Load MiniCPM-V Model

Load MiniCPM-V Model is the setup node for the pack's MiniCPM-V vision-language integration. It downloads OpenBMB's MiniCPM-V model, loads it into VRAM, and hands you a single MINICPMV_MODEL object that the MiniCPM-V Q&A node then uses to look at your images and videos and answer questions about them. Nothing else in the pack touches it; you wire the model object straight into the Q&A node.

The three choices that matter

• model_version - MiniCPM-V (Full) (the original openbmb/MiniCPM-V) or MiniCPM-V-2_6-int4 (7GB VRAM). The int4 build is the sensible default for most people: roughly 7GB VRAM for a quality/performance balance, versus the full model which is heavier. Unless you have serious VRAM headroom and want maximum quality, take the int4.
• precision - bf16, fp16, or fp32. For the int4 model this only affects intermediate compute, not the stored weights, so bf16 is the right call and it's what the code defaults to internally anyway.
• attention - sdpa (faster, memory-efficient, fine on any modern GPU) or eager (more compatible, slower). Use sdpa; switch to eager only if you're on an old card that misbehaves.

The auth wall (this is where people get stuck)

MiniCPM-V is not a grab-and-go model. It's gated on Hugging Face: you must visit the model's page, click to accept the license, and generate a token. The node checks for HF_TOKEN or a logged-in Hugging Face CLI session, and if there's no token it raises a clear error telling you exactly what to do: accept the license at the model page, get a token from HF settings, then set HF_TOKEN or run huggingface-cli login. Do this before you queue the workflow, or the first run just dies.

On first use it downloads the model into ComfyUI/models/LLM/<model-name> via snapshot_download, and loading uses trust_remote_code=True - that's normal for this model family and a known quantity from OpenBMB, but it's worth knowing it's happening. The int4 path pulls in bitsandbytes for 4-bit loading. Once loaded, the node caches the model on the class, so it's reused across runs until you restart ComfyUI.

Install

Everything lives in drmbt/comfyui-dreambait-nodes. ComfyUI Manager (search "comfyui-dreambait-nodes"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/drmbt/comfyui-dreambait-nodes

then restart ComfyUI. The requirements pull in transformers, accelerate, bitsandbytes, torchaudio, librosa and more - this is the heaviest node family in the pack, so expect a real pip install and a real download on first run.

If the download fails partway, check your HF token and disk space; snapshot_download resumes, but a stale partial models/LLM folder occasionally needs deleting and re-running. And if the community's experience with MiniCPM-V is any guide, the Q&A node works fine for describing and captioning - it's the model download and VRAM budget that bite people, not the node logic.