<h3 id="new!">NEW!</h3><h3 id="dopamine-lite"> DOPAMINE lite</h3>BE CAREFUL NOT TO USE THIS AT 1, I WOULD SUGGEST STARTING AT 0.6 AND GOING UP FROM THERE. it can create horrors at high levels. Its the same as dopamine but has been polished with the same dataset with some more training. It has terrible sameface (it decided to merge all the faces into 1 face out of the many many in the dataset). But it has really insane style. Just replace the face if you hate it.<h3 id="dopamine">DOPAMINE</h3>Finally a Krea2 epoch fit for Weird Science. Do not go above 0.8 (or 0.65 on krea2 raw) unless you like nightmares. Settings in the pics. It can do 'normal' photos and absolutely batsh*t insane photos. You could even use this as a 'zoom in' lora with how it behaves. It tends to zoom in and try to fit everything in the prompt and often in creative ways (sometimes looks 'collage-y' but mostly normal if the prompt is normal). It also tends to make everyone's eyes bright too for some reason, which I don't like, but we accept what the noise gods give us and adapt. Decent to mid seed variety. Sameface problems as in all of my loras, as I used a rotating mix of about 15 people in my photography with the training photos. Mix up your prompts/use a seed/noise enhancer to move seeds.This could also be used for conceptual photography, in fact its really good at that but im posting it here in weird science for how weird it is.I've made some comfy images to load in comfy, for those who use comfy, for 12gb vram and 16gb vram and 24gb vram.This one tends to flatten things out and do creative compositions even in tight areas. For instance if you put 'person playing guitar' it will zoom in on the hand strumming the guitar but if you put 'person playing guitar, airplane overhead' it will zoom out a little, or even get an airplane in there while keeping a really tight zoomed in composition. Anyways ive loved playing with this one, but its very WEIRD -- DO NOT GO ABOVE 0.8.If you are in forge, this definitely helps avoid mutations and make the lora come through better (though takes a lot longer)<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/1c50b7c8-80de-47e4-92cd-9480d88587b4/original=true/1c50b7c8-80de-47e4-92cd-9480d88587b4.jpeg" />It also works great with the res2lyf samplers, if youre in forge neo you can add them with this <a target="_blank" rel="ugc" href="https://github.com/MisterChief95/sd-forge-extra-samplers">https://github.com/MisterChief95/sd-forge-extra-samplers</a>The dim is big, and like the first krea2 training the eyes/face can get messed up at full strength (it was trained at 608px, slightly higher than the conceptual photography training, but still catches a certain aspect of my style). Trained on real, PG-Rated photography. No ai in the training images.You may not like the style, just look at the images to see if you do or not :). The examples are all mostly full strength (0.8) but you can add the lora to another to add texture or switch up composition/seed variety. I've posted comparisons of with and without on the page. Its got the same sameface problems as the others so you'll have to substitute your own :). Its strength is mixing up composition and adding texture imo.Dopamine also messes up eyes a lot for some reason sadly, but i use the excellent adetailer <a target="_blank" rel="ugc" href="https://civarchive.com/models/330727/full-eyes-detection-adetailer">https://civarchive.com/models/330727/full-eyes-detection-adetailer</a> , which works in forge.Try WIDE and TALL, some compositions really shine in either or both. Also square works too, but the training images were all traditional 4:3 / 3:4 images for the most part.At high strengths it will do some things a lot better and a lot of things worse. You'll just have to experiment :).For forge users: Many of these images use Sigma Rescale Standalone by lz-55, based on RES4LYF’s rescaling tool.Download the <a target="_blank" rel="ugc" href="https://github.com/user-attachments/files/31612204/sd-forge-sigma-rescale-standalone-v1.0.0.zip">original extension ZIP</a>. Extract the extension folder from the ZIP into your Forge extensions folder, then restart Forge.Once installed, load the original PNG with its metadata intact into PNG Info → Send to txt2img. This restores Sigma Rescale’s enable toggle and Start/End values. My settings vary between images, so use the recorded values rather than the extension’s defaults.Make sure to match the remaining generation settings, including models, LoRAs, sampler, scheduler, steps, and seed. Images that use Hybrid Scheduler also require that separate extension, which is not included in this download.<h3 id="blowout"> BLOWOUT</h3>(USE 5+ distilled CFG at least!)This is a SUPER BRIGHT epoch (trained with Flux dev). It is a bit too bright on regular Flux dev, which is why I didn't post it---but i've been playing with it on <a target="_blank" rel="ugc" href="https://civarchive.com/models/1827475/flux-1-krea-dev-fp8">Flux Krea</a> with better results. Flux dev loras have to be REALLY bold to have an effect on Krea--and this one is in fact very bright/bold.Trained on concept photography and light trails photography.Run at lower strength to brighten/add detail to photography prompts. Has sameface, so may not mix well with character loras. See example images.Faces may be muddled especially if subject is a small part of the frame--just use adetailer or inpainting to fix.This is a style lora and trained on a wide variety of images and is not intended to reproduce any likeness.<h3 id="notes:">NOTES:</h3>This page is now my LIGHT TRAILS series :). ALL OF THESE ARE TRAINED ON REAL PHOTOGRAPHY--a subset of images I used for <a target="_blank" rel="ugc" href="https://civarchive.com/models/1532569/conceptual-photography">Conceptual Photography</a>. For both of these series I've gone through about 65 trainings and thousands of epochs testing my favorites over the last 2 months and I will slowly release my favorites, I have it narrowed down to about 10 or so--i'm such a digital hoarder it's difficult for me to choose which to post, but i'll start with my very favorites and go from there...who knows...i may just be getting started or this may be the end...stay tuned!I have chosen my favorites based on SEED VARIETY and CREATIVITY and also ability to overwhelm Flux biases. One thing I hate about flux is running a prompt and they all look the same. I'm trying to pick epochs where that's not the case.Run at higher strength at your own risk, or use the recommended.I have presented some example images in the post for each lora as UNEDITED AND NON-CHERRYPICKED GRIDS to show seed variety/creativity. Each pic in the grid uses the same exact prompt.These are also trained at a high dim in an attempt to override flux's biases, so the file size is large. Someone more skilled and experienced could probably figure out a way to do this without having such large files, but thats my experience level at the moment. Tested on my 3090 locally with Forge UI with Euler/Simple and most around 5(distilled)CFG, on FLUX DEV FP8.Settings for each are in the metadata of the images. See image gallery for examples of what these can do.

Weird Science

BE CAREFUL NOT TO USE THIS AT 1, I WOULD SUGGEST STARTING AT 0.6 AND GOING UP FROM THERE. it can create horrors at high levels. Its the same as dopamine but has been polished with the same dataset with some more training. It has terrible sameface (it decided to merge all the faces into 1 face out of the many many in the dataset). But it has really insane style. Just replace the face if you hate it.

Files with SHA256: 0415bb1f087997277bbb8b3fdc42a96128525d53863c6cedb270f14bfec4178f