Realism From HaDeS XL Lightning

This checkpoint is a merge of Realism from HaDeS v3.5 EVO and the 8-steps release of SDXL Lightning (ByteDance/SDXL-Lightning · Hugging Face).

The idea was obviously to release a version of Realism from HaDeS XL that everyone could use, even users with less powerful hardware. But in practice, maintaining a dual balance between an 8-step checkpoint and a normal SDXL 1.0 checkpoint without losing quality has not been an easy task.

Recommended Settings for the Lightning Version:

Width/Height: 1024x1328

Sampler: DPM++ 3M SDE

Steps: 8-10

CFG: 1-2

VAE: Already baked in.

Negative Prompt: No negative prompt needed, add afterwards the things you don´t want to see in the image.

HiRes: 4x_NMKD-Siax_200k with 5 Steps and 0.15 Denoise + 1.5 Upscale

The use of FreeU and SelfAttentionGuidance is optional but highly recommended due to the great improvement in Lightning models. Next, a little help about some settings that works especially well in this checkpoint, from here you can tweak it to your liking:

FreeU

B1: 1.34

B2: 1.19

S1: 0.67

S2: 0.27

SelfAttentionGuidance

Scale: 0.5

Blur Sigma: 2

Recommended Settings for the Standard Version:

Width/Height: 1024x1328

Sampler: DPM++ 3M SDE Exponential

Steps: 32-40

CFG: 4-16 (The lower the CFG, the more realism and less details.)

VAE: Already baked in.

Negative Prompt: No negative prompt needed, add afterwards the things you don´t want to see in the image.

HiRes: 4x_NMKD-Siax_200k with half the regular steps and 0.15 Denoise + 1.5 Upscale

Finally, just comment that I still recommend the entire version of Realism from HaDeS XL v3.5 EVO in terms of quality. For the rest, I hope you like the model and I can't wait to see your creations on this wall. Thumbs up if you like it and reviews are highly appreciated.

Update v3.5 EVO

This update marks an evolution in quality and details, thanks to an integrated add-on that I will also publish separately. I hope you like it, enjoy it and fill this wall with a multitude of images of your creations.

Update v3.0

This version 3.0 marks the steps to follow for the future of Realism from HaDeS XL. In this update, the negative prompts are purely aesthetic and are only necessary to improve some specific detail in some images. To demonstrate this, all sample images have been created without a negative prompt and without rescaling.

Tips, recomendations, favourite settings:

Sampler: After testing with a multitude of samplers, DPM++ 2M Karras is my favorite without a doubt. Although DPM++ 3M SDE Exponential gives results with better colors, at high CFG, it also gives many errors.

Steps: 20-50, with 30 being the setting I use the most. CFG: I normally use a scale between 10.5 - 14.5 (more than 14.5 the results are bad, with too many errors), although the model works well between 5 and 14.5, it will depend on the type of images you are looking to create.

VAE: It's already included.

Hires.fix: Not necessary but it can be useful sometimes. I sometimes use 4x-UltraSharp and 4x_fatal_Anime_500000_G.

Upscaled by: It depends on your PC, for my garbager configuration I use 1.5, but on other machines I tried 2 and 2.5 and it rescaled perfectly.

Give this model a lot of love, the reviews, five stars, questions and comments, your images filling the wall are appreciated, and as always... see you in the next update.

Realism from HaDeS SDXL v2.0

Update v2.0

This v2.0 brings many new features, especially in the image quality.

Negative prompts can improve the image somewhat but are no longer mandatory or necessary.
Refineer not necessary
VAE included

Realism from HaDeS SDXL v1.5B

Update v1.5B

Tips and Settings I usually use:

Sampler: DPM++ 3M SDE Exponential, DPM++ 3M SDE Karras

Steps: 32-50

CFG: 7 or 7.5, although the model works well between 5 and 8.5, I do not usually change the scale.

VAE: You don't need any, it already has it.

Hires.fix: Not necessary but it can be useful sometimes. I usually use 4x-UltraSharp but I usually test with 4x_fatal_Anime_500000_G and it usually gives me good results too.

Upscaled by: It depends on your PC, for my garbager configuration I use 1.5, but on other machines I tried 2 and 2.5 and it rescaled perfectly.

Realism from HaDeS SDXL v1.0

Update v1.0

Some bugs fixed.

Realism from HaDeS SDXL v0.9c

Introduction to the v0.9C

The SDXL version of Realism from HaDeS it's almost ready. I thought I wouldn't see the light at the end of the tunnel. Well, it's not a real 1.0 version yet, but it's very close and I wanted to share it with everyone.

A little retrospective before going into more technical details. I started developing this model at the same time as version 5 of Realism from Hades for SD 1.5. I quickly realized, to my dismay, that my PC could not train in SDXL, so I put the project aside for a couple of weeks.

Then I discovered runpod.io and dreamlook.ai and decided to invest time and some money in training this checkpoint since I was using SDXL more and more.

I had problems with a thousand things, with the optimal resolution of the model, the VAE that I should use, I got errors trying to mix the Dreambooth and dreamlook.ai models, I tried a million different solutions, I had to ask for help many times on Discord. .. But... Here we are, and I am very satisfied with the result. The knowledge acquired during this arduous journey has no price.

Tips and Settings I usually use:

Sampler: DPM++ 3M SDE Exponential

Steps: 30 or more

CFG: 7 or 7.5, although the model works well between 5 and 8.5, I do not usually change the scale.

VAE: You don't need any, it already has it.

Hires.fix: I usually use 4x-UltraSharp but I usually test with 4x_fatal_Anime_500000_G and it usually gives me good results too.

Upscaled by: It depends on your PC, for my garbager configuration I use 1.5, but on other machines I tried 2 and 2.5 and it rescaled perfectly.

Finally, I just hope you enjoy it, I'm looking forward to seeing many of your images below and thank you in advance for trying it. Hoping that you give as much love to this version as you do to the Stable Diffusion 1.5 ones. And of course, reading your feedback to complete v1.0.

if you like my work and want to support me, or are interested in information about my upcoming projects or updates to my models, visit: https://ko-fi.com/devildonia .

Hide