LLM Text Processor

If you've been re-wording your prompts by hand while your image model waits, this is the node that kills the copy-paste step. LLM Text Processor is a single-node pack that runs a local GGUF model through llama.cpp and hands the finished text back inside your workflow. The name is honest about the other thing it is: no API, no key, no cloud. Every token is generated on your machine.

Why you'd reach for it: it fits the direction image prompting went. Modern LLM-encoded models read your prompt as an instruction, so having another local LLM write that instruction is translation between two things speaking the same language - the community converged on this exact "have an LLM enhance the prompt" pattern, and it stopped being a browser-tab thing and became a node. This pack is a deliberately small take on that: one node, prompt writing, rewriting, translation, captioning, extraction. Not an ecosystem, just the one thing.

How it works

The node doesn't load the model into Python. It shells out to llama.cpp's llama-cli, which the pack downloads itself - pinned to a specific release (b8840 at the time of writing) and vendored under the pack's vendor/llama.cpp folder. It builds the CLI command from your sampling inputs, writes your prompt to a temp file, and for vision it saves each image to a temp PNG and passes --mmproj plus --image. Then it parses llama-cli's output: reasoning blocks wrapped in [Start thinking] / [End thinking] get split into the REASONING output and stripped from the final answer, and the speed line gets captured as PERF. Temp files are cleaned up even on errors, and it respects ComfyUI's interrupt button.

Inputs and outputs that matter

The two dropdowns you'll set first are model and mmproj. Models are discovered recursively from ComfyUI/models/LLM; the projector dropdown shows only files with "mmproj" in the name, and it's required if you connect an image. Everything else is familiar sampler territory, but a few deserve attention:

• prompt - what you ask the model to do. Defaults to Describe this image in detail. and supports dynamic prompts.
• max_tokens - 2048 default, up to 32768. This is the setting that bites (see below).
• reasoning - off, on, or auto, controlling whether the model thinks before answering.
• system_prompt - a dropdown fed by .txt files in ComfyUI/models/LLM/prompts. Drop a captioner.txt there and it appears as an option.

The three outputs are RESPONSE (reasoning stripped), REASONING (the thought block, if any), and PERF (llama.cpp's speed line). Wire RESPONSE into your text encode or prompt input and you're done. There's also an enable_processing toggle that turns the node into a pure passthrough - handy for A/B comparing generated vs. hand-written prompts, and it skips all model checks.

Installing it

ComfyUI Manager, search LLM Text Processor, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/KingManiya/ComfyUI-LLM-text-processor

Then drop GGUF files anywhere under ComfyUI/models/LLM (for example ComfyUI/models/LLM/My-Model/model-q4_k_m.gguf). The node only needs numpy and Pillow, which ComfyUI already ships - the real download is llama.cpp itself, done automatically on first run.

The big gotcha: it's Windows x64 CUDA 13 for now

Here's where people get burned. The automatic llama.cpp setup only supports Windows x64 + CUDA 13. On Linux or Mac the shipped code hits an unsupported-platform error and stops - there's no manual binary-path override in the pack as of 1.3.0, despite the README implying otherwise. If you're not on that exact setup, check the repo for newer releases before committing to it. On Windows, a failed setup is almost always internet/GitHub access - run the node again and watch the console, which logs download progress, total size, and speed so a slow connection doesn't look like a freeze.

Common issues

• No models in the dropdown - nothing with a .gguf extension under models/LLM. Place one and refresh/restart.
• Image input fails - mmproj set to none, or a projector that doesn't match the model. The node catches the embedding-mismatch error and tells you which n_embd values collided, so the fix is usually "pick the mmproj that belongs to your GGUF."
• Empty or cut-off response - raise max_tokens. With reasoning on, the model spends part of the budget thinking before it ever reaches the answer.
• Out of memory - lower ctx_size first (8192 default), then a smaller quant.

The README also has sensible starter presets for Qwen, Gemma 4, and gpt-oss - copy those numbers and adjust from there. It's a genuinely useful little node if your platform matches; a hard pass if it doesn't.