Generate Wallpapers Just by Writing in Japanese: The Mechanism of "Kabegami Studio" Completed via Local AI
Simply by entering "A city street at night after the rain," your PC draws a wallpaper for you. We will explain in detail the mechanism of "Kabegami Studio," a tool that makes such an experience possible.
- A local LLM analyzes Japanese scenes and translates them into prompts for image generation.
- Everything is completed within your own PC without using external APIs, ensuring privacy and cost-efficiency.
- To overcome VRAM constraints, a memory management system was introduced that swaps between the text AI and the image generation AI.
- The mechanism embeds the prompt into the generated image, allowing settings to be restored later.

Mechanism for Converting Japanese Scenes into "Instructions for Images"
Until now, to create ideal wallpapers with AI, it was necessary to think of "prompts (instructions for the AI)" in English and in great detail. However, it is not easy to directly connect content written intuitively in Japanese to image generation.
Therefore, we introduced a translation process utilizing a local LLM (a large language model that runs on your own PC). Specifically, we use the gemma4:12b-it-q8_0 model. This AI reads Japanese sentences such as "A city street at night after the rain" and translates them into "tags (a list of keywords)" that the image generation AI can easily understand.
A major feature of this tool is that it does not use any external APIs (mechanisms that borrow other companies' AI servers via the internet). Since all processing is performed using only the resources within your own PC, it is designed so that you can try and error multiple times in your preferred environment without worrying about information leaks.
- Text AI: gemma4:12b-it-q8_0 (Analyzes and translates Japanese scenes)
- Drawing AI: Stable Diffusion WebUI Forge (Uses SDXL-based models)
- Flexibility: It is also possible to generate only the prompt first to check and revise the content before drawing.
Memory Management to Maximize Limited Resources
When running advanced AI in a local environment, the biggest challenge is the capacity of "VRAM (Video Memory)". VRAM is like a dedicated workspace for image generation AI and other models to perform calculations.
In this configuration, gemma4:12b (q8) requires about 12.8GB of VRAM, and high-resolution enhancement for the SDXL model requires about 9GB of VRAM. If you try to run these simultaneously, they will exceed the capacity of a typical 16GB VRAM environment, resulting in errors.
To solve this problem, this tool is designed to automatically perform a process that "unloads the text AI from memory" just before drawing. The mechanism hands over the memory occupied by the text AI to the image generation AI at the moment the text AI finishes writing the prompt. Because of this, generation is always performed one image at a time in order, but it can operate stably even with limited specs.
- Required amount for gemma4:12b (q8): approx. 12.8GB
- Required amount for SDXL high-resolution enhancement: approx. 9GB
- Countermeasure: Perform drawing after evicting the text AI from memory (generation one by one)
Ease of Management and Reuse After Generation
It would be a waste if you couldn't reproduce your favorite images after generating them. In Kabegami Studio, generated images can be managed collectively in a gallery function. We have made it possible to smoothly register to favorites, save in PNG format, and reuse prompts.
Particularly convenient is the fact that the "prompt (instruction)" is embedded inside the saved PNG file. This makes it easy to restore the settings of that image later and generate different variations.
- Gallery Function: Favorite registration, PNG saving, and prompt reuse are possible
- Prompt Embedding: Settings can be restored from saved images
- People who want to run AI on their own PC but are worried about memory constraints
- People who are not good at thinking of English prompts and want to give intuitive instructions in Japanese
- People who want to manage and reproduce generated images later
- Local LLM
- A large language model that runs using your own PC's performance without using cloud services (such as ChatGPT).
- VRAM
- Memory installed on a graphics board, corresponding to the workspace when running image generation AI.
- Prompt
- Instructions or keywords used to tell the AI "what you want it to draw."
FAQ
Can I run it with 16GB of VRAM?
According to official information, gemma4:12b (q8) requires about 12.8GB of VRAM, and SDXL high-resolution enhancement requires about 9GB of VRAM. Therefore, it is not possible to run these simultaneously, but by using a mechanism that unloads the text AI from memory before drawing, it is possible to generate images one by one even in a 16GB environment.
"Kabegami Studio" is a tool that achieves the experience of generating wallpapers from Japanese scene descriptions by cleverly switching between local AIs. It solves memory constraints through technical ingenuity (swapping processes), balancing the sense of security of completing everything on your own PC with easy management by embedding prompts.