「AIは同じキャラが出せない」はおしまい！新機能Reference-onlyなら学習なしでも画像だけで差分作成からアイコラも出来ちゃうぞ！！【Stable Diffusion】

In this video, the presenter introduces the new 'Reference Only' feature in Control Net version 1.1.153, which allows for the creation of character images without the need for a trained file. The feature is showcased as being able to strongly inherit character elements like hair and eye color, even without specific prompts. By combining Reference Only with Tagger and Chara LoRA, users can easily generate cosplay images or different patterns of a character. The presenter demonstrates how to use Reference Only to create a cosplay image by combining the original character's neck with another character's body, and fixing the pose with an open pose in a control net. The video concludes by emphasizing the powerful capabilities of Reference Only and its combination with Tagger as a game-changing feature for character image creation.

Takeaways

• 🆕 The new feature 'Reference Only' allows for the creation of derivative images without the need for a trained file, making it easier to generate unique images.
• 🔄 'Reference Only' will be automatically implemented in Control Net version 1.1.153 and can be found in the Pre-Processor.
• 🎨 This feature is particularly adept at inheriting character elements such as hair and eye color, even without specific prompts.
• 🧬 By using 'Reference Only', one can accurately reproduce character elements across different images, with a high reproduction rate from the neck up.
• 📈 The feature is beneficial for quickly increasing the number of materials when learning from a small dataset.
• 🤩 It enables the mass production of images with different patterns while maintaining character consistency, even without a trained file.
• 🔗 When combined with Tagger, 'Reference Only' can precisely reproduce facial features, making it ideal for creating cosplay images.
• 🧞‍♂️ The combination of 'Reference Only' and Chara LoRA allows for easy cosplay without the need for additional models like LoRA.
• 🧑‍🎤 By using two 'Reference Only' instances, one can combine the face and body from different images to create a cosplay image.
• 👚 It's possible to use 'Reference Only' with an outfit extracted from a fashion brand's website to apply that style to a character.
• 🌟 'Reference Only' is considered the strongest feature of Control Net version 1.1, offering a powerful tool for image creation and customization.
• 📚 The feature is highly recommended for use due to its ability to enhance the creative process and produce high-quality, character-consistent images.

Q & A

What is the new feature called Reference Only in the Control Net version 1.1.153?

Reference Only is a feature that allows for the creation of images with specific character elements, such as hair and eye color, without the need for a trained file. It's particularly useful for generating a variety of images that maintain the character's distinctive features.

How does the Reference Only feature differ from Shuffle in terms of character guidance?

Shuffle focuses more on the texture of the input image, hairstyle, and background, while Reference Only has a stronger ability to inherit character elements like hair and eye color, even without explicit support from the Prompt.

What is the role of Tagger in the process described in the script?

Tagger is used to analyze images and transfer the analysis results to the Prompt. When combined with Reference Only, it ensures that the face and other character-specific elements are accurately represented in the generated images.

How can Reference Only help in creating a DEKAPAI sub-image?

By strongly inheriting character elements from a single image, Reference Only can accurately capture the character's features and generate different images, making it easier to create a DEKAPAI sub-image.

What is the significance of combining Reference Only with Chara LoRA for cosplay images?

Combining Reference Only with Chara LoRA allows for the creation of cosplay images by inheriting the character's features strongly from the neck to the top, making it possible to easily cosplay without needing a trained file.

How can two Reference Only images be used to create a cosplay image without LoRA?

By using the original character's raw neck with the body of other characters and leveraging the powerful style inheritance ability of Reference Only, a cosplay image can be created without LoRA, simply by combining the two images.

What is the advantage of using Tagger in conjunction with Reference Only?

The combination of Tagger and Reference Only ensures that the face and other character-specific elements are accurately represented in the generated images, making it a powerful tool for creating images with specific character features.

How does the Reference Only feature change the process of creating iCola-like images?

With Reference Only, the need for a review file is reduced, as it allows for the creation of iCola-like images with just two images, making the process more efficient and less reliant on a large amount of training material.

What is the importance of the Control Net version 1.1.153 update for users interested in character-based image generation?

The update introduces the Reference Only feature, which significantly simplifies the process of creating character-based images, allowing for greater flexibility and creativity without the need for extensive training files.

How does the Reference Only feature facilitate the mass production of different pattern images?

Reference Only allows for the reproduction of character elements from a single image, enabling the mass production of images with different patterns while maintaining the character's identity.

What is the impact of the Reference Only feature on the cosplay community?

The Reference Only feature makes it easier for cosplayers to create images that accurately represent their chosen characters, potentially leading to more diverse and accurate cosplay representations.

Can the Reference Only feature be used with images from fashion brands for creating character outfits?

Yes, by extracting a sample image from a fashion brand's website and analyzing it with Tagger, the Reference Only feature can be used to apply the outfit to a character, allowing for the creation of new character designs.

