Multi Selector Double CLIP

If you've ever loaded an SDXL or Flux-class model the old way, you know the pain: one loader for the UNet, and then a second loader for the pair of text encoders - because these models don't have one text encoder, they have two. SDXL runs a pair of CLIPs; the Flux generation runs a T5 alongside a CLIP. Multi Selector Double CLIP exists to collapse that whole setup into one node: pick a UNet, pick two CLIPs, pick a VAE, and get a single model_info bundle describing all of it.

It's the "bundle the whole model stack" node for dual-text-encoder models, and it's the middle sibling in the pack's selector family - Double, Triple, and Quad CLIP versions for one, two, three, or four encoders, plus the flexible variant that lets you pick the count at runtime.

How it works

Six dropdowns, all fed by your model folders:

• unet_name - the diffusion model (from your diffusion_models/checkpoints).
• weight_dtype - how the UNet loads (default default; other options for quantized setups).
• clip_name_1 / clip_name_2 - the two text encoders (from models/clip or text_encoders).
• clip_type - the loader type for the CLIP pair, defaulting to sdxl. This is the one to think about: it tells the loader how the two CLIPs are meant to be combined. sdxl covers the classic SDXL dual-CLIP pairing; switch it to match whatever architecture your model expects (T5+CLIP for Flux-class, and so on).
• vae_name - the VAE.

When you run it, the node packages the selections into a model_info bundle and hands you that as the single output. Nothing loads yet - the bundle is a description, and it's consumed downstream by Load Models, which expands into the actual loaders and applies the clip_type correctly.

The inputs that matter

Honestly, only two need your attention on a first run:

• clip_type - get this right or your dual-CLIP model loads wrong. Default sdxl is right for SDXL checkpoints; match it to your model family.
• unet_name - the actual model, which is the thing you'll swap most often.

The rest are straightforward picks from dropdowns.

Where it fits

If you're building a reusable SDXL or Flux workflow, this replaces a small pile of loader nodes with one selector. Change the model by editing dropdowns instead of rewiring the graph. And because the same model_info bundle also feeds Sage's metadata nodes, your saved images record exactly which UNet, CLIPs, and VAE ran - which is the pack's whole deal.

Installing

Standard Sage Utils install:

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

Restart ComfyUI, or grab "Sage Utils" via ComfyUI Manager. No models of its own - it organizes the ones you already have.

Where people get tripped up

The clip_type mismatch is the #1 trap: an SDXL model loaded with a wrong CLIP type fails at encode time with a shape error, and it's easy to blame the model instead of the selector. Also keep in mind this is a two-CLIP node by design - if your model needs three or four encoders, that's the Triple/Quad siblings in the same pack. And remember the bundle doesn't load until it hits Load Models; if you connect this straight to a sampler and see nothing, you're missing the loader in between.