BLIP Model Loader

This is the "load" half of a two-node pair. On its own it does nothing you can see - it loads a BLIP model into memory and hands it to the BLIP Analyze Image node, which is the node that actually turns a picture into words. If you want ComfyUI to look at an image and tell you what's in it (a caption) or answer a question about it (visual question answering), this loader is where that starts.

Why would you want that? Image-to-text is the quiet backbone of a lot of workflows. You feed a reference image into BLIP, get a text description out, and pipe that straight into your positive prompt - instant img2img-by-description without typing anything. It's also how people auto-caption a folder of images before LoRA training, or build "describe this, then regenerate it in a new style" loops.

How it works

BLIP (Bootstrapping Language-Image Pre-training) is Salesforce's older-but-reliable vision-language model. The loader's job is just to pull the weights and initialize them. Good news from the pack's own changelog: since October 2023 BLIP ships inside WAS Node Suite - you no longer need to clone a separate BLIP repo like the old instructions said. The model file downloads automatically on first use from a default URL and lands in ComfyUI/models/blip/checkpoints/. First run is slow while it fetches; after that it's cached.

If you ever need to point it at a different caption model or a mirror, the pack reads a was_suite_config.json file at its root where you can override the download URL.

The inputs and outputs that matter

Since the loader is a setup node, there isn't much to fiddle with. You pick which BLIP task you're loading for - captioning versus interrogation (the question-answering mode) - and that's essentially it. The one output is a BLIP model object. It has exactly one destination: the blip_model socket on BLIP Analyze Image. Wire it there, put your image into the Analyze node, and the text comes out of that node, not this one.

So the mental model is: Loader → Analyze → (STRING) → your prompt encoder. This node is step one of three.

Installing it

WAS Node Suite is one big pack; you install the whole thing, not this node alone. Easiest path is ComfyUI Manager: open the Manager, search WAS Node Suite, hit install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/WASasquatch/was-node-suite-comfyui

then install the pack's requirements against your ComfyUI Python and restart. The BLIP weights come down on their own the first time you run the Analyze node.

Common issues

One heads-up worth internalizing before you install anything from this pack: WAS Node Suite is retired - the author stopped active development back in December 2023 and only merges the occasional PR. It still works and is still installed everywhere (it's one of the ecosystem's classic general-purpose toolkits), but nobody's fixing dependency drift for you.

That matters specifically for BLIP, and here's the exact trap from the community: WAS pins/downgrades some packages to keep older stuff like BLIP working, and then when you update ComfyUI, ComfyUI upgrades those packages back - and the pack throws "Import Failed" on the next launch. The author has confirmed this is the mechanism. The fix is almost always to reinstall the pack's requirements.txt against the actual ComfyUI environment (activate the venv first, or run the included install.bat which finds it for you), not against system Python. If the whole suite fails to import, none of its nodes appear - so a missing BLIP Model Loader usually means an import error further up, not a missing node.

The other one: if the auto-download fails (network, or a HuggingFace hiccup), the checkpoint folder stays empty and Analyze errors out. Check that ComfyUI/models/blip/checkpoints/ actually got populated. If you're on modern hardware and want sharper captions, honestly, newer VLM nodes (Florence-2, the various Qwen-VL wrappers) have overtaken BLIP - but BLIP is light, fast, and fine for feeding prompts.