Can't tell the difference between MJ versions?

Can't tell the difference between MJ versions?

What does MJ edition mean?

MJ will regularly release new algorithm models, and different versions of the models are suitable for different types of images. At present, the latest version is V5 released on March 16, 2023, but the default version is V4.

How to switch the currently used version?

Method 1: Enter/set the instruction, and select the version in the setting. The default is V4 version.

Method 2: Add [-]+model version after the prompt content, such as-v 5.

Can you understand what OV5 has updated?

The range of picture stylization is increased, and the response to prompt words is more accurate.

Improve the picture quality, optimize the dynamic range (rich picture detail range), add more picture details, display the content more accurately, and delete unnecessary words.

And the information weight of the picture prompt words is improved.

Support seamless tiling effect of [-tile]

The aspect ratio supporting [-AR] is greater than 2: 1.

Support [-iw] to adjust the information weight of prompt picture and prompt text.

Prompt words pay more attention to colloquial description.

In V5 version, prompt words are more relevant to the picture content, which requires us to pay more attention to the description of the picture content. In addition, the algorithm optimizes the function of natural language processing (NLP), reminding us to use a complete paragraph to describe, rather than a single vocabulary.

The default resolution of the generated picture is higher.

The default four-grid map in V5 version is upgraded to the resolution of 1024x 1024, while the previous four-grid map in V4 version is 5 12x5 12, which can only be achieved by upgrading with [U button].

The following figure shows the image sizes of different versions after frequency upgrading (only for the default square of 1: 1), and the red data indicates the size of the current version after frequency upgrading, that is, the size of the large image generated after clicking the u button.

At present, the default image resolution of V5 version 1024x 1024 is already the maximum resolution, so the u key only cuts the image without optimizing the details of the image.

Detailed frequency up-conversion and maximum frequency up-conversion operations are aimed at the old version of V 1 to V3, and other subsequent new versions add relevant operations by default.

Use more stylized parameters.

Previously, in V4 version, if animation and other tips were added, the style of the generated pictures was basically the same. However, in V5 version, the default stylization parameter is 1000, and the generated pictures will be more realistic, and the effect will be more like photos than art paintings. Therefore, in V5, adding stylized parameters to the prompts as much as possible will make the generated pictures more in line with the style we described.

√ iw parameter regression, pad chart is more useful.

Image weight parameters are regressed in V5, which was previously only supported by V3 version. This parameter is mainly used to set the weight of image prompts relative to text prompts. The default value in V5 is 1, and the numerical range is 0.5~2.

Remix mixed mode is more powerful.

The mixed mode is to continue to compose the generated picture and then regenerate the picture. The content, parameters, version and model, and picture proportion of the new picture can be modified. In the mixed mode of V4 version, even if the prompts are exactly the same, the algorithm will generate a new understanding, so the generated pictures will be very different. In V5, if the prompt words are not modified, the newly generated pictures will not be much different.

Repair the disorder of hands and feet

The confusion of hands and feet, which has been criticized in V4, has been optimized in V5 (although there are still occasional dislocations).

Tiling parameters supporting seamless tiling effect

Tile used to be only applicable to old versions of V 1 to V3. This parameter is used to generate repeated collage images, such as textiles, wallpaper textures and other patterns.

Image generation is twice as fast?

On March 30th, MJ officially announced that it has updated the V5 version algorithm (currently considered as V5 version), which will not consume additional GPU resources. In theory, the image generation speed is twice as fast, and the image quality will not be greatly affected. But recently, the server is full, which leads to frequent delays. At present, there is no difference in speed when using it.

V4 was released in February last year (12), and it is also the default version of MJ at present. Although V5 has improved in detail and performance, it is not stable. According to the current experience, it is still recommended to choose according to your own usage scenarios.

V5 is suitable for generating pictures that need high definition and detail reduction, such as photos, and V4 is suitable for generating illustrations or traditional images with strong artistic style.

Before the release of V5, V4 has been constantly optimizing the algorithm, among which the obvious changes are divided into three sub-versions: V4a, V4b and V4c, and the latest V4c is the default. You can add style parameters after V4 to switch and check the image generation effect of the sub-version.

Test /testp test mode

Test/testp is not an official product. It is used to test and collect feedback in public communities, in which TEST focuses on digital art and testp focuses on photographic images. The compatibility of test model with hints and parameters is poor, so more creative pictures can be generated by adding creative parameters.

Niji model is a two-dimensional picture model made by MJ and Spellbrush, and it is good at all kinds of animation styles in algorithm, especially in dynamic performance, action scene and character performance.

Spellbrush's team focuses on the field of secondary AI, and those who are interested can follow them disharmoniously.