【AI角色一致性】“cref” 参数 Midjourney新功能实操详解

The script introduces a new feature for AI character consistency, allowing for the stable generation of the same character. The narrator, who has been criticized for poor control, is excited to demonstrate this feature. The video will explain the 'Mijurney character reference' or 'CRIF' feature and its usage tips. The feature is similar to the 'style reference' function, but it focuses on character traits. To use it, after writing keywords, one should add '-cref' followed by a URL link, not the words themselves. There is also a new parameter '-CW' to adjust the weight of the reference URL, similar to the 'image weight' parameter but with a range of 100 to 0, defaulting to 100. A higher value means more detailed references, while a value of 0 focuses only on facial features. The feature works best with characters generated in 'Miji', but the narrator also tested it with other images with satisfactory results. The function is currently limited in precision and cannot perfectly replicate details like dimples or logos, but the narrator is optimistic about future improvements. The feature supports both 'Niji journey' and 'Miji journey' models and can be used with the 'Swift' function. It's important to use the MIGI6 model for this feature, as other versions are not supported.

The narrator proceeds to demonstrate the CRIF and CW weight features with a practical example using the character 'Uchiha Itachi' from the anime series 'Naruto'. The character is first generated using 'Niji journey', and several images are produced to select a suitable reference image. The narrator then tests the CRIF and CW weight features by attempting to change Itachi's outfit to a white shirt with different CW values: 0, 50, and 100. The results show that a CW value of 0 retains facial details well while changing the outfit, a value of 50 balances the new keywords with the original character details, and a value of 100 focuses heavily on the original image, resulting in minimal changes despite the new keywords. The narrator also tests the feature without specifying the character's name, only using the CRIF parameter to change the outfit to a pink shirt with a cigarette. The results indicate that even without specifying the character's name, the AI can generate consistent facial features through the reference image. The narrator emphasizes the importance of adjusting parameters correctly and concludes by paying tribute to the manga artist Akira Toriyama, creator of 'Dragon Ball', expressing gratitude for the childhood memories and wishing him rest in peace.