PixArt-Sigma-1024px_512px-animetune

Pruned Model fp16 (1.15 GB):inference Model .Full Model fp32 (2.29 GB): inference Model fullsize.Pruned Model bf16 (9.1 GB):Diffusers model for fine tuning + onetrainer config data.Training Data (9.5 KB):comfyui workflow.12/4 : I've added workflow_v02. It allows adding style tags separately from the base prompt, achieving good quality even with short prompts. It’s motivating and more enjoyable—I really like it!● I trained with a dataset of 400,000 images.The 53-epoch training is complete.Excluding details and fine concepts, I think the 512px model performs better in terms of anatomy and style.The 1024px model might ultimately provide superior inference, but reaching that point involves many failures and takes too much time. Continuously generating 1024px images without knowing the outcome is frustrating. This model is designed to support that process.This model has several potential uses:● Since it has fast inference speed and tag compatibility, it can be used for prompt testing before running inference on a 1024px model.● Use the 512px model to create a good composition, then upscale with the 1024px model.● Merge the 512px model with the 1024px model to generate around 768px, balancing speed and detail. (Practicality is uncertain, as it may not work reliably.)I’ve also uploaded a few workflows for reference.The standard size for this model is 512pxA ratio like 512x768 like SD1.5 is suitable.However, with a long side of 768px, it slightly breaks down. If stability is important, it’s better to base it on 512x512px and adjust the aspect ratio, such as 448x576.768px 1024px is not trained, so the result will be disastrous.The base model is very high quality even at 512px!Usually, models in the middle of pre-training or lite versions lack sufficient learning or aesthetic appeal, but this base model is different. It is the most aesthetically pleasing I have seen so far.Honestly, for creating images focused on 2D characters, there’s little difference between 512px and 1024px. Unless it’s a concept that clearly requires high resolution, 512px should be sufficient.Some tags may still be unresponsive with the 512px model but I believe this will be resolved as training progresses.●If you can't come up with a prompt, try using the prompt auto-generation below.<a target="_blank" rel="ugc" href="https://huggingface.co/spaces/KBlueLeaf/TIPO-DEMO">https://huggingface.co/spaces/KBlueLeaf/TIPO-DEMO</a>●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.●There are likely tags that can be used as negative prompts.Tags related to styles, like the ones listed below, can be included in negative prompts to switch to a different style.1980s (style)-1990s (style),00s,10s,simple background,anime screencap, realistic,figure etc...●If you find it troublesome to come up with prompts that produce stable quality, using prompts like the ones below might help stabilize the output. Ironically, tags like these end up becoming quality tags. lol" nikke, azur lane, blue archive, kancolle, virtual youtuber, arknights, girls' frontline"

pixartSigma1024px_512pxV03.safetensors

pixartSigma1024px_512pxV03_trainingData.zip

pixartSigma1024px_512pxV03.zip

512px_v0.3

1024px_v04

512px_0.7

512px_0.6

512px_v0.5

512px_v0.4

512px_v0.2

512px_v0.1

1024px_v0.31

1024px_v0.2

1024px_v0.1

1024px_v0.0

512px_v0.0

Files with SHA256: 01abbf8b21199a65832730e6f20ec23420f25965eafca0d50158410f71ea2dc1