Realism from HaDeS XL v12

Quick Improvements:

More artistic: Recognizes many more styles.

More flexible: Compatible with more samplers and schedulers.

Less rigid: Much more creative.

More tokens: Responds very well to long prompts.

🖼️ Realism from HaDeS XL — 2nd Anniversary Special Edition

Two years ago, the very first version of Realism from HaDeS XL set out to redefine what realism could look like in AI art. Since then, the model has grown, evolved, and been shaped by both innovation and community feedback.

This Anniversary Edition celebrates that journey — carrying forward all the strengths and versatile styles of v10, while refining and enhancing realism to address the concerns some users raised with v9. The result is a sharper, more consistent, and more immersive model that honors its roots while pushing the boundaries of visual fidelity even further.

🙏 A heartfelt thank you to everyone who has supported Realism from HaDeS XL throughout these two years. This release is dedicated to you — the community that has kept the model alive and growing.

🖼️ The sample images included here are a collage of the highest-rated works from across all previous versions, serving both as a celebration of the journey so far and a showcase of the legacy this model has built.

Whether you’re looking for faithful photorealism, cinematic moods, or artistic versatility, this release is both a tribute to the past and a leap into the future of the HaDeS ecosystem.

🖼️ Realism from HaDeS XL v10.2 - Artistic Precision & Versatility

Realism from HaDeS XL v10.2 takes the v9 foundation and pushes it further — delivering unmatched prompt adherence, broader stylistic range, and an even higher level of artistic and technical quality. This version bridges pure photorealism with creative versatility, making it equally capable of producing lifelike editorial portraits, cinematic storytelling, and fine-art inspired compositions.

Finetuned with an expanded dataset and advanced balancing, v10.2 offers cleaner outputs, better composition control, and the ability to seamlessly adapt to a wide variety of styles — all while preserving the hallmark realism that defines the HaDeS XL line.

🌟 Highlights:

Unmatched prompt adherence – Captures the intended mood, style, and details with high fidelity.
Versatile style handling – From ultra-realistic photography to painterly or cinematic looks, all in one model.
Refined anatomy & posing – Improved accuracy in proportions, hands, and facial features.
Enhanced texture and detail – Lifelike skin pores, hair strands, fabric textures, and lighting depth.
Cinematic depth control – Rich bokeh, authentic DoF, and soft-focus falloff for emotional impact.
Artistic edge – Handles creative, editorial, and fine-art compositions without losing realism.

🔧 Technical Notes:

Based on Realism from HaDeS XL v4.5 with a significantly expanded dataset.
Fully compatible with ComfyUI, Automatic1111, Forge, SDNext, and InvokeAI.
Works best without a negative prompt – let the model express its full spectrum.
Recommended: Steps around 30, CFG 4–7, Sampler: DPM++ 2M SDE Karras, DPM++ 2M, Euler, 832x1216, 896x1152.
Face Detailer and Hires are recommended

Whether you need lifelike humans, stylized editorial shoots, or versatile artistic renders, v10.2 sets a new bar for realism with creative freedom.

🖼️ Realism from HaDeS XL v9.0 – Ultra Photorealism

The ultimate evolution of realism continues.

Realism from HaDeS XL v9.0 brings a new era of ultra-detailed, cinematic-quality photorealism to your generations. Finetuned with an advanced dataset and enhanced visual balance, this model delivers strikingly lifelike portraits and scenes with natural lighting, rich textures, and incredibly accurate human anatomy.

This version builds upon the success of v8.2, refining every aspect to achieve more consistent results, better composition control, and a stunning editorial or magazine-grade aesthetic.

🌟 Highlights:

Medium-format quality inspired by real Hasselblad camera profiles
Exceptional skin texture fidelity with true-to-life pore detail and subsurface scattering
Improved handling of body proportions, hands, and facial structure
Cinematic depth of field, authentic bokeh, and soft-focus falloff for high-end portraiture
Optimized for high resolution vertical framing, ideal for storytelling visuals and character design
Works best without a negative prompt – just let the model express its full potential
⚠️ NSFW-capable by design – The model natively supports tasteful and explicit content generation with high realism and detail. However, no NSFW examples will be posted here – that’s for users to explore on their own.

🔧 Technical Notes:

Based on Realism from HaDeS XL v8.2
Fully compatible with ComfyUI, Automatic1111, Forge, SDNext and InvokeAI
No need for LoRAs – just plug and prompt
Best results at steps: 20-30, CFG: 5-7, sampler: DPM++ 2M SDE Karras

Whether you're crafting deeply human characters, high-end fashion editorials, or emotionally resonant portraits, v9.0 is your new Stable Diffusion XL gold standard.

Realism From HaDeS XL v8.2

Hey, I'm back.

I've been quite busy with my own matters and haven't posted or uploaded anything for quite a while.
But that's about to change since I'm finishing a pretty important project that has taken me a long time.

