Dirty Realism Krea2 UNCENSORED!!!!
About this version
Dirty Realism Krea2 V3
in this version 3, I uploaded the INT8 and INT8-CONVROT quantized models, as well as the FP8 version.
Version 3 is focused on one goal: bringing AI-generated images even closer to real photography.
This update delivers a major improvement in high-resolution generations, producing cleaner results with significantly fewer artifacts while preserving the raw, documentary-style aesthetic that defines Dirty Realism. The grain has been carefully refined, making it feel more natural and photographic instead of distracting, resulting in images that are cleaner without losing their authentic character.
Overall realism has taken another step forward. Materials, lighting, skin, textures, and environmental details now appear more convincing, creating images that feel less like AI generations and more like genuine photographs captured in the real world.
What's Improved
Significantly better performance at higher resolutions.
Major reduction of visual artifacts and unwanted distortions.
Refined photographic grain for a cleaner, more natural look.
Improved realism across lighting, materials, and textures.
More believable photographic results while preserving the signature raw aesthetic.
Better overall consistency across a wide variety of prompts.
This release comes shortly after Version 2 for a reason. Over the past days, the model has been refined through up to 10 hours of daily testing, training, and fine-tuning, allowing rapid improvements to be validated and released as soon as they proved consistently better. Rather than waiting months between updates, the goal is to deliver meaningful improvements to the community as soon as they're ready.
Dirty Realism Krea2 V3 is the most polished and realistic version yet, delivering cleaner images, stronger high-resolution performance, and an even more convincing photographic look while preserving the imperfect authenticity that makes real photography feel alive.
Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 5 denois 0.10 Hires CFG 1,
Width and height: 896*1152 or Hires 1152*1440
CFG Scale: 1
Steps: 8
The sampling method i use is ER SDE-Simple, but you can try a few more, such as Euler-Beta Uniform or Euler A-Simple
Important update: in this version 2, I uploaded the INT8 and INT8-CONVROT quantized models, as well as the FP8 version.
About V2
Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 5 denois 0.20 Hires CFG 1,
Width and height: 896*1152
CFG Scale: 1
Steps: 8
The sampling method i use is ER SDE-Simple, but you can try a few more, such as Euler-Beta Uniform or Euler A-Simple
Dirty Realism Krea2 V2
The next evolution of authentic AI photography is here.
Dirty Realism Krea2 V2 is not just an update. It's a complete refinement of the philosophy behind the model. Every aspect has been revisited with one goal in mind: making AI images feel less artificial and more like genuine photographs captured in the real world.
Version 2 pushes realism further than ever before. The model produces richer textures, more convincing lighting, natural skin, authentic facial features, and a level of environmental detail that makes every image feel lived in rather than rendered. Instead of chasing flawless beauty, it captures the subtle imperfections that our eyes instinctively associate with reality.
The signature raw aesthetic has been significantly enhanced. Images now feature more believable camera behavior, realistic exposure, organic shadows, subtle sensor noise, natural motion blur, imperfect framing, and the tiny inconsistencies that transform an AI generation into something that could easily be mistaken for a real photograph.
What's New in Version 2
Dramatically improved raw photographic realism.
More natural skin tones with realistic pores, texture, and imperfections.
Better facial anatomy and human expressions.
Enhanced lighting with richer highlights, softer transitions, and realistic shadows.
Superior material rendering including fabric, concrete, metal, glass, and weathered surfaces.
Increased environmental storytelling with scenes that feel naturally inhabited.
Improved depth, atmosphere, and cinematic composition.
More convincing handheld camera aesthetics and documentary-style photography.
Better consistency across a wide variety of prompts.
Stronger performance in portraits, street photography, urban exploration, interiors, lifestyle, fashion, industrial scenes, and cinematic storytelling.
Why Download Dirty Realism Krea2 V2?
If you've ever felt that AI images look too clean, too polished, or instantly recognizable as synthetic, this model was created specifically to solve that problem.
Dirty Realism Krea2 V2 focuses on authenticity over perfection. It embraces the subtle chaos of real photography: imperfect lighting, spontaneous moments, natural textures, environmental wear, and the visual imperfections that make images emotionally believable.
Whether you're creating cinematic scenes, documentary photography, fashion editorials, architectural visualization, urban environments, or everyday moments, this model consistently delivers images with remarkable realism and character.
Every generation feels like it has a story behind it.
Not because it tries to imitate reality.
Because it understands what makes reality convincing.
Download Dirty Realism Krea2 V2 and experience a model designed for creators who believe the most realistic images aren't the perfect ones... they're the ones that feel impossible to fake.
Most AI image models chase perfection. Dirty Realism Krea2 does the opposite. It embraces the beautiful imperfections that make photographs feel alive. Tiny flaws, natural lighting, realistic textures, imperfect framing, authentic skin, motion blur, sensor noise, and candid compositions all come together to create images that look captured, not generated.
Built on Krea2, this model is designed to produce scenes with genuine atmosphere instead of sterile perfection. Every image feels like it has a story behind it, whether it's a spontaneous street photo, an old smartphone snapshot, a documentary frame, or a cinematic moment frozen in time.
Why you'll love it:
Natural, believable realism instead of overprocessed AI aesthetics.
Exceptional lighting and texture reproduction.
Authentic skin, clothing, and environmental details.
Beautiful imperfections that increase photographic credibility.
Strong performance for portraits, street photography, documentary, lifestyle, urban scenes, and cinematic compositions.
Consistent results while maintaining a raw, organic look.
If you're tired of images that scream "AI-generated," Dirty Realism Krea2 was built for you. It delivers photographs with personality, atmosphere, and the subtle imperfections that convince the eye.
