Diffusers CLIP Loader

Every modern model has a text encoder - the thing that turns your prompt into the conditioning the denoiser understands. In a diffusers-format model folder, that lives in text_encoder/ (and text_encoder_2/, text_encoder_3/ on bigger architectures), and Diffusers CLIP Loader is the node that pulls it out without you having to convert anything or hunt down a separate CLIP file.

What it does

Pick a model folder from the dropdown, and the loader reads its model_index.json, figures out the architecture, then loads the right text encoders:

• SD 1.5 / SD 2.1 - the single text_encoder/ directory.
• SDXL - text_encoder/ plus text_encoder_2/.
• SD3 - all three, including text_encoder_3/, which is the T5 and often ships sharded across multiple files that the loader reassembles through its index.json.
• Flux - the CLIP encoder plus the big T5 in text_encoder_2/, handled via its index file so sharded checkpoints work.

It also registers the model's own embeddings/ folder as the textual-inversion directory, so any baked-in embeddings stay available. The output is a standard ComfyUI CLIP object - wire it into your CLIP Text Encode node exactly like you would with any other loader.

The inputs

Just two, which is why this node is so easy to live with:

• sub_directory - the dropdown of diffusers folders ComfyUI can see (same list as every node in this pack).
• clip_type - stable_diffusion, stable_cascade, sd3, stable_audio, sdxl or flux. The one that matters is flux: that's what switches the loader into T5-inclusive mode, and getting this wrong is the classic way to end up with garbage conditioning. For SD3 you want sd3. For SDXL it barely matters - stable_diffusion and sdxl resolve to the same CLIP type internally - so leave the default.

Install & models

Same story as the whole pack - ComfyUI Manager, search DiffusersLoader, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Scorpinaus/ComfyUI-DiffusersLoader

Restart, and you're done. No extra dependencies. Model folders go in ComfyUI/models/diffusers/ or any extra base path you add to extra_model_paths.yaml under a diffusers: key.

Troubleshooting

• Wrong clip_type - if you're getting nonsense prompts out of a Flux or SD3 model, this is the first thing to check. Flux wants flux, SD3 wants sd3.
• "Selected directory does not exist" - the folder vanished from under the dropdown, or you moved the model after the list was built. Refresh the node.
• Flux T5 is a memory hog - that's the model, not this node. The text encoder is huge; if it OOMs, consider running the whole model in fp8 or offloading.

Honest take: this is the least glamorous node in the pack, and that's fine. If you're using the Combined loader you'll never touch it directly - but if you're assembling a workflow by hand and want a fully diffusers-native pipeline, knowing this one exists saves you from converting or relocating a CLIP you already have.