Guiding Instruction-based Image Editing via Multimodal Large Language Models
Guiding Instruction-based Image Editing via Multimodal Large Language Models
Instruction-based image editing improves the controllability and flexibility of image manipulation via natural commands without elaborate descriptions or regional masks. However, human instructions are sometimes too brief for current methods to capture and follow. Multimodal large language models (MLLMs) show promising capabilities in cross-modal understanding and visual-aware response generation via LMs. We investigate how MLLMs facilitate edit instructions and present MLLM-Guided Image Editing (MGIE). MGIE learns to derive expressive instructions and provides explicit guidance. The editing model jointly captures this visual imagination and performs manipulation through end-to-end training. We evaluate various aspects of Photoshop-style modification, global photo optimization, and local editing. Extensive experimental results demonstrate that expressive instructions are crucial to instruction-based image editing, and our MGIE can lead to a notable improvement in automatic metrics and human evaluation while maintaining competitive inference efficiency.
| Input | Instruction | InsPix2Pix | LGIE | MGIE | GroundTruth |
|---|---|---|---|---|---|
 |
turn the day into night |
 |
 |
 |
 |
 |
make the forest path into a beach |
 |
 |
 |
|
 |
make the frame red |
 |
 |
 |
|
 |
as if the shop was a library |
 |
 |
 |
|
 |
make it the vatican |
 |
 |
 |
|
 |
turn the sunset into a firestorm |
 |
 |
 |
|
Apple ML-MGIE, or Multimodal Large Language Models Guided Instruction-based Image Editing, is an advanced technology developed by Apple. It leverages the power of multimodal large language models (MLLMs) combined with diffusion models for instruction-based image editing. This technology aims to edit and generate images by understanding textual instructions from users, showcasing the potential of AI in creativity and image processing.
ML-MGIE operates by integrating two core technologies: Multimodal Large Language Models (MLLMs) for cross-modal understanding and response generation, and diffusion models for high-quality image generation. This integration bridges MLLMs and diffusion models in image editing, providing superior performance compared to existing technologies like InstructPix2Pix. ML-MGIE uses an algorithm called LLaVA to derive expressive instructions for enhanced instruction-based image editing, enabling it to understand and execute complex image editing tasks from concise human instructions.
Compared to existing instruction-based image editing technologies, ML-MGIE demonstrates significant advantages. By combining MLLMs with diffusion models, it not only understands more complex instructions but also generates higher quality images. The use of the LLaVA algorithm further enhances its comprehension and execution abilities, surpassing existing technologies in performing complex image editing tasks.
ML-MGIE can generate responses to visual content through language models, meaning it can understand image content and generate relevant textual descriptions or answer questions related to the image. This capability is particularly useful in providing image descriptions, augmented reality applications, and visual data analysis.
ML-MGIE exhibits strong capabilities in cross-modal understanding, linking information across different modalities (e.g., text and image) for comprehensive understanding. For example, it can enhance scene understanding by analyzing image content alongside relevant textual descriptions. This cross-modal comprehension is vital for improving human-computer interaction, enhancing search engine results, and creating more intelligent educational tools.
A significant application of ML-MGIE is guiding instruction-based image editing. It can edit images according to user instructions, such as changing the color, shape, or size of objects within an image. This is achieved by integrating multimodal large language models with diffusion models, where ML-MGIE shows superior performance compared to technologies like InstructPix2Pix. This capability can be applied to automated image editing tools, improving the efficiency and accuracy of image editing.
Currently, the specific usage instructions for Apple ML-MGIE have not been made public. However, based on the available information, users will be able to guide ML-MGIE in image editing by providing natural language instructions. The detailed code and usage methods will be published after the internal review is completed.
Apple ML-MGIE is currently available for a free trial. You can click "try it for free" at the top of this website to access the free trial.
The primary application scenario for Apple ML-MGIE is image editing. Users can guide ML-MGIE in editing images by providing natural language instructions, such as "change the background color to blue" or "add a smiling face icon at the top right corner of the image."
The advantages of Apple ML-MGIE mainly lie in its strong cross-modal understanding, visual perception response generation capabilities, and superior performance in image editing. However, there is currently no clear information available regarding the disadvantages of ML-MGIE.
Apple's ML-MGIE is a multimodal large language model for guided image editing that uses LLaVA (Language for Visual Arts) to generate expressive instructions for enhanced instruction-based image editing. It is the first work to combine multimodal large language models with diffusion models for image editing, showcasing superior performance compared to InstructPix2Pix.
For other image models, several are worth recommending and comparing:
