Introduction

I decided to run an in-depth test analizing SDXL checkpoints for realism. I used the best photo-realistic models for AI image generation in SDXL. My list of checkpoints is not huge, I tested many checkpoints and I discarted many of them. At the end I will test 30 different models. I am still not sure all the checkpoint in my list are worth using them, and probably there are many other out there that could be much better than those in my list (if you know any good checkpoint for realistic images let me know, thanks!). I exluded from the list the realistic-pony checkpoints. Some of them are really good, but they are still based on Pony and that is a different approach to AI image generation.

For these tests I used SD.Next, as I personally think it’s the best Stable Diffusion generation tool around, based on Auto1111, with a fantastic modern UI. I just love it! I installed it as a package of StabilityMatrix, another incredible and very useful software for anyone who is exploring Stable Diffusion generation models.

You can view some of my artworks generated with SD.Next in my Portfolio page.

The List

These are the 30 checkpoints I used for the test:

1. AlbedobaseXL_v21
2. ColossusProjectXLSFW_10bNeodemonFP16
3. CopaxTimelessxlSDXL1_v122
4. CyberrealisticXL_v21
5. EpicrealismXL_v8Kiss
6. FluentlyXL_final
7. IcbinpXL_v6
8. Imaginarium_v10
9. JibMixRealisticXL_v140CrystalClarity
10. JuggernautXL_juggernautX
11. JuggernautXL_v8Rundiffusion
12. LeosamsHelloworldXL_helloworldXL70
13. MOHAWK_v20
14. Nightvisionxl_V900
15. Omnium_v11
16. PhotopediaXL_45
17. RealismEngineSDXL_v30VAE
18. RealisticDigital_v40
19. RealisticStockPhoto_v20
20. RealvisxlV40_v40Bakedvae
21. sdXL_v10 (the base SDXL model, just for comparison)
22. SDXLFaetastic_v24
23. SdxlUnstableDiffusers_nihilmania
24. SdxlYamersRealistic5_v5Rundiffusion
25. STOIQOAfroditexl_XL31
26. STOIQONewrealityxl_XL40
27. Svsn001_v10
28. WildcardxXL_v4Rundiffusion
29. WildcardxXLFusion_fusionOG
30. Zavychromaxl_v90

First test – the settings

For each test I ran, I used the same parameters for all the XX checkpoints. So the only thing that was changed was the base model I used to generate the image.

For the first set, a closeup portrait of a girl, I used the following settings:

Positive prompt: “photographic portrait of a girl, 1girl, 25yo, short red hair, hazel eyes, wearing a white sweatshirt, closeup portrait, blurred background, bokeh, city streets, extreme details, hyper realistic, perfect skin, detailed skin, visible pores, 8k, photographic style, 80mm lens, analogic photographic camera”

Negative prompt: “cartoon, 3d, 2d, painting, unrealistic, low res, low details”

Image size: 1024 x 1024

Sampling method: DPM++ 2M – Karras

Steps: 40 – CFG scale: 6 – Seed: 1136332535

First set results

This first test shows that not all the checkpoints I used will return good results. First of all the basic Sdxl_v10, as expected, is really poor for photo-realistic images. It was expected since it’s a very generic model, good for everything, but it excels in nothing. Some other are probably too “colorful” or have too much saturation. Many have unrealistic skin (too perfect), the eyes are not realistic (this is why there are many LoRA’s for improving eyes’ details) or the hair have a “plastic look”.

But it’s mostly a matter of taste and needs. Maybe you are looking for a hyper-saturated image, with bright-amber eyes, then you have a couple of models you can use that will give you that style. Other models generated the image of a normal girl, not a photo-model with professional make-up on, with a more natural look, with little skin imperfections. These, to me, are the best to start with if you want to achieve the highest level of realism.

Personally, from this first test, these are the top-5 I would recommend:

1. IcbinpXL_v6
2. LeosamsHelloworldXL_helloworldXL70
3. PhotopediaXL_45
4. RealvisxlV40_v40Bakedvae
5. WildcardxXL_v4Rundiffusion

Second Test – The settings

The second set will still use a girl, but this time will be a half-body portrait with a “recognizable background“. I used the following settings:

Positive prompt: “half body photographic portrait of a girl, 1girl, 25yo, short red hair, hazel eyes, wearing a blue-jeans jacket and a white t-shirt, upper body portrait, blurred background, New York city streets, at night, Empire State building, car lights, detailed skin, visible pores, incredible lighting, masterpiece, hyper realistic, insane details”

Negative prompt: “cartoon, 3d, 2d, painting, unrealistic, low details, closeup portrait, deformity, low quality, low resolution, blurry, unclear, wrong, watermarked, noisy, soft, deformed, ugly, bad anatomy”

Image size: 768w x 1152h

Sampling method: DPM++ 2M – Karras

Steps: 40 – CFG scale: 6 – Seed: 3820469872

Second Test results

This time there are many intersting things that come out. First of all there is more consistency, many of the different-model generations are very similar to each other (a few are almost identical, look the last three!). Only a few used correctly the “Empire State Building” clearly in the background. A few are again too “irrealistic”. Overall the results are good, but this time you need to check the small details, like the eyes, or the mouth, that in a half-body portrait tend to be too low-detailed, deformed or just wrong.