In the meantime, I bring you an update of Realism from HaDeS, this time version 8.2, which I've been personally using over the past few months to create images related to the project I mentioned above. Now that I’ve practically finished it, I’m taking the opportunity to share it with all of you.

I hope you like it! Show it some love and share your images on the wall. Thanks!!

Recommended Settings:

Width/Height: Recommended 832x1216

Sampler: Euler, DPM++ 3M SDE Karras, DPM++ 2M SDE Karras

Steps: 33 is a good balance between speed and quality

CFG: Recommended between 5-7

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 + Upscale x2

Realism From HaDeS XL v7.0.3

Features updated:

Huge prompt adherence model, responds to both simple and complex prompts. (You can generate high-quality images with just normal language, or alternatively use specific words separated by commas.).
Faster than other checkpoints.
Good results at 20, 30, 40 and 50 steps
Good quality at low and high resolutions
Recognizes and reproduces many styles

The sample images were all generated at low resolution (896x1152), here is the link to the post with images generated at high resolution (1024x1408): https://civarchive.com/posts/6156068

Recommended Settings:

Width/Height: 896x1152 / 1024x1408

Sampler: DPM++ SDE Karras, DPM++ 3M SDE Karras, DPM++ 2M SDE Karras

Steps: betwwen 10 and 50, 20 is a good balance between speed and quality

CFG: 4.5-7.5

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 0.25 Denoise + 1.5 Upscale (or more if your PC can handle it).

See ya in next update.

Realism From HaDeS XL Lightning V4

This checkpoint is a merge of Realism from HaDeS Lightning V3, an experimental custom 4/8-steps SDXL Lightning Checkpoint (unreleased) and a variety of own tools.

The objective of this update was to improve the representation of different styles, weak point of Realism from HaDeS XL Lightning V3.

Realism from HaDeS XL Lightning thus becomes completely a Lightning model. The CFG has been locked to a maximum of 3 to improve the final result, and although it is still possible to use it as a normal model, it is not recommended due to the time required to configure the parameters for each image. Still, it is possible to use it normally, as long as the CFG does not exceed 4.

Recommended Settings:

Width/Height: 1024x1408

Sampler: Depending on the Sampler we use, the number of steps must be adapted. For example, all samplers adapted to work with few steps (SGMUniform, LCM, Turbo, etc.) will need fewer steps (between 4-8), those that are not adapted but give good results with this checkpoint (DPM++ 2M Karras, DPM++ 3M SDE, to give an example), will need between 8-10 steps. See sample images prompts for some more info.

Steps: 4-10

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 or 4x_foolhardy_Remacri with with half the regular steps and 0.15/0.25 Denoise + 1.5/2 Upscale

Finally, I just want to comment that I still recommend the full version of Realism of HaDeS XL v6 in terms of quality. Version 7.x is in beta testing and will be released throughout the month of August maybe. I hope you like the Checkpoint and I'm looking forward to seeing your creations on this wall. Give it approval if you like it and the reviews are greatly appreciated, as they help me a lot to improve

Realism From HaDeS XL v6.0

Question: Is it possible for a checkpoint to have enough quality to do a good job with both few and many steps?

Answer: Yes, of course it is.

Realism from HaDeS XL v6.0 is a fine-tuned checkpoint that in its latest update returns the maximum quality in image generation, regardless of the number of steps.

- Is it a standard model? Is it a lightning model? Is it a turbo model?

- It's all that.

New Features Implemented:

More obedient model that responds to both simple and complex prompts. (You can generate high-quality images with just simple sentences, or alternatively use specific words separated by commas.)
Generate highly detailed images with as many steps as you want, from just 6 steps to 150.
Use the CFG scale that best suits your needs, from 1 to 12 without loss of color.
Improvement in the model's ability to understand the nuances of lights and shadows that we want to implement.
Advanced Color Adjustment.
Inpainting version available.
"/mode bad prompt:x /mode 3D-rendering:x" trained as a negative prompt for higher image quality, where "x" is a value that is assigned between 0.1 and 2 to give more or less realism to the image.

Recommended Settings:

Width/Height: 1024x1408

Sampler: DPM++ SDE Karras, DPM++ 3M SDE Karras, DPM++ 2M SDE Karras

Steps: from 6 to 150

CFG: 1 to 12.5

VAE: Already baked in.

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

TIP for negative prompt: "/mode bad prompt:2 /mode 3D-rendering:2" return images with higher quality.

HiRes: 4x_NMKD-Siax_200k with half the regular steps and 0.25 Denoise + 1.5 Upscale

See ya in next update.

Realism From HaDeS XL v5.7x

Realism from HaDeS XL V5 is finally ready to be enjoyed by everyone. After much testing, LoRa training and injections of enhancement tools, this latest update can be released.

WHAT'S NEW?

