Rearrangement Notes

I've been trying to make a Lora-friendly merge model for a while now. However, there hasn't been a noticeable performance improvement compared to 1.2v. Personally, I think there's some degradation that happens during the merge process. This also applies to training models with AI-generated images. For the time being, it's expected that there won't be much progress in creating checkpoints, so I've arranged the 1.2v model, which is used by more people, in the front row. I think I'll be able to try a new checkpoint when I have more datasets. I'd also like to merge the merged models I'll create in the future with the checkpoints I created myself.

pony v.0.1 note

The dataset is still insufficient. However, it has reduced the overfitting problem to some extent compared to the previous version. And it is slightly more Lora-friendly, but still not Lora friendly. Since the dataset is small, there is a high probability that unwanted objects will appear when generating images. Please use negative prompts appropriately. I realized how difficult it is to train checkpoints. This model has many imperfections, but I think it has its own advantages. I hope you have fun working on it.

Trigger words: ceiiai ceiistyle

pony v.0.0 note

v0.0 This model is a later version than the previous model, but its version number is smaller. This is because there is a possibility that this model may not be placed on the front page in the future. While I was making the character Lora with images I created myself, I wondered what would happen if I made a checkpoint with these images. I don't know how to describe the result. First of all, it is not Lora friendly. I have always thought that being able to handle various Loras is one of the conditions for a good checkpoint. This model is stubborn about its own style. In particular, there are many cases where the style Lora does not harmonize with this model. Lora with a lot of detailed details is also likely to not harmonize with this model. This checkpoint actually showed better results when adding details to a simple model. I think the dataset of this model is quite biased. So the results are often unstable, and the negative prompts may need to be changed slightly depending on the case. Nevertheless, the reason I am releasing this model is simple.

It's just fun. :)

I don't know how I will use this model in the future. Many of you may be disappointed with this new version. I was also quite satisfied with the performance of Ceii Mix v1.2. However, when I started learning basic model training, I wanted to have a trained model instead of a merge model. Haha.

Now the title has changed.

Ceii Anime [PDXL] is now released.

If you use trigger words, the results will be closer to the style of the images I trained. I hardly used any other tools except Automatic1111 for the example images. I didn't apply Adetailer either. I trained the model based on the images created with ai, so there were some cases where the hand output was unstable and I modified it. I wrote other image posts besides the example images, but I didn't apply Lora in many cases.

Trigger words: ceiiai ceiistyle01

It seems that the dataset images contain a lot of wavy hair. If you don't want wavy hair, I recommend adding wavy hair to the negative prompt. In addition, I recommend using negative prompts such as realistic and source_3D depending on the case. You will probably find yourself adjusting negative prompts as you look at the results.

Have fun working on it :)

pony v.0.1b note

WARNING: This model(v0.1b) is highly unstable. I mostly used the same prompts I used for training when creating the example images, so I didn't notice the problem. Using prompts that weren't used for training can result in unstable and distorted images. I suspect that my model training destroyed a lot of the learned parts of the base model. I thought about deleting the model, but since some people have already downloaded it, I decided not to delete it and just post a warning. I recommend that you do not download this model to avoid wasting your time and resources. I will release a new version as soon as possible.

This was a model that I was hesitant to upload, but I'm sharing it because the results are interesting. It wasn't my intention, but a lot of the model's details and 3D effects were removed. I thought it might be overfitting, but I'm not sure because there are many parts where the image output is good. I'm learning that training checkpoints is very different from training Lora. I think the dataset I prepared is still lacking and there are many things to improve. I plan to continue working to create better checkpoints. Using trigger words generates results that are closer to the images used for training.

Trigger words: ceiiai ceiistyle

Old Version note

The purpose of this model is to generate 2D images by applying various Lora. Since I prefer a sense of volume, I did not want to completely remove the 3D effect of the model. Also, since I use a lot of Lora in my work, I needed a Lora-friendly model that is neither 2D nor 3D-biased.

pony v1.0 note

In order to maintain a slight sense of volume when applying this model with 2D Lora, the effect of the 3D model was intentionally left on when tags for animation were applied.

Therefore, when using this model alone, I personally do not prefer it. However, I think I got good results when I combined it with my personal favorite Lora.

This model was created through combination with the following models.

1. reweik_PonyXL v01.1 (by reweik)

https://civarchive.com/models/404802/reweikponyxl

2. 3x3mixXLtypeB (by ichigodaifuku) v01

https://civarchive.com/models/431350?modelVersionId=480549

3. DucHaiten-Pony-XL (no-score) v4.0 (by DucHaiten)

https://civarchive.com/models/376450/duchaiten-pony-xl-no-score?modelVersionId=531417

Embedding was not applied to the example image, but I prefer unaestheticXL_Alb2.

Follows the typical Pony score tag.

pony v1.1 note

This is an update for fine tuning.

Although I personally prefer strong contrast, previous versions tended to create a contrast effect by overly emphasizing bright areas. I attempted a merge to fix this, and it seems to have resulted in satisfactory fine tuning.

Because it was an attempt to adjust the differences in details, it was difficult to find a suitable model, but lizardon1024's post helped me a lot.

https://civarchive.com/articles/5300

Special thanks to lizardon1024.

This update was achieved by merging the existing model with the following model:

1. Rev Engine PonyXL

https://civarchive.com/models/502370/rev-engine-ponyxl?modelVersionId=558377

(by razzz)

2. Pilgrim_Ponyxl

https://civarchive.com/models/348373

(by devpilgrim)

pony v1.2 note

I prefer a drawing that is slightly different from the NAI style. However, due to personal needs, I needed to create a drawing that is a little closer to the NAI style. With this update, I created a model that is a little closer to the animation style without greatly damaging the details. I think it is still a Lora-friendly model that fits the purpose of applying 2D Lora while leaving the 3D effect.

In addition, in the previous v1.1 version, since the VAE was not baked with the model, those who tried to create the drawing by copying the prompt of my work had an issue where the work was output in grayscale. In this update, the grayscale issue was fixed by baking the VAE with the model.

This update was made by merging with the following models:

1. Homemade Pony (v1.0)

https://civarchive.com/models/600594/homemade-pony

(by NightCity)

2. WAI-ANI-NSFW-PONYXL (v6.0)

https://civarchive.com/models/404154?modelVersionId=627330

(by WAI0731)