This test, on my personal opinion, do not show real winners, all images need at least a little fine-tuning to achieve better realism. In this case there are many LoRA’s that could help to improve general details, skin and eyes details and better lighting. By the way, these are the top-5 I would recommend:

1. CopaxTimelessxlSDXL1_v122
2. EpicrealismXL_v8Kiss
3. MOHAWK_v20
4. RealisticStockPhoto_v20
5. RealvisxlV40_v40Bakedvae

Third Test – The settings

The third set will use a LoRA, the one I created (you can read about in a previous post). The portrait this time is at wider angle, to allow more background and less blurred. I used the following settings:

Positive prompt: “1girl, (half body portrait:1.5), KiaraLora, short red hair, green eyes, medium breasts, walking in New York city, white shirt, (bluejeans:1.3), at sunset, summer, blurred background, city lights, New York skyscrapers, professional photography, photo realistic, extreme realism, 80mm lens”

Negative prompt: “green pants, 2d, 3d, drawing, painting, illustration, cartoon, anime, manga, text, words, font, letters, low quality, oversaturated, undersaturated, low details, overexposed, underexposed, grayscale, monochrome, bad photography, bad art, blurry, malformed, mutilated, poorly lit, visible nipples”

The LoRA was set to 1.0 strength, probably I should have tried with a lower value to achive a better realism.

Image size: 888w x 1184h

Sampling method: DPM++ 2M – Karras

Steps: 50 – CFG scale: 6 – Seed: 1980978714

Third Test results

This test showed more differences among the images, but some details are repeated in some of them . Not all of them used the LoRA correctly. Kiara face is not how it should be. But again, this may be because my LoRA is not trained correctly and also because I set its strength to 1.0. A few, on my opinion, generated a head that looks too big for that body, probably it’s the LoRA.

The third test came out nice. New York city at sunset in the background has been correctly used in every image. In this case there are many LoRA’s that could help to improve general details, skin and eyes details and better lighting. By the way, these are the top-5 I would recommend:

1. CopaxTimelessxlSDXL1_v122
2. MOHAWK_v20
3. Omnium_v11
4. sdXL_v10
5. Zavychromaxl_v90

Fourth (and last) Test – The settings

The fourth set will face a different task, a black and white photography of an old fisherman. The portrait wants to emphasize the wrinkles and white beard of the old sailor, the texture of his wool sweater, the dramatic contrast of the dark clouds in the sky. With a recognizable element (Eiffel Tower, we are now in Paris!). I used the following settings:

Positive prompt: “a black and white photograph of an old fisherman, with a long white beard, wrinkles, wearing a wool cap and a wool sweater, on a sailng boat, on the river Seine, in Paris, Eiffel Tower, white sails, dark clouds in the sky, photo realistic, insane details, extreme realism”

Negative prompt: “color photograph, saturated colors, 2d, 3d, anime, toon, painting, drawing”

Image size: 1024 x 1024

Sampling method: DPM++ 2M – Karras

Steps: 60 – CFG scale: 6 – Seed: 3186687898

Fourth Test results

The last test was really good! All the images are excellent. And you can get many different “flavour” from each checkpoint. Some have a really nice contrast (I like the basic sdxl_v10 results a lot eve it is not among the most realistic ones!), others have a more natural look. Just a few are slightly “fake”, looking a little like a drawing more than a photo.

It’s hard to choose 5, but I would recommend the following:

1. RealvisxlV40_v40Bakedvae
2. MOHAWK_v20
3. Omnium_v11
4. ColossusProjectXLSFW_10bNeodemonFP16
5. CopaxTimelessxlSDXL1_v122

Conclusion

It’s not easy to reach a definitive winner. Mostly because it’s just a matter of taste. Then, also because it depends on what you want to achieve. In these 4 tests, the top three, reaching 3 nominations in the “Top 5” are the following:

1. RealvisxlV40_v40Bakedvae
2. MOHAWK_v20
3. CopaxTimelessxlSDXL1_v122

I think these three deserve further testing. I would also keep an eye on Omnium_v11.

Hope you found this post useful. Please, leave a comment if you like!