Load LFM2.5-VL Model

If your GPU can't fit a 10B vision model alongside the diffusion model you're actually generating with, this is the loader you reach for. Load LFM2.5-VL Model pulls Liquid AI's LFM2.5-VL-1.6B-Extract - a 1.6-billion-parameter vision-language model tuned for extraction - and runs it entirely in-process. No API, no key, no Ollama. The whole thing fits in around 4 GB of VRAM, which is why it's the pack's budget option.

It's part of mamorett/orama, which runs three VLM families locally: this 1.6B LFM2.5-VL, the ~1B MiniCPM-V, and the 10B Step3-VL. LFM is the lightweight end - good enough for solid structural captions, fast enough that you don't feel it in the graph.

How it works

First run, the loader downloads the weights with huggingface_hub.snapshot_download into ComfyUI/models/vision_models/LiquidAI/LFM2.5-VL-1.6B-Extract - real copies, not symlinks into your system HF cache, so the model lives entirely inside ComfyUI's own model tree. Then it builds an AutoProcessor plus an AutoModelForImageTextToText and loads them with whatever dtype and device map you picked. The loaded model is cached in memory for the rest of the session, keyed by (model_id, dtype), so the second queue run doesn't re-download or re-load - you just get the model back.

One tell this is a bleeding-edge integration: the pack's __init__.py patches a transformers naming mismatch for Lfm2VlImageProcessorFast. That's a sign you need a current transformers, not a sign of trouble.

The inputs that matter

• model_id - a dropdown, and for now it's a single choice: LiquidAI/LFM2.5-VL-1.6B-Extract.
• dtype - auto, bfloat16, float16, or float32. auto (default) is right for this model; float16 can save a bit of VRAM if you're desperate.
• device_map - auto (default), cuda, cpu, or mps. cpu technically works; it's also how you make a 1.6B model feel like a 10B one. Keep auto.
• use_flash_attention_2 - off by default, and only turn it on if you've actually installed the flash-attn package, which this pack does not install for you. It's a compile-from-source affair on most setups, so don't flip it casually.

Output is a single LFM2_VL_MODEL socket that feeds the LFM2.5-VL (Liquid AI) 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 just search orama in ComfyUI Manager. The real gotcha here is the dependency floor: the pack requires transformers>=5.7.0 (LFM2.5-VL needs its Lfm2VlProcessor), and if your ComfyUI venv has an older one, the node import will fail with a confusing error. Update transformers first, then install the pack. First model load also downloads several GB, so don't panic when the queue seems stuck - watch the console for the download progress.