This is 8bitdiffuser! A model trained on ~2k images
This page is for the 64x model!
PLEASE when using this model, use one of the AnyLoRA forks, Qteamix, Angelbreed or Animekawa. it works best with stylized anime / cell shaded models, while it can work on many other models, it might not produce the "pixel perfect" results
When generating images scaling them down 8x will produce pixel perfect results! you can downscale them via billinear or nearest neighbor, but i have had the BEST results when using astropulse's Pixelize plugin for A1111, it is completely free and can run even on lower end devices.
make sure to use an 8X downscale factor in post processing!
the 64x model, while a bit worse at generating even pixels, still operates greatly when downsampling images to native size, it also works much better at full body character portraits. with all v3 versions and above, high resolutions (768x and 1024x) are highly reccomended
Note for versions v4 and above! i used some filtering on the dataset, so if it gives you problems with NSFW content, use the v3 or earlier models.
Note that even doe this page specifies a 64x model, u can generate different resolutions of pixel art by just scaling the picture!
256x256 will generate a 32x32 pixel art
512x512 will generate a 64x64 pixel art
1024x1024 will generate a 128x128 pixel art
etc etc...
best settings for this LoRA is:
Steps: 10-30 (On lower step counts, pixels will look very noisy, however if you downsample this to native pixel size this issue is resolved, increase step count if you want a great generation without downsampling)
CFG scale: 4-10
LoRA scale: 0.85-1.25 (the lower you will set the scale, the more it will look like traditional art and the pixels will become uneven, but the higher the scale is it will appear much noisy)
Sampling method: DPM++ 2M SDE karras
Recommended prompt: "pixel_art" (Add these to the start every generation, this isn't "required" but I found it gets the best results)
Resolution: 512-1024 (on lower resolutions, the pixels will be less even, on higher ones the artifacts will be less noticeable but will take much more time)
Description
a LOT more training data on this version, also works much better with lower LoRA scale, this version was trained on 5000 images instead of 1000.
FAQ
Comments (7)
Amazing Lora, it gives me perfect results every time. Thank you! Are you going to make one for 128x too?
Thanks! regarding the 128x
Its very hard to train a 128x one (i have tried before, it works semi well but needs more training data and time then i have), but you can try generating images at 756x and 1024x, and then you can scale it down to 128x perfectly (aldo warnin, the LoRA can start to become unstable at larger sizes so ul need to prompt carefully)
a good example i can give is this generation i made: https://civitai.com/images/3361864 look at the prompting and settings if you need, its about 1024x756 which generates pixel art of 128x94
@extraordicord Yeah, generating a bigger canvas at the cost of stability of the prompt works.
I also tried generating 512x image without the Lora > upscaling it x2 > sending it to img2img > adding the Lora to the img2img prompt with denoise that fits best - and it works too https://civitai.com/images/3387841
Actually, you know, I just tried doing the same, but adding the Lora to the prompt of the Hires. fix step, without sending it to img2img and it works exactly the same, but quicker and more conveniently https://civitai.com/images/3388062
So, yeah, your Lora works wonderfully well as a post-effect of sorts too, really good job ❤
@daxivi6026 Hello! thought id reply to this thread, with the new 2.1 version of the lora out generating larger scaled images will now cause much less instability (aldo it is still present). and will also just generally work better with the style
sadly i dont think the instability is something i can solve but i think i got it to a level where u can generate about 5-10 images and get at least 1 good result every time.
@extraordicord Love it. It's gacha at times, but when it works, it works. I've made a comparison https://civitai.com/images/3548251 although it's cherry-picked, it definitely handles styles and perspective better. Haven't compared how it fares with highres fix, but I feel like it will work with a flat 2D style with hard lines and low details.
@daxivi6026 another (aldo, very late) update, with V3, high resolutions will work MUCH MUCH better! along with just the model operating much less gacha like. so less cherry picking is needed!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.