Blip_Loader

Blip_Loader exists for one reason: so CSGO_Sampler can describe your content image for you. Instead of writing a prompt that matches the subject in your content image, you let BLIP look at it and write the caption itself. It's the "LLM" that the pack's README calls optional - genuinely optional, and the source of this pack's most confusing gotcha.

What it is

It's a thin loader around Salesforce's blip-image-captioning-large, a 2023 vision-language model that's still a default lightweight captioner in a lot of pipelines. One input, two outputs. The single required input, blip_repo, is just a HuggingFace repo id - default Salesforce/blip-image-captioning-large. On first run it downloads the weights into your HF cache (needs internet; there's no ComfyUI model folder involved). After that it's cached.

The outputs are blip_processor and blip_model. Note the schema types them as MODEL, which is a lie in ComfyUI terms: the first is a transformers BlipProcessor and the second a BlipForConditionalGeneration. They only fit the two optional inputs of CSGO_Sampler - don't try to wire them into anything normal, because nothing else accepts them.

The gotcha

The moment both outputs are connected to CSGO_Sampler, BLIP captions your content image and that caption replaces your prompt field. The text you typed is silently ignored. This is the classic "why isn't my prompt doing anything" trap with this pack - if the output doesn't match what you typed, unplug the Blip_Loader and type the prompt yourself.

It's also not free to run: BLIP-large is roughly a gigabyte of extra weights loaded on top of the already-heavy SDXL pipeline the sampler runs, and the captioning adds a forward pass before generation even starts. On a tight GPU, skipping BLIP and typing your own prompt is the leaner call.

Install

Same pack, same routine: ComfyUI Manager → search ComfyUI_CSGO_Wrapper, or git clone https://github.com/smthemex/ComfyUI_CSGO_Wrapper into custom_nodes, then restart. It runs on transformers, which this pack's commented-out requirements.txt assumes you already have - if it fails on import, pip install transformers.

Where people get burned

• No internet, no BLIP. It downloads from HuggingFace at runtime, so air-gapped setups fail at the loader. Just skip it and type the prompt - the sampler falls back to your text.
• Your prompt ignored - covered above; the cause is a connected Blip_Loader.
• "MODEL" is a lie. The outputs aren't interchangeable with checkpoints or CLIP models. They fit exactly one socket: CSGO_Sampler's optional blip_processor / blip_model inputs.