CLIP Selector

Same trick as the Checkpoint Selector, applied to text encoders. CLIP Selector lets you pick a CLIP model from a dropdown and hands back a MODEL_INFO bundle describing it - name, hash, Civitai data - without ever loading the weights. In a Sage Utils workflow this is how your metadata records which text encoder conditioned the image, at zero VRAM cost.

Two inputs, and the second one is where the model-family thinking happens. clip_name is the dropdown of CLIP models in your models/clip folder. clip_type is the loader type - and its default, chroma, is a hint about what world this pack lives in. The CLIP loader types track the text-encoder families: Chroma and Flux-lineage models use the T5/CLIP setup, SDXL uses the dual CLIP pair, older SD 1.5 models use single CLIP. If you're encoding for a Chroma-class model (the crowd-funded Flux rework that's become the default recommendation for uncensored work on a modern base), chroma is what you want. If your checkpoint is SDXL-based, you'll usually reach for the Dual CLIP Selector instead, since SDXL wants its text-encoder pair.

The mechanism: the node queries the pack's model-info system (hash + cached Civitai lookup, stored locally) and outputs a clip_info bundle. Nothing touches the GPU. That's the entire point - the actual encoding happens in your CLIP Text Encode node, which loads the encoder it needs; this selector just documents the choice for your metadata constructor.

The selector-vs-loader mental model from the Checkpoint Selector article applies here too: selectors choose and describe, loaders load. In a typical Sage metadata workflow, CLIP Selector feeds Construct Metadata while your real CLIP loading happens where the conditioning is built. Because it's just data, it's cheap to keep in the graph permanently - which is the design intent.

A practical note about the model-info cache: the first time you select a model, the pack may hit Civitai to fetch details, then it's cached in sage_cache_hash.json / sage_cache_info.json under comfyui/user/default/SageUtils/. If a model shows sparse info, run the pack's Model Scan & Report once to warm the whole cache.

Install

ComfyUI Manager → search Sage Utils → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/arcum42/ComfyUI_SageUtils.git
cd ComfyUI_SageUtils && pip install -r requirements.txt

Restart ComfyUI. No downloads, no keys, only dynamicprompts as a dependency. If you don't care about stamped metadata, you can skip this node - but if you're building the full Sage metadata stack, it's the piece that records your encoder choice.