Introduction
The Illustrious Realism Experiment (IRE) is an attempt at a realistic Illustrious model that does not sacrifice its character knowledge, prompt adherence capabilities, and compatibility with Illustrious-based LoRAs. This was mostly accomplished by block merging WAI-illustrious-SDXL v17.0 with CyberRealistic CyberIllustrious v11.0 and a mix of SDXL models (see the credits and version history sections for additional information) using Hako-mikan's SuperMerger; in addition, a custom LoRA trained on a personal dataset of high quality photos has been merged into each version.
The latest version, 6.0, reflects major changes to the block merging process. The result is greatly improved compatibility with DPM samplers. However, IRE remains very experimental (hence the name). Hopefully, future versions will have further improvements to realism and visual fidelity.
Usage Guidelines
Summary:
VAE: already baked in
Prompting: use Booru tags (see below for important tags!)
Sampler: euler a (for v6.0, DPM++ 2M SDE seems to work well)
Scheduler: karras
Steps: ~30
CFG: 2-5
Resolution: 1024x1024, 896x1152, and 832x1216
For the best results, you should use Booru tags and place the following tags at the start of your prompt:
masterpiece, best quality, realistic, photorealisticIncluding the tags below in the negative prompt may also lead to better quality:
bad quality, worst quality, sketch, flat color, lowresThe preview images used the euler a sampler with the karras scheduler at 30 steps, but feel free to experiment with other options (the DPM++ 2M SDE sampler, for example, can produce sharper images). The recommended CFG scale is 2-5, with a lower CFG tending toward more realism, in my opinion. Good resolutions include 1024x1024, 896x1152, and 832x1216.
While the model should be compatible with most Illustrious-based LoRAs, results will vary depending on the LoRA, and you may need to lower the LoRA's strength and/or the CFG scale.
The model has the default SDXL VAE baked into it.
Credits
IRE is based on extensive block merging of various models:
Free free to create new merges using this model, but please do not use them for commercial purposes
Version History
v6.0: Improved visual fidelity and compatibility with DPM samplers by making substantial changes to the block merge process.
v5.0: Incorporated BeautyFoolXL2025 v2.0 and Reality BoundXL Axiom v14 into the merge.
v4.1: Improved LoRA compatibility by making some slight changes to the merge formula.
v4.0: Attempted to improve visual fidelity and realism by making significant changes to the block merging formula.
v3.0: Reduced graininess by integrating CyberRealistic CyberIlustrious v11.0 into the merging process.
v2.0: Made some improvements to visual fidelity by incorporating RealVisXL V5.0 and training on a higher quality dataset for the LoRA merged into the model.
v1.0: Initial version based on a block merge of WAI-illustrious-SDXL v17.0 with UltraEpicAi Realism v2.0.
Description
Some improvements to visual fidelity by incorporating RealVisXL V5.0 and training on a higher quality dataset for the LoRA merged into the model.
FAQ
Comments (6)
Love how grainy it is. Not sure if that's intended but my generations come out grainy!. Most illustrious models tend to look very similar and yours feels unique. Thank you for the work!
I've actually been trying to reduce the graininess, but without much luck. Not sure what's causing it!
@Maecenas keep it up, I also thing it's an issue which needs solving.
@TheodorSid v3.0 reduces the graininess quite a bit. The next version should be even better, hopefully!
@Maecenas I think you've got a DMD/CFG boost merged in there somewhere. Using APG I hit a wall way faster on this checkpoint than comparable illustrious finetunes were the unavoidable high CFG aliasing starts to kick in.
@codingnewb80604 To be honest, I'm not sure what a lot this terminology (DMD/CFG boost, APG, CFG aliasing) means in the context of this model. The model is based on a merge of the models listed in the description and incorporates (at very low strength) a style LoRA to improve visual fidelity.