BLIP Loader

Two nodes in this pack load BLIP: this one, and DownloadAndLoadBlip. They do the same job - hand you a BLIP_MODEL to plug into BLIPCaption - but this one is the plain version. It reads whatever's already in your BLIP models folder and lists it in a dropdown. It doesn't fetch anything. If that dropdown is empty, that's not a bug - it means the folder is empty, and either DownloadAndLoadBlip needs to run once first, or you need to place a checkpoint there by hand.

Reach for this node specifically when you already have the weights (from a prior DownloadAndLoadBlip run, or placed manually) and want a loader with no network step in the graph - useful on a locked-down or offline setup, or just to keep a workflow's execution path predictable.

How it works

It scans the folder ArtVenture's BLIP nodes use for checkpoints, lists whatever .pth files it finds, and loads whichever one you pick into a BLIP_MODEL object. That's the entire job - no download, no captioning, just "load this file into memory and hand out a reference to it."

The inputs and outputs that matter

One input:

• model_name - a dropdown of BLIP checkpoints found locally. Empty means nothing's there yet.

One output:

• BLIP_MODEL - wire it into BLIPCaption's optional blip_model input. Load once, reuse across multiple captioner nodes rather than letting each one load its own copy.

How to install it

Comes with the Art Venture pack. ComfyUI Manager: search comfyui-art-venture, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

then pip install -r comfyui-art-venture/requirements.txt and restart. This node itself downloads nothing - see below for getting the checkpoint in place.

Common issues & troubleshooting

The model_name dropdown is empty. There's no BLIP checkpoint in the folder yet. Easiest fix: run DownloadAndLoadBlip once (it fetches the weights automatically), then switch to this node for subsequent runs - or keep using DownloadAndLoadBlip directly if a download-capable graph is fine for your setup. On an offline or sandboxed environment where auto-download can't reach the internet, place the .pth file in the BLIP models folder by hand.

Should I even bother with the loader-only version? If your environment has reliable network access and you don't mind the download step being part of the graph, DownloadAndLoadBlip is simpler - one node does both jobs. This one exists for cases where you want the load step separated from the fetch step, or where fetching isn't possible at all.

Is BLIP even a good captioner? For quick, in-workflow captions - seeding an img2img prompt, a rough alt-text, a sanity check on what a model "sees" - it's fine and it's fast. For LoRA training-data captions specifically, the community consensus has moved firmly away from it: BLIP produces formulaic, often inaccurate captions, and the standard advice is JoyCaption or Florence 2 for natural-language captions, WD14 for Danbooru-tag training, or captioning by hand for small datasets. Use this loader for workflow captions; don't lean on it to caption a LoRA dataset.

Node missing after importing a workflow. The pack isn't installed. ComfyUI Manager's Install Missing Custom Nodes pulls in the whole set, including both BLIP loaders.