Midjourney v6 News: SREF v2 Released
TLDRIn the latest update, Midjourney's Style Reference Engine (SREF) has been upgraded to version 2, promising a more precise understanding of style and less leakage of non-style elements into generated images. David Holtz, the developer, explains that the functionality remains similar, but improvements are internal. The new version allows for style weighting from 0 to 1,000, and the old version is no longer available. The video provides a deep dive into the feature with examples, comparing the effects of different prompts and style references on image generation. The results show that the new version is adept at applying stylistic elements without introducing unwanted features, even when combining styles. The video also explores the impact of varying the style weight, demonstrating how it can subtly or drastically alter the generated images. The presenter encourages viewers to share their thoughts on the new version and to subscribe for more content.
Takeaways
- 📈 The Midjourney style reference (SREF) has been updated to version 2, which is said to be more precise in understanding style and preventing non-style elements from leaking into generated images.
- 🔍 David Holtz mentions that the new version should be better at handling style, although the functionality remains largely the same as before.
- 🚫 The old version of SREF is no longer available, and users can now use the updated version with the 'sref' and 'ssw' parameters for style weight.
- 🌐 The script provides examples of image generation using different prompts and style references to illustrate how the new version works.
- 🎨 The use of the word 'eldritch' in the prompt, along with a style reference, results in images with darker colors and more tentacles, demonstrating the influence of the style reference.
- 📉 When the word 'eldritch' is removed and only 'fantasy' is used as a prompt with a style reference, the resulting images are more aligned with the 'eldritch' style, indicating the precision of the new SREF version.
- 🐱 An example of applying the style reference to a 'drawing of a cat' prompt shows how the style can be subtly integrated without completely overtaking the original subject.
- 🔄 The script discusses the behavior of the style weight at different breakpoints, showing how increasing the style weight can drastically change the generated image, from a slight change to a full transformation into a 'tentacle monster'.
- 🔍 The comparison between the default style weight and higher values reveals that extreme ends of the scale can lead to very different results, often losing the original subject matter.
- 🤔 The speaker acknowledges the complexity of the different scales and their combinations, suggesting that it's challenging to understand and predict their behavior.
- ⚖️ It is suggested that the best results are often achieved with the default values or moderate adjustments, rather than extreme settings.
- 💭 The video concludes by inviting viewers to share their thoughts on the new style reference version and to subscribe for more content.
Q & A
What is the latest update about Midjourney's style reference?
-The latest update is that Midjourney's style reference has been updated to version 2, which is reported to be much more precise in understanding style and better at not leaking non-style elements into the generated images.
How does the new version of style reference work?
-The new version of style reference works in a similar way to the previous one. Users can still use it with the 'sref' command and have the 'ssw' available as the style weight, which ranges from 0 to 1,000.
Why was the seed number set to 777 for the image generations?
-The seed number was set to 777 for each of the images to ensure that the generations are always set to the same seed number, allowing for consistent comparison and to show that any changes in the output are due to modifications made to the prompt or style reference.
What was the effect of adding the word 'eldrich' to the prompt 'fantasy'?
-Adding the word 'eldrich' to the prompt 'fantasy' resulted in images with darker colors and slightly more tentacles, indicating that the style reference is capable of influencing the stylistic elements of the generated images.
How does the style reference version 2 differ from the older version?
-The version 2 of the style reference is said to be better at applying styles accurately, with the colors being spot on and not introducing unwanted elements like tentacles, which the older version might have included even if not specified.
What happens when the style weight is increased from 0 to 1,000?
-As the style weight increases from 0 to 1,000, the influence of the style reference on the image becomes more pronounced. At lower weights (e.g., 10 or 50), the changes are subtle, but at higher weights (e.g., 200, 500, 1,000), the images become increasingly dominated by the style reference, to the point where the original prompt's elements may be unrecognizable.
What is the significance of using a consistent seed number for image generations?
