真人换脸？用Midjourney Cref 参数去垫图 到底有多像？Cref+Sref参数会更有效吗？Cref是否适用于照片人脸迁移？最新Midjourney真人面部迁移功能测评

The video begins with a recap of the CREF (Character Reference) parameter for the MJ platform, which was previously demonstrated using a picture of Barack Obama. The host addresses viewer comments about the CREF parameter not producing images similar to the reference images. To address this, the host uses Asian faces as reference images to test the CREF parameter's effectiveness. The video establishes criteria to measure similarity using ComfyUI's facial data recognition nodes, which provide values (EUC and COS) to quantify similarity. The host then outlines the testing methodology, which includes submitting multiple images with the CREF parameter, redrawing faces after initial output, and combining CREF with SREF parameters. The video proceeds to demonstrate the process using a shared account for testing and provides a link for viewers interested in obtaining a shared account.

The host presents the results of using the CREF parameter with different numbers of reference images for facial similarity. The data shows that the similarity values (EUC and COS) do not necessarily decrease with more reference images, and sometimes, fewer references yield better results. The video discusses three groups of images: female, male, and children, noting that the male and children groups perform better than the female group in terms of facial similarity. The host also shares personal opinions on the visual appeal of the results and the limitations of the CREF parameter when used with simple prompts. The video concludes that the CREF parameter's performance in creating realistic Asian faces is not ideal, and there is no clear correlation between the number of reference images and the resulting similarity.

The video continues with further testing of the CREF parameter. The host explores the use of a hidden CW value in the CREF parameter, which when set to 0, focuses more on generating facial features, and combines this with local redraws to see if it improves the similarity. The results show that while the numerical values slightly improve with the CW parameter adjustment, the visual outcome is not necessarily better. The video also tests the combination of CREF and SREF parameters for image generation. The host concludes that the combination does not significantly improve the facial similarity compared to using CREF alone. The video wraps up by suggesting that manually iterating through images ('brushing through images') might be the most effective method to achieve a good facial resemblance with the CREF parameter. The host expresses their understanding that the CREF parameter is more effective with MJ-generated images rather than real-life photos and thanks viewers for their support.