March 2

Improving the quality of your images with Loras. Part 1: Improve Image quality only by Checkpoints

Improving image quality is a topic of constant discussion and debate about which method is the best and most effective. To do this, I decided to make a series of lessons in which I will clearly show all the options available to you and what results they can give. After that, you can choose the best one for yourself or combine everything together. In the first part of a series of lessons, I want to talk about how you can improve image quality using only a change (checkpoint).

To create an image, we always need to first determine the control point of the picture (checkpoint), also called the style of the picture or, more simply, the model. The model used can become the basis for the creation, since it already contains elements such as clothing, environment, characters, drawing style, etc. The model can be used without additional parameters and settings. But each model also has a number of limitations that may interact poorly with different Loras.

For this lesson the following checkpoints were used:

Our lesson will be based on this prompt. It will not change throughout the lesson! This is an important aspect, because we want to see exactly how the quality and the image itself will change depending on changes in the checkpoin, and not on Lora or rewriting the prompt.

As a starting point, we define the following parameters:

Model (checkpoin):

Loras:

Prompt:

  • masterpiece, best quality, 1girl, gothic, black lips, solo, black hair, makeup, twin tails, hair buns, (Flannery), detailed hands, detailed face, detailed body, multicolored hair, bangs, eyeshadow, hair bow, two-tone hair, black background, front view, looking to the side

Link:

https://www.seaart.ai/explore/detail/cnecenle878c73dvk710?u_code=46RK7R

As a result of using these input data, we will get the following result. This image was obtained using a simple promt and minimal Lora settings. As you can see, images with these settings turn out to be of very good quality and with minimal deviations in the anatomy and individual parts of the character’s body.

https://www.seaart.ai/explore/detail/cnecenle878c73dvk710?u_code=46RK7R

Now that the basic input data has been established, we can move on to the practical part of our lesson, namely studying the influence of checkpoint on image quality.

Successful work

First of all, I want to look at those images that turned out well and do not have significant display problems. In this list I want to show that most of the existing Loras can be easily changed without fear of something going wrong. You can use the list given above (checkpoints, which I use most often in my work) or use others. But then the results may be different.

Anime checkpoins

The first in line of good works will be in the cartoon style.

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnecmede878c738lerig?u_code=46RK7R

As you can see, the image quality was not affected and in the end the image itself turned out to be very high quality. There are no problems with the anatomy or proportions of the character.

https://www.seaart.ai/explore/detail/cnecmede878c738lerig?u_code=46RK7R

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnecn1le878c73dvqmq0?u_code=46RK7R

The image has not lost so much in quality as it has changed in style, this is normal. The most important thing is that the proportions and the basic concept were not affected and remained at the desired level.

https://www.seaart.ai/explore/detail/cnecn1le878c73dvqmq0?u_code=46RK7R

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnecq5de878c73db7ajg?u_code=46RK7R

Another successful change to checkpoint while maintaining excellent image quality and level of detail.

https://www.seaart.ai/explore/detail/cnecq5de878c73db7ajg?u_code=46RK7R

Realistic-3D checkpoins

Now let's look at successful works in a realistic style.

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnecqote878c73dvtc10?u_code=46RK7R

Here we see a complete difference from previous works. This is normal, since checkpoint is designed specifically for a realistic style. But it’s important to note that Lora’s lists have an anime element. But as you can see, this did not lead to the end result being spoiled or any defects appearing in the drawing.

https://www.seaart.ai/explore/detail/cnecqote878c73dvtc10?u_code=46RK7R

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cneco8te878c73db5mgg?u_code=46RK7R

As you can see, the style of this drawing borders between 3D and 2D. At the same time, we can see small anatomical defects, but they are insignificant and with further modernization of the prompt we can get rid of them.

https://www.seaart.ai/explore/detail/cneco8te878c73db5mgg?u_code=46RK7R

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnechcte878c73db00ug?u_code=46RK7R

Another great option for modification is to use ReV Animated. This checkpoint is one of my favorites and frequently used. He always shows excellent results and precision in the execution of image details.

https://www.seaart.ai/explore/detail/cnechcte878c73db00ug?u_code=46RK7R

Unsuccessful works

Now let's move on to considering works that, due to various reasons, did not work out or the result of the work was not satisfactory for me. In this list, I want to show you that you need to be very careful when choosing checkpoins, because although most of them may suit you, some may not give the result that you expect. And this means that you don’t need to waste your time on them and you need to change the checkpoint to another.

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnecknte878c738lddl0?u_code=46RK7R

Working with such input parameters turned out to be more strange than beautiful. Everything in the drawing looks very strange, starting from the face and ending with individual parts of the character’s body.

https://www.seaart.ai/explore/detail/cnecknte878c738lddl0?u_code=46RK7R

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cneckvde878c738ldftg?u_code=46RK7R

Overall the image may appear normal, but it is not. Primarily due to the fact that the composition was changed. Part of the checkpoint does this and you need to come to terms with it and change the prompt to get the desired result.

https://www.seaart.ai/explore/detail/cneckvde878c738ldftg?u_code=46RK7R

To obtain this work the following was used:

Model (checkpoin):

Link:

https://www.seaart.ai/explore/detail/cnecpite878c73db6ppg?u_code=46RK7R

Overall the image looks ok, but the checkpoint style itself is quite specific to work with and requires more fine tuning to work with.

https://www.seaart.ai/explore/detail/cnecpite878c73db6ppg?u_code=46RK7R

The rest of the experiments showed little change with the other results. Therefore, I don’t see the need to show them separately. The most important thing is that in practice you saw possible changes in the image when changing only the checkpoint parameter without changing other parameters. The lesson turned out to be very long, and there will be 2 more parts in which we will look at two more possible options for changing image quality.

To summarize this lesson, the following points can be highlighted:

  • Choosing a checkpoint is a very important part of creating your image..
  • Changes to the checkpoint parameter can result in either small changes or major problems in your drawing. Be very careful with this.
  • Don't be afraid to experiment with checkpoints and their combinations, because this is the only way you can get better results.

If you want more tutorials like this or a better discussion of this topic, then write about it in the comments.

You can also see examples of how to use prompts in my telegram group, where I share ready-made algorithms and image creation tweaks every day.