Realism from HaDeS – v6.1 HQ

Realism from HaDeS v11 HQ

This update is specially made to generate images at 768×768, Use Realism from HaDeS 10.4 if you want to use 512×512 generation.

This update introduces a slight modification to the Stable Diffusion 1.5 layer and allows images to be generated at a higher resolution than the native one with very few errors.

• Improved realistic effect

• Possibility of generating images at 768×768 instead of 512×512

• Greater realism

• Multiple LoRa handling support

• Better color management in automatic 1111 and Forge

• Eye warping bug fixed

• Improved hands

• Minor bugs fixed

• Simple or elaborate prompts

My Favorite Settings:

Sampling Steps: Around 30

Resolution: 768×1152

Sampling methods: Whichever you prefer, I am personally experienced with DPM++ 2M

SDE Karras, DPM++ 2M Turbo, Restart, DDIM, but the usual ones work fine.

Hiers.fix Upscaler: 4x_fatal_Anime_500000_G, 4x-UltraSharp, R-ESRGAN 4x+

Hires steps: It depends, depending on the desired result and the complexity of the image.

Denoising Strength: Between 0.35-0.45

CFG Scale: From 4 to 14. A value between 4 and 7 gives better realism.

Realism from HaDeS v10.4 HQ

This update fixes a lot of bugs in v10.3 as well as replacing several checkpoint concepts.

• Reduced hyperrealistic effect

• Multiple LoRa handling support

• Color saturation bug fixed

• Eye warping bug fixed

• Improved hands

• Greater realism

• Minor bugs fixed

• Simple or elaborate prompts

My Favorite Settings:

Sampling Steps: Around 30-40 (Between 50-75 for complex prompts with a lot of detail)

Sampling methods: Whichever you prefer, I am personally experienced with DPM++ 2M SDE Karras, DPM++ 2M Turbo, Restart, DDIM, but the usual ones work fine.

Hiers.fix Upscaler: 4x_fatal_Anime_500000_G, 4x-UltraSharp, R-ESRGAN 4x+

Hires steps: It depends, depending on the desired result and the complexity of the image.

Denoising Strength: Between 0.35-0.45

CFG Scale: From 4 to 14. A value between 4 and 7 gives better realism.

Realism from HaDeS v10.3

It’s true, there were not going to be any more updates… But… I was not satisfied with the previous and last update. So I took advantage of some things learned to modify some bugs and improve some things. And here is the result:

• better faces

• better colors, lighting and shadows

• Armor from HaDeS v3 injected (experimental)

• Some experimental improvements

Favorite Settings:

Sampling Steps: Around 30 (Between 50-75 for complex prompts with a lot of detail)

Sampling methods: Whichever you prefer, I am personally experienced with DPM++ 2M Turbo, Restart, DDIM, but the usual ones work fine.

Hiers.fix Upscaler: 4x_fatal_Anime_500000_G, 4x-UltraSharp, R-ESRGAN 4x+

Hires steps: It depends, depending on the desired result and the complexity of the image.

Denoising Strength: Between 0.15-0.35

CFG Scale: From 4.5 to 14.5. A value between 7 and 10 for characters gives better results with fewer errors.

Instructions:

1. Turn off face restoration If you haven’t already disabled it. Simple or elaborate prompts (The simpler and more fluid the prompt, the better results). Easier than previous versions with better results.

2. Recommended configuration to get the most out of it? There is no single configuration that is best. But I share some that I personally like the most and that usually give me very good results:

Sampling Steps: Around 30

Sampling methods: UniPC, DPM++ 2M SDE Karras, DPM++ 2M

Hiers.fix Upscaler: 4x_fatal_Anime_500000_G, 4x-UltraSharp, R-ESRGAN 4x+

Hires steps: It depends, depending on the desired result and the complexity of the image.

Denoising Strength: Between 0.35-0.45

CFG Scale: From 4.5 to 14.5. A value between 7 and 10 for characters gives better results with fewer errors.

3. Once again I would like to thank everyone who supported this project. There will be no more updates (unless some major advancement for stable diffusion 1.5 comes out that is worth it).

As always, enjoy it, give it lots of love and fill this wall with your images!!

Release 8.1 HQ

Test, share, give it a lot of love, fill this wall with incredible images and above all… Have as much fun as I had improving this checkpoint day after day.

See ya in the last update…

Release 7.1 HQ

I’m currently immersed in training Realism from HaDeS SDXL v1.0, so I don’t know how many more updates are left to the SD 1.5 model. Probably one or two more, so here’s version 7.1 HQ. Why 7.1 and not simply 7.0? The reason is that since version 6, there is a full 5GB model that I currently use to train some of my LoRa. The x.1 versions are pruned versions.

Instructions:

1. The same recommendations as the previous version, this being a direct evolution of Realism from HaDeS 6.1 HQ. Turn off face restoration. Simple or elaborate phrases or prompts. Even easier to use than the previous version with even better results.

2. Recommended configuration to get the most out of it? There is no single configuration that is best. But I share some that I personally like the most and that usually give me very good results:

Sampling Steps: Around 30

Sampling methods: UniPC, DPM++ 2M SDE Karras

Hiers.fix Upscaler: 4x_fatal_Anime_500000_G, 4x-UltraSharp, R-ESRGAN 4x+

Hires steps: It depends, depending on the desired result and the complexity of the image.

Denoising Strength: Between 0.35–0.45

CFG Scale: 7–7.5

3. I hope you enjoy this release and please review and share your creations and feedbacks. Give it a like if you have fun creating images with it and see you next update.

Release 6.1 HQ

Following the evolution of Realism from HaDeS, the logical step was to touch hyperrealism, and the result of this update suggests great potential for the future. Also, if this checkpoint continues to receive love in the form of positive feedback, there will soon be a SDXL version.

Instructions:

1. turn off face restoration

2. write your best prompts, it doesn’t matter if they are simple phrases or elaborate prompts, this update makes everything much simpler, and practically everything will give you good results.

3. Any recommendation? Well, I think that to get the best out of a checkpoint, each person must use their own tricks, but if you don’t know where to start you can take a look at the prompts in my sample images if they help you. Feel free to use and modify them as you wish.

4. I hope you enjoy this release and share your creations. Give it a like if you have fun creating images with it and see you soon

Release 5.0

After the 4.5 ReDuX, here comes the 5.0. This release is more refined and precise and if the 4.5 was easy to use, this one is even more so. Write prompts naturally and you will achieve great results.

Update 4.5 ReDuX

This update changes the rules of the game. Write a prompt with a simple or complex description without separations or commas. The model has been refined to the point of being very flexible in all (or almost all) styles. I hope you share all your creations so we can continue improving the model.

Release 1.0

The main idea was to mix to get a model to be able to try and dress all my LoRa. As for that, I think the result is more than satisfactory.