A Comparative Analysis of Illustrious AI Realism Models
In recent months, Illustrious-based models have gained significant traction in AI art community. As Illustrious models become increasingly popular, creators and enthusiasts are asking: How do current Illustrious realism models perform when generating consistent imagery, how faithfully do they adhere to prompts and loras - and which one currently leads the pack?
This comparison examines four models using the Illustrious base model:
2DN by advokat (4.978 downloads)
Realism_Illustrious_By_Stable_Yogi (DMD2) (2.799 downloads)
Rillusm 2.0 by ZyloO (2.306 downloads)
Fasercore_V1_FP16 by Faser_Media (826 downloads)
Methodology
To ensure a fair and consistent evaluation, all models were tested under identical conditions:
Images generated at 3:4 aspect ratio (888x1184)
Identical prompt set across all models
CFG scales and schedulers are used as recommended by model creators (using mean values when cfg ranges were provided)
To test Lora-adherence, i used Ichi_sk's Oni Girl Momo Lora. I consistently applied a Lora to test Lora adherence capability. This Lora was only trained using anime / illustration type images so that any realism model can be checked on how well it adapts to non-realism loras.
"Best of Three" approach: selecting the best result from three random seed generations
Enhanced through 9 targeted inpainting passes: 3 for the face and 3 for each hand to refine details in these critical areas.
Prompt: I consistently used a prompt containing detailed clothing instructions, instructions on lighting, background setting and character posture. Models usually start to struggle with longer and specific prompts to some degree.
Positive Prompt: oni horns, cute oni girl, tanlines, round glasses, short black hair, black fur coat, red-white striped shirt, black thighhighs, white high-low skirt, light brown eyes, jumping, medium breastsize, forest at dawn, dynamic warm lighting, winter, light smile, <Stable_Yogis_Illustrious_Positives>
Negative Prompt: <Stable_Yogis_Illustrious_Negatives-neg>
The prompt included specific character details (oni horns, tanlines, round glasses, short black hair, specific clothing), setting elements (forest at dawn, dynamic warm lighting, winter), and posture instructions (jumping).
Performance Analysis
Let's first look at what results the process achieved!
2DN by advokat (4.978 downloads)
2DN showed some strengths in prompt adherence: You can clearly see tanlines, round glasses and short black hair. The winter theme as well as the warm lighting found it's way into the imagery as well. Unfortunately, even under a best of three approaches, 2DN did not provide a black fure coat and neither consistently a red-white striped shirt with a white high-low skirt either. A good portion of key prompts for character details were not consistently generated. The hands and face detail-passes achieved good results.
While the image has a really nice image aesthetic, it does not achieve high realism but drifted into high detail illustration.
Realism_Illustrious_By_Stable_Yogi (DMD2) (2.799 downloads)
Stable_yogi's Realism_Illustrious Model features a more amateur-photo look. It partially features tanlines and successfully generated round glasses, short black hair, a black fur coat, a red-white striped shirt as well as the white high-low skirt. Face and hand inpainting-passes worked well as well. However, as you can see, the image did have more trouble getting the lighting correctly done (we asked for warm lighting, forest at dawn). The overall imagery is much colder than expected by the prompt.
Nontheless, Stable_yogi's realism model features a high degree of realism, albeit their amateur-photo look may not be to everyone's aesthetic liking.
Rillusm 2.0 by ZyloO (2.306 downloads)
ZyloO's Rillusm model brings out the best of his pony experience and provides a great illustrious model: the model fulled respected all clothing prompts, portrays lighting instructions very well and has a great overall aesthetic. The model seems to still have some minor issues with eye details and the smoother texture combined with professional, hard lighting can subtract from a desired realism aesthetic.
Fasercore_V1_FP16 by Faser_Media (826 downloads)
Faser_Media is the least downloaded model of the four, but look at what Fasercore did! While it did not manage to implement the tanlines, it implemented all other prompt instructions to a great degree - the warm lighting can be seen throughout the imagery, the clothing details have all been respected with a stunning realism aesthetic. Fasercore shows a good distinction from both amateur-photo looks and high-detail illustrations by featuring a realistic but asthetically pleasing image style.
Conclusion
Your experiences may vary when using these models so i recommend you also try all of them if you have the time. While 2DN is currently the most popular model of the four by downloads, it also showed the weakest prompt adherence to multiple character details with the highest tendency to an illustration aesthetic. 2DN still has nice looking images, but it may not be the best choice for more complex prompts and those looking for the most realistic image generations using illustrious.
Stable_Yogi's Realism model shows off a high level of realism by catering to a more amateur-style photo look with great prompt adherence. This great model is only dampened a bit by the lighting in my testing. If you like the amateur-photo style, this is your best choice!
ZyloO's Rillusm has a great aesthetic and i'm excited to see how the King of ponyrealism will improve this model in the future. Rillusm seems to still have minor issues with eye details and has room for improvement on skindetail for high realism. Try it and watch out for model updates!
Fasercore was the underdog of this model comparison and i'm stunned by how good the result it. Fasercore's model page shows a lot of less-polished images but it's much greater than the download numbers or the preview images might let you think. It managed almost perfect prompt adherence, a great image asthetic and shares the top spot for realistic details together with stable_yogi's model. If you are looking to create highly realistic images using illustrious, do not miss out on Fasercore!
While the four tested models show clear differences in aesthetics, prompt adherence and image realism, illustrious models tend to have a much higher prompt adherence than pony models and handle character details, postures and clothing objects really well overall. If you mostly use pony and have not yet tried illustrious models, i hope you got an idea about how great some of them are at the moment.
I hope this short comparison analysis offers some perspective on what models to try in the ever growing landscape of image generation base models. See you around!