feat: add sd3.5 medium support (+skip layer guidance)

[stduhpf]

Closes#450

How to Use

Download weights

• Download sd3.5_medium from https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/blob/main/sd3.5_medium.safetensors
• Download clip_g from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_g.safetensors
• Download clip_l from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/clip_l.safetensors
• Download t5xxl from https://huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/blob/main/text_encoders/t5xxl_fp16.safetensors

Run

SD3.5 Medium

For example:

.\bin\Release\sd.exe -m ..\models\sd3.5_medium.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Medium\"' --cfg-scale 4.5 --sampling-method euler -v

Skip Layer Guidance (SLG)

Variant of Classifier-Free Guidance (CFG) that seems to improve SD3.5 medium's output quality quite significantly (especially anatomy and hands for example), at the cost of slower inference speed (some steps require 3 evals of the model instead of 2).

> sd.exe --model ..\models\sd3.5_medium-q8_0.gguf --clip_l ..\models\clip\clip_l.safetensors --clip_g ..\models\clip\clip_g.safetensors --t5xxl ..\models\clip\t5xxl_q4_k.gguf -p 'A cute anime girl waving at the camera.' --cfg-scale 4.5 --sampling-method euler --steps 30 -t 24 --vae-tiling -W 512 -H 768 --seed 42
vs
> sd.exe --model ..\models\sd3.5_medium-q8_0.gguf --clip_l ..\models\clip\clip_l.safetensors --clip_g ..\models\clip\clip_g.safetensors --t5xxl ..\models\clip\t5xxl_q4_k.gguf -p 'A cute anime girl waving at the camera.' --cfg-scale 4.5 --sampling-method euler --steps 30 -t 24 --vae-tiling -W 512 -H 768 --seed 42 --slg-scale 2.5

	CFG	SLG
output
sampling time (30 steps)	73.46s	79.44s

[dsignarius]

Working nicely on amd igpu using vulkan and cpu

tried Q8_0 and fp16

fp16 model, vae and t5xxl:

[leejet]

Thank you for your contribution.