⭐ Star CLIP Prompt Refiner

The name is wrong, in a good way

It says CLIP and it isn't CLIP - there's no contrastive text-image model in here. ⭐ Star CLIP Prompt Refiner loads a text encoder out of models/text_encoders as a language model and lets it write your prompt, using the same pipeline as ComfyUI's native Generate Text node (clip.tokenize → clip.generate → clip.decode), with the loader baked in and a stack of system-prompt presets on top.

Why you'd reach for it: on Z-Image, Klein, Krea 2 and most 2026 models the text encoder is a small Qwen3 instruct LLM. Reusing that same file as your prompt writer means no second model, no Ollama server process, no llama-cpp-python build to match your system, no API key, no refusal on the far end - and ComfyUI already handles loading and offloading it.

The trade-off: it isn't a chat window - one shot per run, no history. And the consensus on enhancers is well-earned cynicism: LLM-written prompts carry noise, and a prompt you wrote by hand usually beats one a model invented. It's best when the LLM polishes and structures rather than doing the thinking for you.

How it works

It loads your chosen checkpoint through comfy.sd.load_clip, then feeds the model a chat-formatted token stream with system_prompt replacing the system prompt in the model's chat template. Your sampling settings go to clip.generate, and the returned text is split on </think> - reasoning before it, answer after - hence two outputs.

IS_CHANGED is the quiet part worth knowing. With use_seed on it caches like any other node: queue twice with identical inputs, get the identical prompt back. Turn it off and it re-runs on a fresh random seed every queue.

The inputs that matter

• clip_name - dropdown of everything in models/text_encoders. Defaults to qwen3-vl-4b-heretic_int8.safetensors; if that file isn't there it falls back to the first entry in your folder, so don't be surprised by what's pre-selected.
• prompt - your rough idea.
• system_prompt_preset - the reason to use this over the native node. 29 choices: Custom plus 28 shared presets (Flux2(Klein) SysPrompt, Krea2 Image Prompt Refiner, Character Sheet Expert SFW/NSFW), read from json/startext.json - the same file Star Easy-Text-Storage and Star Ollama Prompt Helper write to. Add one there, refresh, it appears here.
• system_prompt - used when the preset is Custom. Leave it blank and you silently get the Flux2(Klein) SysPrompt preset: a long, opinionated style guide about subject hierarchy, lighting-as-geometry and hex codes, written specifically for Qwen3 encoders. Read it before trusting it.
• temperature (0.7), plus top_k, top_p, min_p, repetition_penalty and max_length (512). Temperature is the one you'll actually touch: ~0.4 for template-shaped output, 0.9 when you want it to riff. max_length is a hard ceiling, and 512 tokens isn't always enough with thinking on.
• thinking - leave it on. The author's own note: disabling it injects an empty think block into the template for Qwen3-VL heretic models, which then echo your prompt instead of answering.

Optional inputs: clip_override reuses a CLIP already loaded in the graph (dropdown ignored, no second copy of a 4B model in VRAM). image1 grows a slot as you connect, up to four, for vision-language encoders; batches flatten into individual reference images. video (24 FPS assumed, subsampled to 1 FPS) and audio only matter for models that accept them.

Outputs are result - the prompt, into a CLIP Text Encode or any text widget - and think_output, the reasoning. Read the second one when the first looks off; it tells you whether the model understood you.

Installing it

ComfyUI Manager, search Starnodes, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

That requirements list is for the rest of the pack; this node needs none of it. What it needs is a text-generation-capable model in models/text_encoders. The default expects the heretic build of Qwen3-VL 4B in int8 - heretic is an abliteration method, refusals surgically removed, which is the only reason presets like the NSFW character sheets work. Comfy-Org publishes stock encoders into the same folder (qwen3vl_4b_bf16.safetensors, ~8.5 GB). Point this at a plain CLIP or T5 and you get errors, not a smarter prompt.

Where it goes wrong

It just repeats your prompt. Empty-think-block symptom - turn thinking back on. Still repeating? Lower the temperature and check that a Custom preset didn't get left blank and then overridden.

Chatter leaking into the prompt. Preset problem, not a bug. Pick one that says polish rather than expand, or end your Custom system prompt with "output the prompt and nothing else". The common complaint about every local enhancer - hand it an already-decent prompt and let it finish the job.

Prim or refused output. You're on a stock instruct encoder. Swap to an abliterated build - Qwen3-VL's refusals are a fine-tuning artifact, and running locally is the whole point.

VRAM. Two models load per workflow now - diffusion plus a 4B encoder. ComfyUI manages the offload; on a tight card use the int8 file and clip_override when that encoder already exists upstream.