-Using a consistent seed number allows for a controlled experiment where the only variable that changes is the prompt or style reference. This helps in isolating the effects of different parameters on the image generation process.
How does the style reference interact with the original prompt?
-The style reference interacts with the original prompt by overlaying or modifying the stylistic elements of the generated image according to the style specified. It can enhance or alter the atmospheric qualities, colors, and other stylistic features based on the weight assigned to the style reference.
What is the recommended approach for using the style weight?
-The best results are achieved when the style weight is set at either the default value or at moderate values. Extreme ends of the scale are not recommended as they can lead to overstylization or loss of the original prompt's essence.
Why is it difficult to understand the behavior of all the different value scales in combination?
-The difficulty arises from the vast number of possible combinations of different scales. Even if one were to select just four breakpoints for each scale, the number of combinations would exceed 4,000, making it impractical to show or check every possible outcome.
How does the presenter suggest viewers engage with the content?
-null
What is the presenter's final note regarding the exploration of style reference combinations?
-The presenter acknowledges the complexity of exploring all possible combinations of style reference parameters and suggests that viewers might find the default or moderate values to be the most effective for their image generation needs.
Outlines
🎨 Mid Journey Style Reference Update - Version 2
The video discusses an update to the Mid Journey style reference, which has been upgraded to version 2. The presenter, Th, explains that this feature has become more precise in understanding style and better at excluding irrelevant elements from image generation. The style reference operates similarly to its previous version, utilizing the '--sref' and '--ssw' parameters to adjust the style weight. Th demonstrates the functionality of the updated style reference through various image generation examples, comparing the results with different prompts such as 'fantasy' and 'eldrich'. The video also explores how the style reference influences the generated images at different style weights, from subtle changes at weight 10 to extreme transformations at weight 1,000. Th concludes by inviting viewers to share their thoughts on the new version and their usage of different scales in their work.
📺 Viewer Engagement and Content Subscription
In the second paragraph, the presenter encourages viewers to subscribe to the channel for more content like the one they just watched. The speaker also expresses appreciation for likes and views, signaling the end of the video with a thank you note and a continuation prompt, suggesting more content to come.
Mindmap
Keywords
💡Midjourney v6
💡SREF v2
💡David Holtz
💡Style Weight
💡Prompt
💡Seed Number
💡Eldritch
💡Fantasy
💡Tentacles
💡Cat Monsters
💡Combination of Scales
Highlights
Midjourney v6 introduces an update to the style reference (SREF) to version 2, which is claimed to be more precise in understanding style.
David Holtz mentions that the new version should prevent non-style elements from leaking into generated images.
The functionality of style reference remains the same, with changes occurring on Midjourney's end.
The style weight can be adjusted from 0 to 1,000 using the 'sref' and 'ssw' parameters.
The old version of the style reference is no longer available.
Examples are provided to demonstrate how the style reference works with various prompts.
The prompt 'fantasy' generates atmospheric and slightly painterly images.
Adding 'eldrich' to the prompt 'fantasy' results in darker colors and more tentacles.
Using the entire grid as a style reference for 'eldrich' produces images with accurate colors and a controlled number of tentacles.
The older version of SRE might have included unintended elements like tentacles.
The style and eldrich styles are noted to be quite similar, which can affect the outcome.
An 'eldrich drawing of a cat' prompt results in a unique image that maintains some of the original drawing's style.
The style weight's impact on the image is demonstrated with varying levels, from subtle changes at weight 10 to extreme transformations at weight 1,000.
At style weight 50, the image begins to show more tentacles, indicating a shift from the default style.
At style weight 200, the image becomes extreme with a significant increase in tentacles.
At style weight 500, the original elements of the cat or drawing are no longer recognizable.
The complexity of different value scales and their combinations make it challenging to predict outcomes.
The best results are suggested to be at the default value or moderate scales, avoiding extreme ends.
The presenter invites viewers to share their thoughts on the new style reference version and their usage of different scales.
The video concludes with a call to action for viewers to subscribe to the channel and like the content.