I made LiliPixel v1: High quality pixel art image generator DiT under a fully open license
retained text
Examples in the post. After seeing the top dataset on HF two weeks ago, I got inspired into training a DiT model from scratch, for the first time. (and also to avoid burnout as for the last few months I've been training the same LLM lol) **Story behind the model and it's makings:** This is technically the sixth varaint of the model, went through so iterations to get a model which makes respectable images. At first I trained a simple DiT on just one dataset for 3 epochs (the pixel art 1.4M cc0 dataset) with no text encoder (simple tokenizer) it was fine but the model could only generate 16x16 images, then I made a new model that also supported 32x32, but.. prompts were limited to keywords from the dataset so I added Voyage 4 nano embedding model as a text encoder, which improved prompt adherence and understanding by a lot, and afterwards I added a 64 pixels dataset and we landed here. The model was trained across 12 epochs (so \~30 million non unique images seen) at epoch 7 I altered the data mix to have more 64 x 64 images per epoch vs lower res, and continued that until epoch 12 where I landed on the final ready to be used model. **Tldr**; trained this model by myself over 126 hours on my rtx 4080 and released it under an open license. The model supports rgba outputs at 16x16 and 32x32 res and rgb only at 64x64 (no transparent background) Let me know what you think! If you have any kinds of questions, feel free to ask and I'll try to answer all of them! The model was made with love. **Link to download diffusers ready weights:** https://huggingface.co/Pinkstack/LiliPixel-V1-190M
document fetched 2026-11-20 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1y0hgyz/i_made_lilipixel_v1_high_quality_pixel_art_image