Better lighting
Possibility of using it as a limited lightning model
Handles LoRa better than v4.5
It introduces unique image enhancement keywords similar to what Pony does (An extensive tutorial will be released in mid-June), while working like a conventional SDXL model.
Optimized for animations with AnimateDiff (A basic workflow will be available soon)

Recommended Settings:

Width/Height: 1024x1408

Sampler: DPM++ 3M SDE Karras, DPM++ 2M SDE Karras

Steps: Around 30

CFG: 4.5-12.5 (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, I just want to comment that I greatly appreciate the support received, I greatly enjoy your reviews and the images you share on this wall (I hope to see many more). I hope you enjoy this release and give it all your love. See you in v6.x!

Realism From HaDeS XL Lightning V3

This checkpoint is a merge of Realism from HaDeS v5 (unreleased) and an experimental custom 4/8-steps SDXL Lightning Checkpoint (unreleased).

The objective of this update was to improve the weak points of Realism from HaDeS XL Lightning. In this release, the images generated are faster, more colorful and consistent with less steps, thanks to the incorporation of Image Enhancer XL v1 and other improvements.

Realism from HaDeS XL Lightning thus becomes completely a Lightning model. The CFG has been locked to a maximum of 2 to improve the final result, and although it is still possible to use it as a normal model, it is not recommended due to the time required to configure the parameters for each image. Still, it is possible to use it normally, as long as the CFG does not exceed 4.

Recommended Settings for the Lightning Version:

Width/Height: 1024x1408

Sampler: Depending on the Sampler we use, the number of steps must be adapted. For example, all samplers adapted to work with few steps (SGMUniform, LCM, Turbo, etc.) will need fewer steps (between 4-8), those that are not adapted but give good results with this checkpoint (DPM++ 2M Karras, DPM++ 3M SDE, to give an example), will need between 8-10 steps. See sample images prompts for some more info.

Steps: 4-10

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 or 4x_foolhardy_Remacri with with half the regular steps and 0.15/0.25 Denoise + 1.5/2 Upscale

Recommended Settings for the Standard Version:

Width/Height: 1024x1328

Sampler: DPM++ 2M Karras, DPM++ 3M SDE

Steps: Around 30

CFG: 2-4

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 or 4x_foolhardy_Remacri with half the regular steps and 0.15/0.25 Denoise + 1.5/2 Upscale

Finally, I just want to comment that I still recommend the full version of Realism of HaDeS XL v4.5 in terms of quality. Version 5.x is in beta testing and will be released throughout the month of June. I hope you like the Checkpoint and I'm looking forward to seeing your creations on this wall. Give it approval if you like it and the reviews are greatly appreciated, as they help me a lot to improve.

Realism From HaDeS XL v4.5

This update is the next evolution of Realism from HaDeS v3.5 EVO and is intended to be the transition to the next v5.x that will be released in several weeks and which will bring many improvements and new features. This update is focused on getting the highest possible quality from the generated images.

Recommended Settings:

Width/Height: 1024x1408

Sampler: DPM++ 3M SDE Karras

Steps: Around 30

CFG: 4.5-12.5 (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, I just want to comment that I greatly appreciate the support received, I greatly enjoy your reviews and the images you share on this wall (I hope to see many more). Version 5.x is now underway and I promise you that it will surprise you. Unfortunately, it will take several weeks due to the little time I can dedicate to this project, something I do as a hobby and that I love. In the meantime I hope you enjoy this release and give it all your love. See you in v5.x!

Realism From HaDeS XL Lightning V2

This checkpoint is a merge of Realism from HaDeS v4 (unreleased) and the 8-steps release of SDXL Lightning (ByteDance/SDXL-Lightning · Hugging Face).

The basic idea of this update was to work on an improved version of Realism from HaDeS XL Lightning. In this release, the images generated are much more detailed, thanks to the flexibility of the CFG variable, which works between values 2-4 (instead of 1-2). At the same time, it is possible to use several different samplers and obtain very good results (Although I still recommend DPM++ 3M SDE, for me the most stable).

Recommended Settings for the Lightning Version:

Width/Height: 1024x1408

Sampler: DPM++ 3M SDE (Others give very good results too, but I'm still in the testing phase)

Steps: 8-10

CFG: 2-4

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. Although I clarify that in this update, I have not used them yet and they are pending testing. Next, some settings that works especially well in the previous version, 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: Around 40

CFG: 4-12 (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, I just want to comment that I still recommend the full version of Realism of HaDeS XL v3.5 EVO in terms of quality. Version 4.x is being tested and will be released throughout the month of April. I hope you like the model and I'm looking forward to seeing your creations on this wall. Give it approval if you like it and the reviews are greatly appreciated, as they help me a lot to improve this checkpoint.

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 .

Files with SHA256: 1926677a9cec689d1e1d8e09f4cc64d11eb680cd9c4f8ff27464f500e684a2b0