Diffusers Load Checkpoint

Most of this pack's image-generation nodes run through the HuggingFace diffusers library instead of ComfyUI's native sampling, and diffusers wants its models in its own format. UL_DiffusersCheckpointLoader bridges that gap: it takes an ordinary single-file .safetensors/.ckpt checkpoint - the kind you already have in models/checkpoints - and loads it as a Diffusers_Model the pack's diffusers sampler can use. Per the node's own note, it handles sd1.5, sd2.1 and sdxl, cosxl included.

If you're running the JoyType path (the diffusers sampler with a canny ControlNet), this is the loader that feeds it.

How it works

Diffusers normally expects a multi-folder repo layout, not a single blob. This node does the single-file conversion in memory - reading your checkpoint and assembling the UNet, text encoder(s) and VAE into a diffusers pipeline object - so you don't have to keep a separate diffusers-format copy on disk. It detects the architecture (SD1.5 vs SD2.1 vs SDXL) from the checkpoint. That's the whole value: use the checkpoints you already have, in the format diffusers needs.

The inputs and outputs that matter

• ckpt_name - the checkpoint to load, from ComfyUI/models/checkpoints. Empty (None) on this listing because the crawl box had none installed.
• unet_only (BOOLEAN, default off) - load just the UNet, skipping the bundled text encoder and VAE. Useful when you want to supply those separately or save memory, but for a normal run leave it off so you get a complete pipeline.
• dtype (default auto) - precision. auto is fine; fp16 halves the memory footprint, which SD 1.5 barely needs but SDXL appreciates.

Outputs: diffusers_model (Diffusers_Model) into the diffusers sampler, and ckpt_name (STRING) for logging.

How to install it

ComfyUI Manager, search ComfyUI_Anytext, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zmwv823/ComfyUI_Anytext

then restart. Put your SD 1.5 / SD 2.1 / SDXL checkpoints in ComfyUI/models/checkpoints as usual and they'll appear in ckpt_name. The diffusers library ships with the pack's requirements.

Common issues & troubleshooting

Dropdown is empty. No checkpoints in models/checkpoints. This loader reads the same folder as every other checkpoint node - download one and it'll show up.

Load fails or the model looks wrong. The single-file conversion has to recognise the architecture. A merged or oddly-formatted checkpoint can trip it up where a stock SD1.5/SDXL model wouldn't. If a specific checkpoint won't load, try a vanilla one to confirm the node works, then suspect the file.

unet_only gave me a broken pipeline. That switch deliberately drops the text encoder and VAE. If you enable it, you need to supply those elsewhere in the graph - otherwise the sampler has no way to encode prompts or decode latents. For a standard run, keep it off.

Why not just use a native loader? Because the diffusers sampler and the JoyType path need a Diffusers_Model, not Comfy's MODEL type. This loader exists specifically to serve those diffusers nodes; for native ComfyUI workflows, your normal checkpoint loader is the right call.