(Down)Load LLaVA-OneVision Model

The name is doing double duty and it's not kidding about either half. (Down)Load LLaVA-OneVision Model is the first node in every workflow from kijai's LLaVA-OneVision pack, and the only one that costs you real disk space, a real download, and most of your VRAM. Run it once, and it pulls the vision-language model from HuggingFace into ComfyUI/models/LLM/LLaVA-OneVision/, then loads it into memory. Everything else in the pack - the Run node, the folder captioner - plugs into the LLAVAMODEL output this thing hands you.

So why would you reach for it? LLaVA-OneVision is the lmms-lab family of Qwen2-based vision models, the ones people actually use for local captioning and "what's in this image" questions. This node is simply the gateway: pick your model size, pick how much you want it quantized, and it does the download-and-load so the rest of the pack has something to chew on. Kijai is one of the most trusted names in the ComfyUI ecosystem (KJNodes, the Wan video wrappers, the Bernini integration), so this small side-project inherits some of that trust - but it's also, honestly, an afterthought of a pack. The README is literally marked "WORK IN PROGRESS," and the author notes he installed zero new dependencies and it worked. That tells you the real story: this pack leans hard on whatever your environment already has.

The inputs that matter. Four of them, and you'll only touch three most of the time:

• model - four choices: the 7B and 0.5B versions, each in -ov (one-vision, trained to handle multiple images and video-style frame input) and -si (single-image only). If you're captioning one still at a time, -si is the tighter pick; the -ov variants are the ones that reason across a batch of frames.
• precision - fp16 by default, with bf16, fp32, and the quantized fp4/nf4/int8. This is your VRAM lever. A 7B in fp16 is roughly a 15GB download and a real VRAM commitment; nf4 is where people go when their card complains.
• device - cuda, cpu, mps. Note the trap: the code hard-fails on fp16/bf16 with a CPU device, so keep that on fp32 or a quantized setting.
• attention - leave it on sdpa. flash_attention_2 is faster only if you've already got the flash-attn package installed, and eager is there as a fallback when something's broken.

How to install it. ComfyUI Manager, search "ComfyUI Llava-OneVision", install, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-LLaVA-OneVision
# restart ComfyUI

The requirements.txt is minimal - huggingface_hub, pillow, transformers>=4.57.0 - and the code also imports torchvision and bitsandbytes-based quantization, neither of which is in that file. Practically: torchvision ships with ComfyUI, so that's fine. But the fp4/nf4/int8 options will throw an ImportError if you don't already have bitsandbytes in your Python env, because the pack doesn't install it for you. That's the single most common "why won't this load" failure, and it's not in the README.

The gotchas worth knowing. The download happens on your first queue run, not at install time, so the first execution is slow and chatty in the console. Everything is cached in that models/LLM/LLaVA-OneVision/ folder afterwards, so it's a one-time tax. And if a model choice seems to do nothing, remember the output is a bundle, not a tensor - it's meant to be fed straight into LLaVA-OneVision Run or OneVision Caption Folder. Those two, not this node, are where the actual work happens.