Poll: Lightning vs Hyper-SD

[mashb1t]

Quick check for #2812 / #2812 comparing Lightning 4 steps LoRA with Hyper-SD 4 steps LoRA.
Images of each round were created using the same model / config / seed / styles / etc., only difference is the performance setting.

I'll post which one is which in a few days.
A is Lightning, B is Hyper-SD.

You can find the comparison between Extreme Speed (LCM) and Lightning here: #2416

Round 1

A

B

Round 2

A

B

Round 3

A

B

Round 4

A

B

Round 5

A

B

[xhoxye]

sampler_name = 'ddim', scheduler_name = 'sgm_uniform'
Judging by my own test results, this combination of parameters is not optimal.

[xhoxye]

Fooocus v2.3.1
juggernautXL_version6Rundiffusion.safetensors
styles = none
seed = 123

A female magician wearing a magic hat,Sitting at the street coffee table drinking coffee

[wangfeng35]

It seems current version tends to "burn" the image when CFG is high.
According to the team, their CFG-Preserved Hyper-SD15/SDXL that facilitate negative prompts and larger guidance scales (e.g. 5~10) will be coming soon! Looking forward to it.

[wangfeng35]

They just uploaded 8-steps CFG-preserve LoRA for SDXL (support 5~8 guidance scales). 4-steps version will be coming soon.

[Irmagon]

My preset, after ~200 generation:

"default_model": "juggernautXL_v8Rundiffusion.safetensors",
"default_loras": [["Hyper-SDXL-4steps-lora.safetensors",0.8]],
"default_cfg_scale": 1,
"default_sample_sharpness": 1.0,
"default_overwrite_step": 4,
"default_sampler": "dpmpp_sde_gpu",
"default_scheduler": "karras",
"default_performance": "Speed",
"default_styles": ["Fooocus V2"],
"default_aspect_ratio": "1024*1024",

[mashb1t]

@Irmagon thanks for the hint, i can confirm that these settings work way better contrary to the suggested ones.

LoRA weight 0.8
sampler dpmpp_sde_gpu
scheduler karras

all one shot images, no best-of-4, tested with Juggernaut v8 Rundiffusion:

[ILikeToasters]

Looks like they have some new 8-step LoRAs out with 4-step ones coming soon.

[mashb1t]

@ILikeToasters @xhoxye @Irmagon while waiting for the 4step CFG LoRA I've tested the 8step CFG one. You can find a preset draft here:
hyper-sd8cfg.json

This preset is not intended to be implemented as default, but just for evaluation.
The image is generated with quite a high amount of artifacts when using LoRA weight > 0.4, which is why 0.3 is used here. It's less about the actual CFG. Negative prompts do work, but, while being really good, i wouldn't necessarily say that it replaces normal SDXL 30+ step rendering as ByteDance claims on their HF page.

we strongly recommend making the 8-step CFGLora a standard configuration for all SDXL and SD15 models

Happy to hear your thoughts on this one.

ps this is basically the same as just adding a new performance with 8 steps. Code see commit 8ef1a21.
Performance overview with examples => https://gallery.mashb1t.de/gallery/5aO0qQP1NQjp8r1SoVJKUIWc

[chouhai2018]

太实用了,我调好了,1秒1张1024x1024高清

[mashb1t]

Hyper-SD 4 step LoRA is now part of release 2.4.0-rc1.
Why not 8/12 step CFG LoRA? => improvement in image quality is not noticable (subjective) and i'm still waiting for the 4step CFG LoRA for best comparability.
Open ToDo's: see #2955 (comment)