Image Tiles to LLM

Here's the workflow this node completes: your tiled upscaler samples every tile, and you'd like each tile's prompt to actually describe that tile - the sky tile prompting about clouds, the character tile prompting about the character. Image Tiles to LLM is the captioner for that loop: it sends each tile to a vision-capable LLM server (llama.cpp or anything OpenAI-compatible), collects the per-tile descriptions, and outputs them as one newline-separated string. That string is exactly the tagger_text input the pack's ImageTiledKSamplerWithTagger expects. Crop → caption → sample with per-tile prompts → merge. Closed loop.

The server-side setup is the same external-server pattern as the pack's JoyCaption and Qwen3VL nodes: your own llama.cpp running a multimodal model, no API key required. The default apikey is literally sk-no-key-required - that's the tip-off that this is BYO-server, not a hosted service.

The inputs

• tiled_images - tiles from ImageCropTiles or ImageCropTilesByPixels. One request per tile.
• base_url - default http://localhost:58080/v1, the llama.cpp OpenAI-compatible endpoint.
• model - the model name your server reports (default gemma-4-31B-it; set it to whatever your llama.cpp is actually serving - mismatch is the most common error here).
• apikey - placeholder by default; keep it unless your server enforces one.
• temperature / max_tokens / timeouts - sampling and request-timeout controls. 300s default timeout is generous because a vision LLM chewing a big tile can take a while.
• system_role - the system prompt (default <|think|>You are a helpful assistant).
• prompt_template - the user prompt template; put {image_data} where the tile's data URL goes. The stock template is a long, deliberately uncensored "describe everything in clinical detail" instruction - swap it for something gentler if that's not your dataset's vibe. It's just a template; you can replace it entirely.

One output: responses - newline-separated captions, one per tile, in tile order.

How it works

The node needs the openai Python package to talk to the server. If it's missing, the node raises an ImportError telling you to install it:

pip install openai

Each tile is converted to a PNG data URL and sent as a user message; responses are joined with newlines. It processes tiles sequentially, so a 20-tile image is 20 LLM calls.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/mirabarukaso/ComfyUI_MiraSubPack
pip install openai   # the only extra dependency this node needs

or search "MiraSubPack" in ComfyUI Manager and restart.

Notes and gotchas

• Model name mismatch is the #1 failure: llama.cpp returns a model name that isn't gemma-4-31B-it, the request 400s, and you get nothing. Check what your server reports (llama-server prints the model path at startup) and set model to match.
• It's slow by design. One request per tile, sequentially, with generous timeouts. Budget for it.
• Line count must match the tile count when you paste into tagger_text - the KSampler matches lines to tiles by index.
• If you're not running a vision-capable server (a text-only GGUF), you'll get "model doesn't support images" style errors - the LLM has to be multimodal for this node to be anything but an expensive way to read a filename.