Painter Prompt Rewriter

"Write a prompt for this" is the most common thing people hand to an LLM, and video models are the most demanding audience: LTX wants a paragraph, Wan wants structure, and your one-liner won't cut it. PainterPromptRewriter is an in-graph LLM call that takes a short prompt, rewrites it to match a template you specify, and - the useful bit - can look at up to nine reference images while it does. No external API, no key, no network round-trip: it runs on the text-generation capability inside the CLIP model you already have loaded.

What it is

A text category node that returns a single generated_text string. Required inputs are clip (the model providing generation), the prompt to rewrite, a template describing the target format, max_length (default 2048), and a sampling_mode control. Optional inputs add the reference images, a max_image_long_side cap, and three behavior toggles.

How it works

It uses the same interface as ComfyUI core's TextGenerate node - clip.tokenize / clip.generate / clip.decode - so it's running the LLM-style text generation inside the CLIP, not a separate model you have to install. The template is the real power: you paste the exact format your video model wants (the LTX-style long-prompt guide, a Wan prompt skeleton, whatever you've settled on) and the node rewrites your rough idea into that shape. sampling_mode opens a DynamicCombo with the sampling knobs - temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty - which you mostly leave alone until output gets repetitive or bland.

The reference images (up to 9) are downscaled to max_image_long_side (default 960) before being sent, so the model can describe or incorporate what it sees. thinking (if the model supports it) gives it a reasoning pass; use_default_template applies the model's built-in chat template; use_chatml_system is the experimental option that injects your template as a ChatML system message instead of user text - flip it if you're getting ignored instructions.

The inputs that matter

• prompt - what you want said.
• template - the format to force. This is the node's entire personality; spend time here.
• clip - the generation model (any CLIP that supports generate).
• images / max_image_long_side - optional visual guidance.
• sampling_mode - advanced; leave defaults until output quality needs tuning.

Output: generated_text → wire into your prompt/CLIP text encode.

Install

Part of the PainterNodes pack. ComfyUI Manager → search "PainterNodes" → install → restart, or:

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

Common issues

• Template ignored - try use_chatml_system (the experimental toggle exists for exactly this), or check your template actually describes the output shape rather than just stating a goal.
• Garbage or repetitive output - raise repetition_penalty slightly, nudge temperature down, or check that your clip actually supports text generation at all. Not every CLIP does.
• Node won't load - it's built on the newer ComfyUI node API; update ComfyUI core. Same story as the rest of this pack's newest nodes.

It's a small node that saves a huge amount of tab-switching - the "go paste my idea into ChatGPT, get a paragraph back, come paste it in" loop, done inline. And because it can see your reference images, it's genuinely better at the job than the generic LLM tab.