Any checkpoint can now be in the generator. Anything already loaded works for everyone; members load the new ones free for now, with a faster lane and Boost to jump the queue.

Any Model, Any Time: Load It Into the Generator Yourself

We rebuilt the trainer. Watch your run as it trains, generate with an epoch before it's done, and compare epochs side by side. It's in beta, and you switch it on yourself in your preferences.

Training Studio: Watch It Train, Generate As It Goes

A training and content contest with over 2 Million Buzz in prizes and 50% off H3 generation for the duration of the contest.

The Civitai H3 Contest: 2 Million Buzz Up For Grabs

I noticed a distinct lack of biomechanical and body horror style loras for SDXL, so I cooked up this abomination. Trained on a variety of cgi and AI generated images in an attempt to get it to work with a wide variety of styles. It can at the very least manage anime and photorealism to a certain degree. While "flesh" and "mutant" seem to be the most reliable trigger words, it seems to trigger off of a variety of horror related keywords so you'll have to play around with it. Also works to some degree with biomechanical subjects (the merged version is much better at generating cyborg and robotic characters). The merged version of the model should be used with lower weight ranges (0.35-0.6) because its a combination of multiple loras, some of which may be over trained.

Anima version of the model trained with a lot fewer AI images and more images created directly by me to improve the concept. The trigger words are still "cybermutant" and "fleshmutant" and as usual I recommend describing what you want in detail (Anima seems to respond especially well in this regard compared to prior versions). I recommend a strength of 0.7 to 0.9 as any higher and the model seems to tend toward a specific style, which I hope to fix in future versions.

Anima Beta

Improved version of the model trained locally on a reduced dataset focusing on a greater variety of styles in order to make the model not alter styles as much. Should be capable of the same concepts as prior versions. The trigger word for robots and robotized zombies is "cybermutant", while the trigger word for body horror eldritch abominations is "fleshmutant". Additional keywords related to the concepts you desire are helpful, such as "glowing green tubes" and "robot zombie" for cybermutant and "body horror" or "extra eyes" for flesgmutant. Check the example images for which tirgger words help with the concepts, or try your own.

Illustrious V1.5

Illustrious version trained on a slightly expanded data set, including nsfw, using civitai's training service. This version uses the same trigger words as prior versions and seems to be compatible with most styles and checkpoints based on noobai and illustrious. The trigger word for cybernetic monsters is "cybermutant" while the trigger word for body horror is "fleshmutant", though I recommend use additional trigger words to get more specific results, like some of the example ones ive included.

Illustrious V1

Flux Version of the model trained on simple tags. This version should allow maximum flexibility in terms of style, as other versions tended to alter the style of the image. The base keywords are "fleshmutant" for body horror creatures and "cybermutant" for cyborg zombies. I also trained some other keywords such as "covered in eyes" and "glowing green tubes", however it should be fairly flexible in terms of what you prompt it with. The additional trigger words provided are only examples, with enough prompting you could probably get different things. It should be capable of whatever the prior version of this model were.

FLUX v1.0

Flux version of the model trained on the same dataset as prior versions. Using the trigger word "fleshmutant" gets you body horror style monsters made of flesh, while "cybermutant" gives you robot zombie style monsters. After some testing, I can safely say that it actually can create some additional concepts like glowing green tubes on things and things covered in eyes. If you apply the keywords to landscapes and the like and add some extra descriptors, it kinda functions like a world morph, so you can create buildings made of flesh or robotic forests.

FLUX Beta v1.0

I wanted to experiment and see if it was possible to create a LORA that would function in both SDXL and PonyXL with similar behavior so I merged my 2 release models of each at 0.75 strength. I'm quite satisfied with the resulting model, though the SDXL portion seems stronger than the pony portion so I probably should have used a different merge ration. Its flexible and can handle most styles and subjects, though youll need to put the pony version at higher strengths to get more consistent results, while you can probably get away with lowering the strength for SDXL.

SDXL/Pony Merge V1

After pruning some images from the dataset and fixing some tags, I think ive gotten the pony version of the model to produce some decently flexible images over a wide variety of styles.

PonyXL Release

Tweaked the dataset of the SDXL version a little bit, hopefully making it a bit easier to generate body horror and cyberpunk concepts without overly long prompts while also maintaining a flexibility with styles and subjects. As the example images show, it works a lit better with upscalers than it does just generating images at a high resolution.

SDXL Release v1.1

So Ive finally gotten the model to a decent state where I can say it is no longer beta. I was able to get a good data set with adequate captions trained at 1024 resolution on civitai and this is the result. Its pretty consistent with creating horror subjects but also works in a wide variety of styles. Strengths closer to 0.5 seem to more stylistically flexible while strengths closer to 1.0 will produce more consistent results. Note that while the keywords I listed do trigger the 2 parts of the model, other horror and cyberpunk prompts may also trigger parts of it. For instance "glowing green tubes" adds a bunch of glowing green tubes to stuff, and "eldritch abomination" produces tentacled horrors. I would recommend playing around with it and seeing what you can get.

SDXL RELEASE v1.0

Trained on a slightly enhanced dataset in hopes that the model can produce a wider range of subjects, including animals. Seems to be mostly successful, but the model seems to work best in realistic style checkpoints than it does in 2D ones, though lowering the weight seems to help with style issues.

Flesh Monstrosity XL