This is the initial release of the LoRA trained by @1041uuu (TOYOI Yuuta) pixel artworks.
⚠️ Pixel perfection is NOT guaranteed. This LoRA is incompatible with the Pixelate x4 VAE. For best results, I recommend generating at 512x512 or higher using base-v01 with 30 steps / 5 CFG and EulerER_SDE / Simple. However, please note that it still will not achieve flawless pixel perfection.
I attempted to clean up the messy or broken pixels using post-processing tools like PixelOE and ComfyUI-AI-Pixel-Art-Enhancer, but in my testing, it was impossible to fix them without losing the original aesthetic and atmosphere of the style. Therefore, for this v0.0.0 release, I have decided to publish it as-is, in its "pixel-imperfect" state.
Recommended Settings
Recommended Weight:
1.0Optional Positive Prompts:
jaggy lines, dithering, no lineart, limited paletteOptional Negative Prompts:
jpeg artifacts, isometric, letterboxed, pillarboxed, dated
Description
Update Notes: v0.1.0
Dataset Changes: Increased the training image upscale factor from x2 to x8.
Training Adjustments: Tweaked the configurations to be more prone to overfitting.
Settings:
AdamW8bit,unet_lr=1e-4,dim=32,alpha=32,LR Scheduler: cosine_with_restarts (100 steps, cycles=1),gradient_accumulation_steps=1.
What to expect in v0.1.0:
Unlike v0.0.0 and v0.0.1, generating at lower resolutions will now result in very low-pixel-count pixel art (giving a true micro-pixel aesthetic).
Because of the high-overfitting training settings, compatibility with other LoRAs has decreased, and you may encounter more style conflicts.
On the bright side, the generation results at resolutions of 1024 or higher have improved drastically—the difference is night and day!