Project-K (Kawai): 5.0.1-250723: projectKKawai_501250723.safetensors

Minor changes to the color of the output image.

It is recommended to use it together with ControlNet.


  • The number of tokens that can be met is reduced.

  • Fine-tuned can lead to unintended problems. If you fine-tune it, set mix-precision to fp-16.


I don't know how to introduce it, but it's been renamed several times. Project-K is an open, free to use and fine-tune AI-art model. It was created by my curiosity. Hope you will like it. Have fun! (●'◡'●).


  • The `masterpiece` and `best quality` tags are not necessary, as it sometimes leads to contradictory results, but if it is distorted or discolored, add them now.

  • The CGF scale should be 7.5 and the step counts 28 for the best quality and best performance.

  • Use a sample photo for your idea. Interrogate DeepBooru and change the prompts to suit what you want.

  • You should use it as a supportive tool for creating works of art, and not rely on it completely.

  • Normal: Clip skip = 2.


As it is a version made only by myself and my small associates, the model will not be perfect and may differ from what people expect. Any contributions from everyone will be respected.

Want to support me? Thank you, please help me make it better. ❤️

Do you want to sponsor computing resources for us? Thank you ❤️. Please sponsor to me on Ko-fi at

FAQs that I made up (might be helpful for you).

Do you have a Discord server?

  • Well... No, I'm not prepared for that. Besides, my English is not very good.

Can I use your model to make money?

  • Of course, feel free to do so. But you have to promise me not to sell my original model. If I'm mentioned in your credits, I'll be happy.

Why is your model's size 5.19 GB instead of 7.14 GB? I have OCD. Ahhhhhhh

  • Well... it relates to the EMA training process. In order to maximize cost efficiency, we needed to train a model using fp16. However, when I trained the model using d8ahazard's dreambooth, enabling fp16 with EMA resulted in such a skewed model. Boom, we have fp16 EMA and fp32 UNet.

Your model seems... not very good?

  • I admit, compared to other models, mine is quite bad. That's why your opinions and feedback are important to me. Please help me improve.

Can you help me improve the model's output quality?

  • Based on my experience, you should use Easy Negative, set Clip skip to 2, and make the Negative prompt similar to the images in the preview. You can also try using the hires fix Lantent once; it works fine on my potato computer. Lastly, use Adobe Photoshop and a bit of luck.

Can you help me add the things I want to your model?

  • Of course, please comment below. I will do it for you for free to the best of my abilities.

Is there a secret?

  • Secrets are not revealed, but you may have already figured it out. Bro, we're quite similar =))).

Of course, there are more questions. Comment below, and I'll answer you or someone else will. Good luck.

Special Thanks: This wouldn't have happened if they hadn't made a breakthrough.

  • Runwayml: Base model.

  • CompVis: VAE Trainer.

  • stabilityai: stabilityai/sd-vae-ft-mse-original · Hugging Face

  • d8ahazard: Dreambooth.

  • Automatic1111: Web UI.

  • Mikubill: Where my ideas started.

  • Chat-GPT: Help me do crazy things I never thought I would.

  • Novel AI, Anything Model, Abyss Orange Model, CentusMix Model, Grapefruit Model: Dataset images. An AI made me thousands of pictures without worrying about copyright or dispute.

  • Danbooru: Help me write the correct tag.

  • My friend and others: Get quality images.

  • And You 🫵❤️


This license allows anyone to copy, and modify the model, but please follow the terms of the CreativeML Open RAIL-M. You can learn more about the CreativeML Open RAIL-M here.

If any part of the model does not comply with the terms of the CreativeML Open RAIL-M, the copyright and other rights of the model will still be valid.

All AI-generated images are yours, you can do whatever you want, but please obey the laws of your country. We will not be responsible for any problems you cause.

Don't forget me.

Have fun with your waifu! (●'◡'●)

I'm not sure if you've read this section, but if you're wondering, "Where are versions 1.0 and 2.0? What happened to them?" Well, don't fret, they've been retired because they were about as effective as a chocolate teapot. I wouldn't want you to waste your time downloading something that's about as useful as a screen door on a submarine.

This model has a bit of a past... it's kind of hard to describe, it's like a mix of comedy and melancholy. I usually call it the "awkward situations" mode. It was trained on an utterly illogical system, an RTX 3050, yes, you heard that right, an RTX 3050. Its speed is, well, let's just say it makes snails look fast. Training for one epoch takes a whopping 4 hours, OMG. And the results? Utterly disastrous. Sometimes, I question why I'm doing this when I could be downloading something better made by someone else. Yet here I am, still at it, up until now.

Simply put, it's all for the thrill, the sensation of creating something, even if it's hilariously terrible. It's like making a lopsided cake and still being proud of it. So, if you're like me, here's a piece of advice: "Where there's a will, there's a way," so keep on truckin' in your career.