2024.07.15
After a series of explorations and preparations for related possibilities in the past few days, the simplified version problem that friends need has now been solved. The latest version provided this time is the smallest simplified version. The actual content is still the same as the original version, but the speed is the fastest, and the visual expression of the picture is as good as all previous versions.
2024.07.13
This update is the third version of 1.0. After seeing a friend's message last time, I realized a problem, which is the size of the cropped version.
I have been busy looking through the archives and series of the past history, recalling the steps of the complete version of 1.0, and finally found a clue. This time I cut from scratch, but the final size is still the same as the last cut version, so I can now be sure that if the original model is large, even if it is cut, the reduction will be limited.
The content style is the same as the previous two versions, but the picture details have changed.
2024.07.11
I noticed that some friends wanted a smaller version, so I updated the 1.0 trimmed version. Although it doesn't look much smaller, the speed is slightly improved. The overall picture performance is consistent with the full version.
2024.07.09
This is the first model I tried. I have been using it for a while. My goal is to make the model easy to draw a satisfactory picture through simple and easy operations. After many tests, it has good performance in all aspects and is close to my expected state, but I feel that I can't bring out its full potential, at least not yet, so I decided to share it. Welcome everyone to use it and give me some valuable suggestions to help me improve it and make it better.
Features: It is very simple and friendly to get started, and does not require too many and complex instructions. No embeddings are required, and a few words can have a good effect. It will be rich at high resolution and also applicable at low resolution. It will be unexpectedly surprising when paired with various VAE and Lora.
Note: Generally speaking, when you start drawing, the default CFG is 7, This value may become sensitive, depending on the content. If you think the contrast is a bit heavy, try to reduce it to around 4, the color will be softer. The value should not be too high, it is recommended to keep it within 10. Even if all the parameters are the same, but the CFG value is different, it will have a great impact on the image.
I am used to using the HD repair option, ADetailer and ControlNet to control and adjust the picture effects I need. For more details, you can refer to the demonstration picture and the information data included in it.
Denoising: 2-5
Steps: 20-30
Sampling steps: 10
Sampler: DPM++ 2M Karras
CFG: 2-10
CLIP: 1-2
Amplifier: 4x-UltraSharp
Unfortunately, I am a novice, I don’t understand too many things, and I am still learning, but I will check the dynamics here regularly. If you have any questions, you can contact me and I will try my best to solve them.
The model prototype comes from many creators, and I salute them. Without their efforts, there would be no community today. My English is not good, and the above text comes from Google Translate, please forgive me.
Description
The size has been optimized at a deeper level, making it faster while also maintaining the expected outstanding performance.
FAQ
Comments (7)
So far im getting good result, will it be possible to have SDXL version anytime?
Thank you for your feedback. Of course, I am still focusing on version 1.5. There is no doubt that not only XL, but also Pony will be part of the plan. I want to make them a long-term series of projects, but it will take me a lot of time and energy. I am working step by step.
@yun447 i like most the traditional chinese ink art prompt from one of your shown images, sd1.5 creates alot of errors and is bad for alot of things but its fast(i wanted to create 1 horseman but instead it create 1 horseman with 7 horses fused together, really not good) for background is really good i will see what i have in generations to upload for you
@antonfawkes33350 Oh, the situation you mentioned happens from time to time. I have also encountered it many times. AI can be really outrageous sometimes, and make some very simple mistakes, but it cannot be corrected. It is stupid, right? It is also ridiculous. But I think this is an inevitable part, because matching the model with the recognition of text is usually a difficult and complicated process, especially when it is all handed over to the equipment and machines. Now the model does not understand the text well. It can be said that it will become very bad under some requirements, which is very frustrating, and it is even more difficult for sensitive multi-color. Through my long-term observation, I found that the prompt words in the generation should be as short and clear as possible. Too many and too messy words and sentences may cause it to crash because it cannot receive too much content. It must be limited. We can only select more concise prompt words in this small range to ensure that it has the best working condition. Yes, the traditional Chinese ink painting style has always been a type of painting that I appreciate and like. If nothing unexpected happens, I also plan to provide a similar lora at some point in the near future. Then, regarding the mistakes you mentioned about knights and horses, I suggest that you use ControlNet to achieve more accurate control of image generation, but before that you need to prepare a picture similar to a knight, or you can find a knight's lora for generation, which is what I often do. You know, this is to improve the accuracy of generation and some detailed control steps. Well, you can upload more pictures or prompt words for me for reference. If I can help you, I will be happy.
@yun447 this is the case for SD1.5, for SDXL based models or ponydiffuse mixes you can have long sentences(and it feels that the models follow your directions better), look at bananastrike lightning it can have a sentences that are followed (look at this image and its prompt https://civitai.com/images/20496656, i find it fascinating that they managed to make it work like that), i have been using adetailer to fix faces, hands, feet and similar, it does the fix automatically, once again thank you so much for this model, it has good results, you should make more and similar, in fact here, take my civitai lightning money
@antonfawkes33350 Thanks for you support, yes I'm also looking at the XL and Pony models to see what they are different from and what they have in common with 1.5, I'll try to see what I can find, but I'm also working on a lot of things that need to be done with 1.5.
fp16 version for speed boost?