Download it, experiment with it, and discover how realistic AI imagery can become when perfection is no longer the goal.
Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 5 denois 0.20 Hires CFG 1,
Width and height: 896*1152
CFG Scale: 1
Steps: 8
The sampling method i use is Euler-Beta, but you can try a few more, such as ER SDE-SGM Uniform or Euler A-Simple
Since it is a turbo model, it is a good idea to use the recommended parameters.
I’ve never really liked putting my models behind early access, but if you’re able to donate some Buzz, it’s appreciated; if not, feel free to enjoy the model anyway. I really look forward to seeing your creations. This is my first Krea2 model, so I imagine there might be a few flaws that we hope to iron out over time.
Description
Dirty Realism Krea2 V3
Version 3 is focused on one goal: bringing AI-generated images even closer to real photography.
This update delivers a major improvement in high-resolution generations, producing cleaner results with significantly fewer artifacts while preserving the raw, documentary-style aesthetic that defines Dirty Realism. The grain has been carefully refined, making it feel more natural and photographic instead of distracting, resulting in images that are cleaner without losing their authentic character.
Overall realism has taken another step forward. Materials, lighting, skin, textures, and environmental details now appear more convincing, creating images that feel less like AI generations and more like genuine photographs captured in the real world.
What's Improved
Significantly better performance at higher resolutions.
Major reduction of visual artifacts and unwanted distortions.
Refined photographic grain for a cleaner, more natural look.
Improved realism across lighting, materials, and textures.
More believable photographic results while preserving the signature raw aesthetic.
Better overall consistency across a wide variety of prompts.
This release comes shortly after Version 2 for a reason. Over the past days, the model has been refined through up to 10 hours of daily testing, training, and fine-tuning, allowing rapid improvements to be validated and released as soon as they proved consistently better. Rather than waiting months between updates, the goal is to deliver meaningful improvements to the community as soon as they're ready.
Dirty Realism Krea2 V3 is the most polished and realistic version yet, delivering cleaner images, stronger high-resolution performance, and an even more convincing photographic look while preserving the imperfect authenticity that makes real photography feel alive.
Upscaler: Im using 4x-ClearRealityV1 upscale by 1.5 Steps 5 denois 0.10 Hires CFG 1,
Width and height: 896*1152 or Hires 1152*1440
CFG Scale: 1
Steps: 8
The sampling method i use is ER SDE-Simple, but you can try a few more, such as Euler-Beta Uniform or Euler A-Simple
FAQ
Comments (23)
I went to download the checkpoint, and I noticed that all three have the same file name. Would you please double-check the file names/verson?
The same file name is just how Civitai works. You have to rename them if you want to download the different quantizations.
@jmblsmit129 Thanks for the clarification; I did make sure to upload the models with the correct quantization names, but I wasn't aware of this new thing with Civitai.
@Sateluco It's a bit annoying to be fair.. if you hover over the names in the download picker you can the original names in the tool tip.
Just tested the V3. I don't know what people merge into the mix that simply destroys the image quality. This is one of those models. My guess is some lora with a lot of grain. I really don't know who started this grainy trend. It just destroys the quality and sharpness.
Yeah, retro aesthetics like CCD style or Polaroid should really just be handled by a LoRA. The base model only needs to focus on nailing the composition and content generation. The model itself is actually quite capable; it's just that the issues you mentioned are indeed holding it back.
how many steps and resolution u recommend for v3
10 steps, i have very good results and the gallery examples are 1152x1440
Very nice. Handeled everything I threw at it.
Nice model, thanks for it! 👍
One remark tho, you say that you use 4x-ClearRealityV1 to upscale your pics. Yet 4x-ClearRealityV1 is supposed to produce aritfacts when there is bokeh, DoF or grain in a picture according to its creator. So are you sure that is the the best choice in your workflow for as far as your model is supposed to aim at a realistic style that includes film grain, bokeh and DoF?
Thanks a lot for the clarification; in v1, I started out using Clear Reality, but later switched to generating directly at high steps. I might need to remove that recommendation until I’ve done more testing.
4-bit versions (or int4-int8 hybrid) will be appreciated.
seems like bad naming for version V3 , it has V2 in name....
In V3 prompt following improved but quality still very bad. Image is noisy and have visible pixels grid.
this model really is one of the best in terms of image quality, shadow and lighting and composition, too bad that V3 has a really rough problem with noise and visible grids, otherwise it would easily the best model to use in most cases.
V3:
Good colors. Good composition.
But something is wrong with textures / details for the first pass (Text-to-Image). I see a very unpleasant repeatable patterns on the textures with smaller details like hairs (head and body hairs).
This is not specific to this model. I see it on many Krea2 Turbo models.
Fortunately it can be easily fixed with SDXL models. See my post below and drop the image into ComfyUI to see the actual workflow.
720x1080 image is the Krea2 SAT-DirtyRealism.
1080x1616 is the HiResFix with CinEro-NG XL v1.
I'm definitely going to try it; it sounds like an interesting proposition.
@Sateluco I'm trying to train the Krea2 with dataset used for CinEro NG v1. It is very sad that I cannot use your model as a base (OneTrainer don't accept it no matter what I was trying to do). Hopefully I will fix that common issue with hairs and repeatable pattern. As for now we can fix it with any realistic SDXL checkpoint.
Yet another tasty showcase ^___^
https://civitai.red/images/138807098
Love this model. I was blown away at what it could do. Can't recommend enough. Home-run.
Does this play well with character LORAs made with RAW MODEL



















