Zephyr Chat

Zephyr Chat is the pack's answer to "I'm tired of writing twenty prompts by hand." You type one theme into it, and it runs Stability AI's StableLM Zephyr 3B chat model locally inside ComfyUI, then hands you back a list of prompts as a string. Wire that into a prompt text field, and one node call replaces an evening of typing.

It's a deliberately narrow job. The default prompt is literally "A list of interesting subjects for photographs are as follows:" - the node exists to be a prompt-lister, not a general assistant. And because it runs the model on your own machine, there's no API key, no per-call cost, and nothing leaves your box. The author's own plan, from the thread where they showed off the grabber node, was to pair a local GPT like this with BingImageGrabber to generate fresh search terms on the fly - a closed loop where one node writes the query for the next.

How it works

The node loads stabilityai/stablelm-zephyr-3b through HuggingFace transformers (AutoModelForCausalLM + AutoTokenizer) on first execution, which means a real model download before anything happens - roughly 6GB for the fp16 weights. It wraps your prompt in Zephyr's chat template, generates up to 512 new tokens at temperature 0.8 with sampling, then strips the assistant header and drops any lines shorter than four characters. The result is one STRING: your list, cleaned up, ready to paste or wire onward.

device_map="auto" means it will happily offload to CPU if your VRAM is tight, but don't expect speed - a 3B model on CPU generates at a crawl. Give it a few GB of VRAM and it's fine.

Inputs and outputs

One input, one output - there's nothing to misconfigure:

• prompt (STRING) - your theme or instruction. Defaults to the "interesting subjects for photographs" list opener.

Output is a single STRING with the generated list, which you wire into whatever prompt input you want (KSampler positive prompt, text concat node, or straight into the grabber's search_term).

Installing it

Same story as the rest of the pack - ComfyUI Manager, search ConCarneNode:

cd ComfyUI/custom_nodes
git clone https://github.com/concarne000/ConCarneNode
# restart ComfyUI

Here's the gotcha this node specifically cares about: the pack's requirements.txt only installs auto-gptq, but Zephyr needs transformers, and that isn't in there. Most ComfyUI installs already have it because some other node pulled it in, but if the whole pack fails to import (all three nodes missing), pip install transformers is the fix. And every install of this node means a 6GB download from HuggingFace on first run - on a slow connection that's a coffee break, not a refresh.

Where people get burned

The big one is caching. Zephyr doesn't override IS_CHANGED, so ComfyUI treats it like any deterministic node: feed it the same prompt and it returns the cached result - it will not regenerate on re-queue. People assume the model is broken. It isn't; you just have to change the input to force a new generation.

Second, set expectations about the writing. StableLM Zephyr 3B is a late-2023 3B instruct model. It's genuinely useful for enumerating subjects and variations - "list twenty photography subjects" is its sweet spot - and mediocre at anything that needs real prose. Temperature is fixed at 0.8 in the code and there's no seed control, so you can't nudge the creativity knob from the UI.

Third, the 2026 context: with LLM-encoded checkpoints (Flux 2 Klein, Z-Image) now reading natural language directly through their own text encoder, a separate "prompt-writing LLM" is less essential than it was in the SDXL era. Where it still earns its keep is SDXL-lineage workflows that want booru-tag style variation - a local chat model turning "cats" into a varied list of tag-filled prompts beats hand-rolling thirty of them.

If you want a modern tiny model that writes better, a current 3B (Llama or Qwen) will beat this frozen 2023 one - but as a zero-setup, no-key prompt-lister that lives inside the graph, Zephyr Chat does the job.