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
TLDRThe video explores the effectiveness of the CREF (Character Reference) parameter in the MJ image generation tool for creating realistic facial depictions, particularly of Asian individuals. The creator tests three methods: using multiple reference images, combining CREF with局部重绘 (partial redraw), and using CREF alongside the SREF (Style Reference) parameter. The results indicate that CREF's performance in mimicking real Asian faces is subpar, with no images meeting the set similarity standards. The video concludes that, currently, the best method to achieve a realistic facial depiction using CREF may be through extensive image iteration rather than the tested methods.
Takeaways
- 🎨 The CREF (Character Reference) parameter is used for facial feature reference in image generation, aiming to make generated images resemble a provided reference.
- 🌐 The effectiveness of CREF was tested using images of Asian individuals, as the previous example with Obama was deemed too distinctive.
- 📊 ComfyUI's facial data recognition nodes were used to calculate similarity scores, with lower EUC and COS values indicating greater resemblance.
- 🔄 Experiments were conducted with varying numbers of reference images (1, 2, and 3) to see if more references improve resemblance.
- 🖌️局部重绘 (局部重绘) was attempted after initial image generation to refine facial features, using the CW value to focus on facial details.
- 🌟 Combining CREF with SREF (Style Reference) parameters was tested to see if it would enhance the resemblance of generated images.
- 🔢 The results showed no clear correlation between the number of reference images and the similarity of the generated images to the originals.
- 👤 CREF's performance in facial feature migration for Asian individuals in a realistic style was found to be suboptimal.
- 🚫 None of the images generated with CREF met the high similarity standards set by the creator (EUC < 0.8, COS < 0.3).
- 💡 The creator suggests that manually 'brushing' or re-generating images might be a more effective way to achieve desired results with CREF.
- 📚 The video aims to provide insights into the use of CREF for facial feature migration and offers tutorials on using MJ's (Midjourney's) features.
Q & A
What is the purpose of the CREF parameter discussed in the video?
-The CREF (Character Reference) parameter is used to improve the similarity of generated images to a reference image by adjusting the facial features and style based on the provided example images.
Why did the creator choose to use Asian faces for the demonstration in this video?
-The creator chose to use Asian faces to test the effectiveness of the CREF parameter in generating images with a more diverse range of facial features, as the previous example with Obama was quite distinctive and may not represent the performance across different facial structures.
How does the video determine the success of the CREF parameter in achieving a similar look to the reference image?
-The video uses a tool called ComfyUI to calculate the facial similarity between the generated images and the reference images by comparing EUC (Euclidean) and COS (Cosine) values, which provide a numerical representation of similarity.
What are the thresholds for EUC and COS values to consider the generated images as similar to the reference image?
-The video suggests that EUC values below 0.8 and COS values below 0.3 are considered as indicative of a good similarity between the generated and reference images.
What was the conclusion drawn from testing the CREF parameter with multiple reference images?
-The conclusion was that there is no clear relationship between the number of reference images provided and the similarity of the generated images to the reference. Sometimes, fewer reference images resulted in more similar images, but this was not consistent.
How did the video address the issue of gender and age affecting the CREF parameter's performance?
-The video found that the CREF parameter performed better with male faces than female faces and better with children than adults, but this was not a definitive rule and the data was limited.
What was the outcome of using the CREF parameter in combination with the CW parameter for local redraws?
-The attempt to improve similarity by using the CREF parameter with a reduced CW value for local redraws did not yield significant improvements in the generated images' similarity to the reference, and in some cases, the results were less visually appealing.
What was the result of combining the CREF parameter with the SREF parameter?
-The combination of CREF and SREF parameters did not show a significant difference in the similarity of the generated images compared to using the CREF parameter alone, suggesting that this method may not be the most effective for improving facial similarity.
What was the creator's final recommendation for achieving a more realistic facial depiction with the CREF parameter?
