01
Input
The process starts with a description provided by the user through an interface designed to work on both desktop and mobile devices.
Project
ai.vjml.es
A self-hosted image generation service built around local inference and an automated pipeline that transforms an initial idea into a request prepared for the generation system.
01 / Overview
Image generation on self-hosted infrastructure
ai.vjml.es explores how to provide image generation through a simple interface while keeping processing within controlled infrastructure. The system automates request preparation before sending it to the local inference backend, separating the user experience from the complexity of the generation process.
02 / Pipeline
Pipeline
02
Prompt generation
The request is automatically transformed into a structured description better suited to the generation process, reducing the need to manually construct complex prompts.
03
Composition
The pipeline organizes and refines the elements of the request before constructing the final input used by the generation backend.
04
Local inference
The prepared request is processed by an inference backend running on self-hosted infrastructure and the result is returned to the interface as a generated image.
03 / Media
ai.vjml.es in use
The same generation and exploration experience is available on desktop and mobile, including information about each generated image.
04 / Architecture
Architecture
The application separates the public interface, request preparation and inference so that each part of the system can evolve independently.
Web frontend
The public interface where generation requests are created and results are presented, designed for both desktop and mobile.
Prompt pipeline
The layer responsible for interpreting the request, structuring it and automatically constructing the input used to generate the image.
Generation service
The service coordinates generation requests while keeping the Web application decoupled from the inference process.
Local inference
Generation runs locally on controlled infrastructure without exposing implementation details of the inference backend.
Distributed deployment
The Web layer and inference workload can run on different systems, allowing each task to be assigned to the most appropriate node.
Iterative development
Component separation makes it possible to experiment with and evolve the pipeline without redesigning the public interface or coupling it to a particular inference implementation.
05 / Development
Prompt pipeline v21
Current development
The current prompt pipeline is being refined to make better use of the information extracted from each request, improve composition and increase variation in generated results while keeping the user input simple.
06 / Links
