Sign In

Stable Diffusion 3.5 with ComfyUI

Stable Diffusion 3.5 with ComfyUI

Stable Diffusion 3.5 (Large) has arrived !

First lets cover what is needed and where to put the files.
(Cover Image used for this article is SD3 Workflow from ComfyAnonymous)

My new "Thoda Workflow Pack for SD three point five" is also available soon with some basic workflows, expect more as things progress over the coming days and weeks.

CLIP MODELS

Place inside ComfyUI/models/clip/
Clip_L: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_l.safetensors
Clip_G: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_g.safetensors
T5XXL: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp16.safetensors

As with a few other models we saw this year, you can reduce the VRAM load (32GB @FP16),
by choosing to use the T5XXL_FP8_e4m3fn_scaled in place of the FP16 version:
https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp8_e4m3fn_scaled.safetensors
note: a different file from the one used with FP8 Flux.

SD3.5 Weights

Place inside ComfyUI/models/checkpoints/
SD3.5_Large: https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main

SD3.5_Turbo: https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo/tree/main

recommended for turbo: "Steps = 4, CFG = 1.2"


Comfy-ORG has provided a combined FP8 checkpoint, which contains all three clip models like most regular checkpoints for ComfyUI.
SD3.5_Large_FP8_scaled: https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/sd3.5_large_fp8_scaled.safetensors

SD3.5 VAE

Place inside ComfyUI/models/vae/
The VAE is included as part of the weights checkpoints, if you need it separately, it is only 166mb.

SD3.5_VAE: https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/vae/diffusion_pytorch_model.safetensors
I Renamed mine for convenience to "SD3_vae.safetensors"

Triple Clip (loaded separately)
using SD3.5 Large (FP16) with T5XXL (FP8): (same seed) 714772221606588

FP8_scaled (combined checkpoint)

using SD3.5 Large (FP8) with T5XXL (FP8): (same seed) 714772221606588

SD3.5-Turbo
using SD3.5 Large (FP8) with T5XXL (FP8): (same seed) 714772221606588


SD3.5 Large "Full Fat" edition

I have added a txt2img and img2img workflow which used only FP16 for the model and the CLIP, where the Clip is offloaded to CPU, this is the highest quality setup, but requires both high VRAM and high system RAM. I saw 48GB loading Triple CLIP's into RAM.

Remember to change the Denoise value when using the img2img workflow !



SD3.5 Medium

All the default medium workflows for SD3.5 Medium have been fixed for quality and consistency, also i have changed the upscaling example to use only the Medium model. They recommended to use 40 steps with 3.5 CFG, but the outputs significantly improve from 20 steps and 7 CFG as the only real change. Added all the creature comforts from my other workflows.

For now I have only added the "Opti" workflows (optimized/fixed examples)

12

Comments