CXH_HG_Model_Load

CXH_HG_Model_Load is the loader for the pack's MiniCPM-V 2.6 route - the "prompt generator" option in the README's ranking of captioning backends. It downloads pzc163/MiniCPMv2_6-prompt-generator into models/LLM/ and hands you a CXH_Hg_Pipe for the two run nodes that consume it: CXH_Min2_6_prompt_Run (single-image prompt generation) and CXH_Min2_6_classifiy (batch image classification).

MiniCPM-V 2.6 is a small vision-language model (~8B) from OpenBMB that's surprisingly capable at following instructions, and the pzc163 fork is a popular community build tuned for writing Stable-Diffusion-style prompts from images. It sits in the middle of the pack's speed table: slower than Florence 2, faster than JoyCaption, and good at turning an image into a usable generation prompt rather than a raw description. The README's example workflow even pairs it with CogFlorence and a Q8 flux GGUF for a fully local image-to-prompt-to-image loop.

What's on the node

It's about as minimal as loaders get - one input, one output:

• Input: model - a single choice, pzc163/MiniCPMv2_6-prompt-generator. No dropdown gymnastics; you're getting what the pack author tested.
• Output: pipe (CXH_Hg_Pipe) - the loaded model + tokenizer, wired into the run or classify nodes.

There's no precision or attention selector. The load itself uses AutoModelForCausalLM with trust_remote_code=True and no explicit dtype, so it comes up in whatever the model defaults to - usually fine on a mid-range card, but the first load is a multi-GB download and a few seconds of patience.

Installing

Standard pack install - ComfyUI Manager (search "Comfyui_CXH_joy_caption") or:

cd ComfyUI/custom_nodes
git clone https://github.com/StartHua/Comfyui_CXH_joy_caption
cd Comfyui_CXH_joy_caption
python -m pip install -r requirements.txt

Restart, then connect this node and run once to trigger the download.

Gotchas

• First run downloads the model to models/LLM/MiniCPMv2_6-prompt-generator - the same LLM/ folder the JoyCaption Llama and Florence 2 variants use, so a big download is expected.
• No caching toggle. Unlike the JoyCaption loaders, there's no explicit unload here - the pipe stays resident once loaded, and the run node keeps referencing it. On a small card, that plus your checkpoint can get tight.
• NVIDIA effectively required. The run nodes feed images to the model via a CUDA-expected path; this pack isn't engineered for CPU.
• If you're only here for captions, the JoyCaption nodes are usually better quality; reach for MiniCPM when you specifically want a generation-style prompt or classification.