SD3.5 Support

[Acly]

Stable Diffusion 3.5: Current State

Plugin version: 1.27.0

Experimental Support

It's new, so there aren't any inpaint, control or other supplementary models out yet to support all the features. But you can try out the base model and LoRA already.

Model Download

Currently SD3.5 Large (8B) is released. It typically comes as a checkpoint with integrated VAE, but no text encoders. That means text encoders have to be downloaded separately.

Checkpoint

There is a smaller fp8 version which required less VRAM:

• sd3.5_large_fp8_scaled.safetensors
• or use the official files at huggingface/stabilityai

The models must be placed into the ComfyUI/models/checkpoints folder.

Text Encoders

SD3.5 supports up to 3 text encoders.

• Clip-L (clip_l.safetensors)
• Clip-G (clip_g.safetensors)
• T5 XXL (t5xxl_fp8_e4m3fn.safetensors)

Text encoder models must be placed into the ComfyUI/models/clip folder.

Note that T5 is optional. If you don't have it, the model will still work, but may have slightly reduced text understanding.

If you already have these models (eg. because you are using Flux), you don't have to download them again. They can be shared. GGUF variants of T5 also work.

Future Work

I didn't get around to do much testing yet. Let me know what you find.

[XoseNomad]

Is it possible to add support for SD 3.5 medium? I was trying to use it but it won't work. It is significantly quicker than large model.

[Neeqstock]

Same.
I have all the clip files in the correct folders.
I get this error with SD 3.5 medium. Didn't try with large because my GPU probably won't be able to support it.
(P.S.: it would be awesome if GGUF versions of checkpoints could be supported :))