Outlines

00:00

🎨 Introduction to Reference Only Feature

The speaker, Robin, introduces the Reference Only feature in Control Net version 1.1.153, emphasizing its ability to create images with strong character inheritance without the need for a trained file. The feature is automatically implemented post-update and is found in the Pre-Processor. The speaker demonstrates the feature by creating an image with minimal prompts and compares it to the Shuffle feature. Reference Only is shown to be effective in inheriting character elements like hair and eye color, and when combined with Prompt and Tagger, it can accurately reproduce the character's face. The speaker also explores the possibility of creating cosplay images by combining Reference Only with different character parts, showcasing its potential for diverse image generation.

05:04

🚀 Mass-Producing Images with Reference Only

The speaker discusses the powerful capabilities of the Reference Only feature for mass-producing images of the original character's whole body. They highlight that this feature is a must-try and a game-changer for Control Net version 1.1, allowing for the creation of a variety of images with minimal input. The speaker concludes by encouraging everyone to experiment with Reference Only and looks forward to the next video.

Keywords

AI

AIは人工知能の略で、コンピューターシステムが人間の思考や行動を模倣すること。このビデオでは、AIが画像を生成し、キャラクターの要素を引き継ぐ方法を紹介しています。

Stable Diffusion

Stable Diffusionは、画像生成のためのAI技術の一つで、高品質な画像を生成することができます。このビデオでは、Stable Diffusionの新しい機能であるReference-onlyを使用して、新しい画像を作成する方法が説明されています。

Reference-only

Reference-onlyは、Stable Diffusionの新機能で、訓練されたファイルなしで画像を参照して要素を引き継ぐことができます。この機能を使うことで、キャラクターの髪色や目の色などの要素を強く引き継ぐことができます。

Control Net

Control Netは、AIが画像を生成する際に使用する技術の一つで、画像の特定の要素をより正確に制御することができます。このビデオでは、Control Netのバージョン1.1.153以降、Reference-onlyが自動的に実装されることが説明されています。

Shuffle

Shuffleは、このビデオの前半で紹介された機能で、キャラクターのテクスチャや髪型、背景などの要素に強い指導力を持ちます。しかし、全体的なハイスピード力はReference-onlyに劣ります。

Tagger

Taggerは、画像の分析結果をPromptに変換するツールで、学習LoRAのビデオで使用されました。このビデオでは、Taggerを使って画像から分析結果を取り出し、Promptに転送することで、より正確なキャラクターの再現を実現しています。

LoRA

LoRAは、AIが学習したデータを使って画像を生成する技術です。このビデオでは、LoRAとReference-onlyを組み合わせることで、キャラクターの顔や身体的特徴を正確に再現しながら、コスプレの画像を簡単に作成できることが示されています。

Cosplay

コスプレは、漫画やアニメ、ゲームなどのキャラクターを模倣して衣装を着ることを指します。このビデオでは、Reference-only機能を使って、オリジナルのキャラクターの画像を簡単にコスプレできることが示されています。

Character Elements

キャラクター要素とは、キャラクターの外見の特徴を指します。このビデオでは、Reference-only機能を使って、髪色や目の色などの要素を引き継いで、異なる画像を生成することができます。

Image Generation

画像生成は、AI技術を使って新しい画像を作り出すプロセスです。このビデオでは、Stable DiffusionのReference-only機能を使って、指定されたキャラクターの要素を引き継いで、新しい画像を生成する方法が紹介されています。

DEKAPAI sub-image

DEKAPAI sub-imageは、ビデオの中で言及された「デカパイサブ画像」の略で、キャラクターの胸の大きなサイズを指します。この機能を使って、キャラクターのこの要素を強化して異なるパターンの画像を大量に生成することができます。

Highlights

Introduction of a new feature called Reference Only in Control Net version 1.1.

Reference Only allows for character element inheritance without a trained file.

Automatic implementation of Reference Only in Control Net version 1.1.153.

Reference Only is located in the Pre-Processor of Control Net.

Demonstration of creating a separate image with minimum spellings using Reference Only.

Comparison of Reference Only with Shuffle in terms of character guidance and texture focus.

Use of Wife Diffusion 1.4 Tagger to analyze and transfer results to Prompt for image generation.

High reproduction rate of character elements from the neck to the top with Reference Only.

Potential for mass-producing images of different patterns while maintaining character identity.

Combining Reference Only with Tagger for accurate facial reproduction.

The ability to cosplay without LoRA by combining face and body from two Reference Only images.

Creating cosplay images by combining the original character's neck with other characters' bodies.

Using Tagger to analyze sample images for outfit application in Reference Only.

The emergence of new designs through AI-specific interpretation.

The combination of Reference Only and Tagger as the optimal tool for character cosplay.

The significance of Reference Only in reducing the need for review files in character cosplay creation.

The ability to mass-produce whole body images and world difference with just an original character image.

Reference Only being the strongest feature of Control Net version 1.1.

Encouragement for everyone to try Reference Only for its powerful capabilities.