Load Checkpoint Mix With Diffusers

At a glance this is just the standard Load Checkpoint node with a longer name: one ckpt_name dropdown in, unet, clip, vae out. And for any normal workflow you'd be right to use the standard node instead. This one exists for exactly one reason - the SelfGuidanceSampler in the same pack runs the UNet through diffusers' scheduling machinery, and it needs a model that's been loaded the diffusers way.

Despite "Mix" in the name, it mixes nothing. What it actually does is load your checkpoint through diffusers' from_single_file() path (StableDiffusionPipeline for SD1.5, StableDiffusionXLPipeline for SDXL), which is what gives you the diffusers-format unet output the sampler expects. Meanwhile it uses ComfyUI's own loader to detect the model type and pull out native CLIP and VAE weights. So you get the same three outputs as a regular loader, but the unet is the special flavor.

It only knows two model families. Detection maps the checkpoint's latent format to SD1.5 or SDXL; anything else errors out with a "not in the supported correspond dict" message. No Flux, no newer architectures - the whole pack is SD-series-only by design.

The one input: ckpt_name, populated from your ComfyUI/models/checkpoints folder. Pick your SD1.5 or SDXL checkpoint and wire it into the pack's standard graph: unet → SelfGuidanceSampler, clip → CLIPConditioning, vae → the encode/decode side. That's the whole setup.

Install. It's part of the pack, so the pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/forever22777/comfyui-self-guidance

or ComfyUI Manager → search "self-guidance," then restart. This node is why the pack pins diffusers==0.26.3 - no diffusers, no load. If you see an import error mentioning diffusers right after install, that pin is the first suspect, especially in an environment with other nodes that want a different diffusers version. That dependency conflict risk is the main reason not to use this loader when you don't need the sampler.