2024/6/24 update

Add approximately 10,000 images to the original dataset after personal preference screening. Fix the issue of low saturation in some generated images caused by the merge of AnimagineXL in version 2.0 (details can be seen in the last preview image).

In the dataset for the 2.5 version training, all images tagged as questionable and explicit have been tagged with nsfw. It is recommended to add this tag when generating NSFW images to enhance the effect, such as 1girl, huge breasts, explicit/questionable, nsfw

2.5 version recommended syntax:

major content e.g. 1girl/1boy, key feature tags, rating tags e.g. general/safe/explicit, quality tags e.g. score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up

2024/6/8 update

The training workflow upgrade is complete and 2.0 version have been uploaded. This version is based on PonyXL and obtained by training on a real-time updated dataset and merging with Animagine XL. The total training duration is approximately 60 hours. The issue with directly generating high-resolution images without hires-fix in the original Pony base model, e.g., extra heads and breasts, has been essentially fixed, but it is still recommended not to use extreme aspect ratios for image generation, with a recommended aspect ratio of less than 1:2 or 2:1.

Compared to 1.0 version, the realistic portraits now more closely resemble individuals of mixed Asian and European descent. If you find the appearance unappealing, you may appropriately reduce the weight of realistic and mature tag.

2.0 version was trained without making any changes to the original quality tags of the Animagine XL and PonyXL. Therefore, it is recommended to set the quality tags as follows:

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, masterpiece, newest, absurdres, very aesthetic

2.0 version has not yet undergone compatibility testing for the LoRA/LyCoRIS based on Animagine XL and PonyXL. Welcome everyone to try it actively and provide feedback on any issues.

The sampler is recommended to use Euler a. Although the generation effect of DPM++ is not unattractive, its stability is slightly inferior compared to Euler a. The CLIP SKIP can be set to 1 or 2, with no significant impact on the results.

Introduction

This is the SDXL version for the Yesmix. This model is based on animagineXL, which is fine-tuned on more than 22,000 images filtered using anime aesthetic predictor and produced by merging AAM XL and AingDiffusion XL. Therefore, this model also follows license Fair AI Public License 1.0-SD. This model can currently generate anime SFW/NSFW images with good quality and stability. Since the generated results of the SDXL model are more diverse than those of SD1.5, welcome to test it with different tags. Currently, at least 6GB of video memory is required to generate images using SDXL model in webui (although the speed is relatively slow).

This model now is available in tensor.art.

If you would like to support me, you can buy me a cup of coffee🍵.

Generating Settings

• Sampler: DPM++ samplers, Euler a are recommended.

• Eta noise delta: 31337 or 0

• Upscale Method: HiresFix+4xAnimeSharp

• Max Resolution without Upscaling: 1024x1024

• VAE: No need to use external VAEs.

• Embeddings: negativeXL is recommended.

• Clip Skip: 1 (SDXL model does not support clip skip).