-The creator recommended that the most straightforward and effective method for achieving a realistic facial depiction with the CREF parameter might be to simply use it and iterate through multiple attempts ('brushing' images) to find a satisfactory result, rather than relying heavily on the other tested methods.
Why did the creator have low expectations for the CREF parameter's performance with real photos?
-The creator had low expectations because the official documentation for the CREF parameter stated that it is most effective with images generated by MJ and not specifically designed for real photos, which suggests inherent limitations when applied to actual human facial images.
What is the creator's speculation about the underlying technology of the CREF parameter?
-The creator speculates that the CREF parameter might be based on IP-Adapter technology rather than FaceID or InstantID, based on personal experience and the observed performance of the parameter.
Outlines
🎨 Exploring CREF Parameters for Realistic Style Transfer
The paragraph discusses the use of the CREF (Character Reference) parameter in the context of generating images using the MJ platform. The creator addresses viewer concerns about the discrepancy between the generated images and the reference images, particularly when using Obama's image as an example. To further investigate, the creator selects Asian faces to test the CREF parameter's effectiveness in creating realistic style transfers. The video introduces a method to quantitatively measure similarity using ComfyUI's facial data recognition nodes, which calculate similarity through EUC and COS values. The creator sets a standard for what constitutes a good match and explains how these values can be used to determine the similarity between images.
📉 Analyzing the Impact of CREF Parameter Variations
This section delves into the impact of using multiple reference images with the CREF parameter. The creator conducts experiments by submitting varying numbers of reference images and observes the resulting facial similarity scores. The results indicate that increasing the number of reference images does not consistently improve the similarity of the generated images to the reference. The creator also notes that the CREF parameter's performance is not ideal for creating realistic Asian facial features, as none of the generated images meet the set standards for similarity. The discussion includes a comparison of different facial features and the influence of gender and age on the results.
🖌️ Testing Additional Methods for Enhancing Facial Similarity
The creator explores alternative methods to improve the facial similarity in generated images. These methods include using the CREF parameter in conjunction with local redraws and the SREF (Style Reference) parameter. The creator tests these methods using the best-performing images from the previous tests. The results show that local redraws with a reduced CW value do not significantly enhance the similarity, and the use of SREF does not markedly improve the outcomes. The creator concludes that the CREF parameter's effectiveness for realistic facial style transfer is limited, especially for Asian faces, and suggests that manually iterating through images ('brushing') may yield better results.
Mindmap
Keywords
💡MJ
💡CREF
💡EUC
💡COS
💡ComfyUI
💡IP-Adapter
💡FaceID
💡InstantID
💡SREF
💡局部重绘 (Local Redraw)
💡写实风格 (Realistic Style)
Highlights
The video discusses the use of the CREF (character reference) parameter in MJ, a tool for generating images.
Viewers reported that images generated using CREF did not closely resemble the reference images.
The video uses Asian faces as examples to test the CREF parameter's effectiveness.
ComfyUI's facial data recognition nodes are used to calculate similarity scores.
EUC and COS values are introduced as metrics for facial similarity.
The video sets standards for what constitutes a 'similar' facial image based on EUC and COS values.
Experiments are conducted to see if submitting multiple reference images improves resemblance.
局部重绘 (local redraw) is tested to improve facial similarity after initial image generation.
Combining CREF with SREF parameters is explored for better facial similarity.
The video concludes that CREF's performance in facial migration for Asian faces is subpar.
No image generated with CREF meets the similarity standards set at the beginning of the video.
The number of reference images does not have a clear impact on the final image's resemblance.
CREF performs better with male faces than female, and better with children than adults.
Using a lower CW value in conjunction with CREF does not significantly improve facial similarity.
Combining CREF with SREF does not yield significantly different results compared to using CREF alone.
The video suggests that manually 'brushing up' images with CREF might be the fastest way to achieve a realistic facial depiction.
The CREF parameter is acknowledged as not being designed for real-person facial migration.
The video creator plans to continue sharing tutorials related to MJ in future content.