PixArt-Sigma-1024px_512px-animetune

Pruned Model fp16 (1.15 GB): Inference checkpoint.Pruned Model bf16 (9.12 GB): Fine-tuning diffusers model+OneTrainer config data.Training Data (66.9 KB): ComfyUI workflow. ■Trained for 18 additional epochs, +210,000step■Training with the 400k dataset helped the model learn more concepts, but it also made the style more like average 2D illustrations, sometimes losing the aesthetic and anatomical stability of earlier versions. More precise prompts may now be needed. For easier, more aesthetic results, v02 or earlier (trained on 70k images) might be a better choice.Compared to the 512px model, the 1024px model offers more compositional freedom due to its higher resolution. However, the 512px model is much more stable overall. It's best to choose based on what you prioritize.I recommend using both models with SD1.5 or SDXL for img2img. It combines PixArt prompt flexibility and creativity with SD's strong styling. My sample images were made with SD1.5 i2i. I’ve also shared my workflow feel free to use it as a reference.I also recommend using the TIPO tag completion extension. It makes it easy to create long prompts and improves the stability of the generation results.●It is recommended to use as specific tags or natural language prompts as possible with this model.Short prompts like "1girl,beach" may not produce very good images.Tags that consistently have high-quality images on Danbooru might help increase image quality.●If you don’t want to waste time, it might be a good idea to use the 512px model first to practice which prompts are effective.●Also, the model has not been trained on quality tags or negative prompts.It has not been trained on images with potentially harmful effects, like sketche or monochrome images.However, all 400,000 images are of high quality, so there is a possibility that any tag could improve quality. The more tags, the better.●Natural language prompts can also be enjoyable.When I want to create beautiful images, I use natural language.When I want to use concepts that the base model doesn't know, I use tags.You can also combine the strengths of both by adding tags into natural language prompts when appropriate.●I’m using only general tags and character names. Tags like work titles, artist names, highres, official art, game CG, and other meta tags have been removed. Exceptionally, 'anime screenshot' has been individually trained. These tags were removed because they were commonly applied to many unrelated images and could overload the tokens, causing the model to learn unwanted concepts.■Breakdown of the dataset:Existing: 70,000Concept enhancement: 200,000Aesthetic enhancement: 130,000

pixartSigma1024px_1024pxV04_trainingData.zip

pixartSigma1024px_1024pxV04.safetensors

pixartSigma1024px_1024pxV04.zip

1024px_v04

512px_0.7

512px_0.6

512px_v0.5

512px_v0.4

512px_v0.3

512px_v0.2

512px_v0.1

1024px_v0.31

1024px_v0.2

1024px_v0.1

1024px_v0.0

512px_v0.0

Files with SHA256: a4634170c3f552f75f535542b055f8e68d928f3840f130999ec9582e239ae5d2