The Conspiracy - LoRA (Delta Green Style)

You are cordially invited to a night at the opera.

The conspiracy began in 1927, following the raid on Innsmouth and the confirmation of what was previously local legends. The conspiracy would go by many names in the years hence, but has always been referred to internally by its security clearance its members claim: DELTA GREEN. Delta Green would be shuttered and revived throughout the decades, embroiled in conflict with both domestic rivals and cosmic threats, but the mission remains. To gather intelligence on and suppress unnatural incursions, and to protect the public from the corrosive knowledge of such unnatural forces. It is a cruel profession, but we live in a cruel and inhuman universe.

Scientia est mors. To know is to die.

Update 23/6/3:

Minor update optimizing the model size from 144 MB to 19 MB. Shouldn't change the behavior in any meaningful way, but it should take up a bit less space on your hard drive.

Update 23/5/29:

I fixed a lot of issues with the LORA for beta2. While the filesize went up, it retains the style across a wider variety of models and is overall more responsive and flexible. This is especially true for the noise offset, and you should be able to make both very bright and very dark images. I still need to do some more testing and dataset refinement before I publish the "final" version. Important note: make sure to use "disfigured" or "bloody face" in the negative prompt, as these elements can emerge unprompted due to the nature of the dataset. It is something that I will work on to alleviate for the final version. I still recommend 0.7-1.0 for weights, but the model can be touchy when it comes to lighting at higher weights.

Original Post
Hey folks,

This is a follow-up to my old TI version of the Denis Detweiler "Delta Green" style (The Conspiracy - TI). It is based on the art for the Delta Green tabletop role-playing franchise, which transposes the Cthulhu mythos to a modern day conspiracy thriller setting. I was happy with the TI, but it didn't really work outside of SD 1.5, so I trained a larger data set (with aspect ratio bucketing). The LORA works better with more models, although I find that strongly "aesthetic" models tend to wash out the oil painting style that Detweiler's artwork is known for. I recommend using it with either more neutral models that were directly trained on 1.5 (like Vodka) or models with a more painterly baseline style (like Dreamshaper or A-Zorya RPG Artist Tools). The LORA has not been tested with any CLIP SKIP 2 models, so no guarantees that it works with those.

23/5/28 EDIT: I did some more testing, and found that Realistic Vision 2.0 is actually much better at retaining the original style than the more painterly models, so you might have good results with applying photoreal-aimed models and merges as well.

I've also found this negative TI to be helpful (Negative Embedding for Deliberate) for the RPG Artist Tools model, and I've used Lykon's personal negative embedding (Fast Negative Embedding) for Dreamshaper.

This initial release is an open alpha release. For weights, I recommend 0.7-1.0, although I'm still testing to find the best prompts. You can use the ddstyle tag, but it doesn't seem necessary and doesn't have much impact. This is my first LORA, and I think that I can improve the dataset quality & hyperparameters. I also want to try training the LORA on other models to see if it can better transfer the style while gaining the improved composition quality of those mixes. Also, right now the model is just trained on the style, and I can work on improving tagging/captioning for better subject knowledge as well.

If you've read this far down, I appreciate it, and I've to hear your feedback.

Thanks,
Machinique