle through. It works best with the LLaVA 1.5/1.6 family GGUFs, per the author - those models are already trained on the kind of image-prompt text this node is asking for.

How it works

Like the other structured nodes in the pack, it runs a constrained chat against the model handle you give it (from LLMLoader), with a JSON schema demanding exactly five suggestions - no preamble, no commentary, no "Sure! Here are five prompts". The randomize boolean swaps the instruction the model receives:

• randomize ON (default) - "create five close, useful variations of the input." Similar-but-different. Good for keeping a subject while remixing framing and style.
• randomize OFF - "create five deliberately different but production-ready ideas." Broad remixes. Better for exploration.

Temperature sits alongside it at 0.15 by default - deliberately low, so the five results are usable rather than unhinged.

The inputs that matter

• prompt - the seed.
• model - CUSTOM handle from LLMLoader; this is the part beginners miss.
• randomize - the mode switch above. This is the input you'll actually fiddle with.
• temperature - 0.15 default; raise it only if the variations come out too samey.

Output

A single STRING - the five suggestions, returned as JSON-shaped text per the pack's structured-output convention. Run it through ViewText to read it, or pipe it into a JsonToText node. From there, pick one and wire it into your sampler's CLIP encode.

Install and gotchas

The standard pack install, plus the llama.cpp caveat that applies to every LLM node here:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI_VLM_nodes
python -m pip install -r ComfyUI/custom_nodes/ComfyUI_VLM_nodes/requirements.txt

Use a llama-cpp-python wheel built for your GPU backend. And a reality check: "five variations" is only as good as the model behind it. A 1.8B prompt GGUF will give you five variations that all sound like the same prompt; a 7B LLaVA will give you five that read like a mood board. If the results feel repetitive, that's the model's ceiling, not the node.