Load MiniCPM-V Model

MiniCPM-V is the one model in this pack that actually has a track record in the community - it's been a favorite small VLM for years, mostly running under Ollama as an external server. Load MiniCPM-V Model is mamorett/orama's take on it: run openbmb/MiniCPM-V-4.6 (or the huihui-ai abliterated variant) in-process through transformers, no Ollama process, no API. At roughly 1 billion parameters on a SigLIP2 vision tower plus a Qwen3.5-0.8B language model, it's the pack's other lightweight - ~3 GB of VRAM - and its headline feature is a visual token compression slider that trades speed for detail.

How it works

Same bones as the other loaders: first use, it downloads weights into ComfyUI/models/vision_models/openbmb/MiniCPM-V-4.6 (real files under ComfyUI's own model tree, not symlinks into the system HF cache), then builds an AutoProcessor and AutoModelForImageTextToText with your chosen dtype and device map. The twist is visual_token_compression: in the "4x (detailed)" mode it swaps in an image processor built with compress_ratio=4, meaning the vision encoder emits more visual tokens for the LLM to read. "16x (fast)" (the default) compresses harder, so fewer tokens, less work, faster generation.

Loaded models are cached in memory keyed by (model_id, dtype, compression mode), so switching between 16x and 4x reloads the model - worth knowing if you bounce between modes mid-session.

The inputs that matter

• model_id - two choices: openbmb/MiniCPM-V-4.6 or huihui-ai/Huihui-MiniCPM-V-4.6-abliterated. The abliterated build runs the identical inference code, it just has the refusals filed off.
• visual_token_compression - the interesting one: "16x (fast)" default, "4x (detailed)" for text-heavy or detail-dense images. If the model keeps missing text in a screenshot, try 4x - that's exactly what it's for. You pay in tokens and time.
• dtype - auto / bfloat16 / float16 / float32. Leave it on auto.
• device_map - auto, cuda, cpu, or mps. The mps option is a nice touch for Mac users, which is genuinely rare in this corner of the ecosystem.
• use_flash_attention_2 - off by default. The tooltip says it's recommended for multi-image tasks, and it can be - but it requires the flash-attn package, which this pack does not install for you, and flash-attn is a pain to build on Windows and older CUDA. Skip it unless you specifically need the speed.

Output is a single MINICPM_V_MODEL socket feeding the pack's MiniCPM-V (OpenBMB) inference node.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/mamorett/orama
# in your ComfyUI python environment:
pip install -r ComfyUI/custom_nodes/orama/requirements.txt

Restart ComfyUI, or find orama in ComfyUI Manager. This model needs transformers>=5.7.0 and torchvision; the README's av/torchcodec extras are for video input, and since this pack's inference only reads the first frame of an image anyway, you don't need them for normal captioning. Keep an eye on the console on first load - the weight download is a few GB and it happens before the model appears.