V3 now released! <s>Electric Art - Version: Initialized</s> Electric Art AI Album Art Style LoRA (Low Rank Adaption) v2.0 Model is now being released for public evaluation. Trigger word: electricartNote: Be sure to save this page etc. I am right now going to begin to rebase the model to ensure it plays nicely with the rest of your LoRA. I am going to further optimize and make a v3.0 and with the release of that I would like to hold a contest for bragging rights or possibly more depending on participation amount! "Most Striking Image - Electric Art LoRA". All you have to do is try the model and upload your results to the model page here on civitai. I will select my personal winner. I would love the community to vote but I am not yet famous enough to have participation for that, let's change this?This is the first version of the model (<s>v2 now available, soon v3 with contest.</s> new and improved v3 now available) and I am hoping for feedback and discussion on how I can improve models I hope to continue creating in the future (as hardware allows, of course).The Electric Art - Version: Initialized LoRA AI album Art Style model was trained using 100 images with captions. The images are of 100 strikingly great electronic based music album cover artworks. The number is now less, to compare many images vs fewer images used during LoRA generation.I am a synthesist and music discovery aficionado. v2 includes new optimization of images used with weighting and pruning. V3 threw out many images to focus on the best ones much more strongly, adjusting the network rank and left to process for more epoch.I chose images based on my personal taste of both a great album artwork cover that will contribute to generating an interesting style once parsed through AI. I also chose basing initially on also loving the content of the music.When using the Electric Art Style model for example under ComfyUI please consider you may need to turn up the CFG value of the sampler first, to make sure you are generating tangible images. Then you can use my model turning the strength(s) down on it, then up until the effect is too strong, so just back off and it should be a sweet spot? Also adjusting CFG as well, to achieve a range of effect. Just be careful or the image will be too mangled, the value you can set the LoRA at depends on the CFG value input first is my experience overall. v3.0 seems to work at 0.6 but you should increase/decrease strength of LoRA as needed to avoid mangled mess or having no effect.Please just experiment and see what you can get out of it, be sure to let me know your experience. No commercial use as the images input to create the Electric Art - Version: Initialized LoRA were not licensed to me that way.

Electric Art LoRA

Electric Art LoRA v3.0: I used the SDXL base this time, so it should have better compatibility. I increased the Network Rank (Dimension) and Network Alpha to 72. This basically in theory should have made things "more intelligent". All of the "normal/rest" images were tossed out for now, I had used weighting but seems like recommended number of images is way lower either way. I kept the best images then also made a version cropping the best area of content on each image and weighting them nearly equal repeats. Using the SDXL model with CFG cranked to as high as 6.0 I could then set my model for instance at 0.6 strength in comfyui but try a range of values to reduce distortion of the image, while keeping the effect.I think I am next switching to flux if the 3080 and kohya are powerful enough.

Files with SHA256: 6dc0eb2f869eca3d0564c9c53022bf4c958dcb622cdfaed7a2c21ea4abba10a7