Upload 3.5 version. Using 'Add Difference' method to add some training content in 1.6 version Yesmix (original). Try to balance realistic and anime effects and make the female characters more beautiful and natural.
3.5 version now is available in tensor.art
Usage for 3.5 version
For child: Control the weight of
cuteto adjust age of characters.
For sister/mature female/おねえ: Control the weight of
mature femaleto adjust characteristic of characters.
For realistic/anime: Control the weight of
realisticto adjust effect.
To prevent loss of prior knowledge, the LECO method is not used in this version to erase negative tags.
Upload 3.0 version. The 3.0 version has the following important updates compared to previous versions:
Support for kl-f8-anime2 VAE and no longer support for NAI's VAE.
The 3.0 version is merged by a base model trained with about 7000 images randomly selected in a dataset that contains more than 10000 images. The base model will be continuously updated with the dataset updates.
Use LECO method to enhance the positive tags and weaken the negative tags in the model, so that good results can be obtained without inputting numerous negative tags.
The positive tags involved in training are:
masterpiece, best quality, extremely delicate and beautiful, highres, original
The negative tags involved in training are:
lowres, ugly, worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, blurry, bad feet, poorly drawn hands, poorly drawn face, mutation, deformed, extra fingers, extra limbs, missing limbs, extra legs, extra arms, malformed limbs, long neck, extra feet
The 3.0 version is currently well compatible with other anime LORAs. It can be used in conjunction with FastNegativeV2 to achieve the best results.
Upload 2.0 version. Building upon previous versions, I further balanced the generation effects of realistic and anime images, attempting to make both more natural and harmonious. As of now, I remain uncertain whether this version's generation results outperform those of previous versions. Currently, this version works quite well with other LoRA models (see preview images). FastNegativeV2 embedding is highly recommended, which can further enhance the clarity of generated images and make them look sharper. NAI's VAE had been baked in this version, so unless necessary, no need to load other VAEs. In addition, this model is not recommended as base model for training other LoRA models.
Upload original version of 1.6 Yesmix. No LoRA models are merged in this version.
Adjust merging proportion to fix hands. Merge following new LoRAs to enhance effect:
Moxin by simhuang to enhance face and style.
Fashion Girl by me to enhance face.
Akiryo's Mai by me to slightly finetune face, body shape and skin texture.
Thus, when this model and the above LoRA are used to generate images, the weight of these LoRAs should be lowered appropriately.
For convenience, the preview images are generated using only tags without any extra modules (e.g. embedding, LoRA, Hypernetwork) and functions (e.g. hires fix). The NAI's VAE had been baked in this model, so there is no need to load any additional VAEs.
Adjust the merging ratios to fix deformed limbs.
This model is originated an idea of creating a model that can precisely generate anime nsfw images. For this reason, I fine-tuned the NAI model with more than 10k nsfw images. Unfortunately, it's overcooked. To reduce this bad effect, I have to merge this model with the following models:
Hassanblend by sdhassan (0.15)
Anyhentai by asdpro123 (0.1)
AbyssOrangeMix2 - Hardcore by Havoc (0.15)
7th-anime-v3 by syaimu (0.2)
Kenshi by Luna (0.15)
Grapefruit by Ikena (0.1)
ColorbombMix by mocker (0.1)
Counterfeit v2.5 by rqdwdw (0.1)
Then this checkpoint was born after above merging, which is a very coincidental work. The name of this model is still unknown to me. Therefore it's named as "YesMix". The NAI's VAE had been baked in this model, so there is no need to load any additional VAEs. Thanks to the excellent models provided by the above authors. Btw, all of the preview images are generated using this model without any extra modules such as LORA and embedding.