Sign In

The Big May of 2023 comparison of ~25 realistic models

The Big May of 2023 comparison of ~25 realistic models

Hello!

I have been hinting from time to time that I am doing for myself (and I guess others :P) a comparison of realistic models (that were available as of May 2023) for the purpose of being BASE models for LyCORIS / LoRa usage.

I have used 77 of lycoris that were waiting for release, I used those 24 models and for each, I've used 20 prompts. Quick math gives us 77*24*20 = 36960

Yes, I have generated 37k images :) It took me around 2 days or so (resolution 512x704).

It was almost 500 images per LyCoris. I write this article at the end of June because it took quite a while to go through those images and for each decide if it was GOOD or BAD. If anyone is interested in how I was doing it. First I used a script to sort it per person so I could go through 500 at a time. I was using Total Commander with images view and I saw around 15-20 images on the screen at the same time. I was clicking on those that I like and when I finished with all 500 I just moved the good ones into a separate folder. Then I quickly looked at the remaining to see if there are some that I liked, sometimes there were some so I moved them as well. Rinse and repeat 77 times.

I really wanted to do this sooner but it turns out that it is a quite tedious process :)

The whole statistic is uploaded to rentry (json file): https://rentry.co/kdaoo but I will of course summarize my results!

This was not only a test of the models but also a test of the prompts, moving on my new samples will be using models and prompts that work best! I will however update the models to the latest versions since a lot has happened during those two months :)

So, pretty much the separation between BAD and GOOD lets me compute the quality percentage. I have to preface it that this is of course subjective selection. What I like and what you like could be vastly different, but in the BAD category usually are failures (anatomical failures, blurry or otherwise technically bad outputs) and also images where there is no likeness to the actual person.

We all know that not every generation will turn out decent, but we expect to have it more frequently. So if the model produces satisfactory content more often - it is a good model.

Out of the all GOOD and BAD outputs I'm computing a ratio and percentage. The higher the percentage - the better. I would say that anything close and above 40% constitutes a decent model. Sometimes there are models that produce epic results but not very often, they will be lower on the list because my main criteria were the amount of GOOD outputs. Such is life :)

So, here is the list of the models from Best to Worst (I will be linking to the main page if the model version is the same, or to the specific version that I've used, in case there is a newer one):

  1. https://civitai.com/models/4201/realistic-vision-v20 (71%) Realistic Vision

  2. https://civitai.com/models/5273?modelVersionId=27530 (52%) WyvernMix

  3. https://civitai.com/models/15003?modelVersionId=32492 (49%) CyberRealistic

  4. https://civitai.com/models/24385?modelVersionId=44105 (48%) MadVision

  5. https://civitai.com/models/28059?modelVersionId=70910 (47%) ICBINP

  6. https://civitai.com/models/41636/pornvision (46%) PornVision

  7. https://civitai.com/models/25694?modelVersionId=48174 (46%) EpicRealism

  8. https://civitai.com/models/21813/edge-of-realism (43%) Edge Of Realism

  9. https://civitai.com/models/47691/artiusv1-5 (43%) Artius

  10. https://civitai.com/models/4559 (42%) Stably Diffused

  11. https://civitai.com/models/7116/photomerge (41%) PhotoMerge

  12. https://civitai.com/models/28059?modelVersionId=42311 (41%) ICBINP (yes, another version)

  13. https://civitai.com/models/5083/photo-s4w3d0ff-blend (40%) [Photo] s4w3d0ff Blend

  14. https://civitai.com/models/23775?modelVersionId=59097 (40%) SuccubusMix

  15. https://civitai.com/models/40369/theallys-mix-iv-verisimilar (39%) TheAlly's Mix IV: Verisimilar

  16. https://civitai.com/models/6431/hardblend (37%) HARDblend

  17. https://civitai.com/models/49463?modelVersionId=54039 (36%) Am i Real

  18. https://civitai.com/models/44747?modelVersionId=49377 (34%) ProGenUberAnalogMix

  19. https://civitai.com/models/18637?modelVersionId=38189 (31%) PhotoSomnia

  20. https://civitai.com/models/30711?modelVersionId=37071 (30%) M4RV3LS & DUNGEONS

  21. https://civitai.com/models/12967?modelVersionId=58537 (30%) Unreality

  22. https://civitai.com/models/50899/huns (30%) Huns

  23. https://civitai.com/models/4922?modelVersionId=6514 (21%) GrapeLikeDreamFruit

  24. https://civitai.com/models/45424/metagod-real-realism (12%) Metagod Real Realism

Honestly, anything that is 30%+ is quite decent. I was initially blown away by Huns as it provided photorealism but in a more gritty way. Turns out that I only liked 30% of the generations but still, the things I got were very nice!

Overall it was a pleasure to see how my models blend with those checkpoints :)

One note about number 20: it is not really realistic, but I like the comic-book feel so it was a nice change to see them from time to time :)

As for prompts, overall this one is the winner:

portrait of sks woman|person by Flora Borsi, style by Flora Borsi, bold, bright colours, rainbow Mohawk haircut, ((Flora Borsi))

followed by: A stunning intricate full color portrait of (sks woman:1), wearing a black turtleneck, epic character composition, by ilya kuvshinov, alessio albi, nina masic, sharp focus, natural lighting, subsurface scattering, f2, 35mm, film grain

and: High detail RAW color Photo of gorgeous sks woman, smile, fit, toned, sexy look, high detail hair, (skin pores, skin imperfections), sharp body, highly detailed body, highly detailed face, sharp focus, 8k high definition, insanely detailed, intricate, masterpiece, highest quality,

All of them have their own negatives, you can find them in my samples.

So, what is next? I've already started doing samples for my upcoming models. This time I'm not doing 35k but a "mere" 9k spread over 120~ models so it is more manageable.

I am using the following list of models (based on the outcome of my previous test, but with updated version and some new contenders):

  "baseModels": [
    {
      "modelName": "absolutereality_v10",
      "modelPrompt": ""
    },
    {
      "modelName": "analogMadness_v50",
      "modelPrompt": ""
    },
    {
      "modelName": "cyberrealistic_v31",
      "modelPrompt": ""
    },
    {
      "modelName": "deliberate_v2",
      "modelPrompt": ""
    },
    {
      "modelName": "epicrealism_pureEvolutionV2",
      "modelPrompt": ""
    },
    {
      "modelName": "icbinpICantBelieveIts_final",
      "modelPrompt": ""
    },
    {
      "modelName": "madvision_v40",
      "modelPrompt": ""
    },
    {
      "modelName": "perfection_v30Pruned",
      "modelPrompt": ""
    },
    {
      "modelName": "realisticVisionV20_v20",
      "modelPrompt": ""
    },
    {
      "modelName": "reliberate_v10",
      "modelPrompt": ""
    },
    {
      "modelName": "wyvernmix_v9",
      "modelPrompt": ""
    }
  ],

If you are interested in seeing how those compare against each other, please leave a comment :)

Thanks for sticking to the end. Bye! :)

46

Comments