Bildgenerierung¶
Bildgenerierung is a web application for generating and editing images from natural-language descriptions. The interface combines text-to-image and image-to-image generation in a single input form and accesses an inference service connected via the OpenAI API and operated as part of the local infrastructure. Two interchangeable image models are currently supported.
At a glance¶
- Generate images from text descriptions — from short keywords to detailed scenes, including text rendered within the image
- Edit existing images — provide reference images and have them modified in a targeted way
- Produce multiple variants in parallel — compare up to four images per run
- Work iteratively — adopt a result directly as a reference image for the next run
- Reach a result quickly via presets — set resolution, image count, level of detail, and prompt adherence through two preset selectors
- Achieve reproducible results — using a fixed seed reproduces identical images
- Download generated images — directly from the gallery as PNG files with descriptive filenames
Highlights¶
The application encapsulates the call to an image model — which would otherwise require a non-trivial direct API request — behind an interface tailored to typical tasks. The following properties go beyond a simple call and have direct effects on result quality and on the interchangeability of models:
- Connection via the OpenAI API — Communication with the inference service uses the standard OpenAI endpoints for image generation and chat completion. As a result, models and servers can be swapped without changes to the application or the workflow.
- Two generation modes in a single interface — Depending on whether reference images are present, the appropriate API path is selected automatically (text-to-image or image-to-image). Switching between modes is transparent to the user and happens in the same input form.
- Connection to the local infrastructure — Inference runs on a server operated as part of the local infrastructure. Inputs, reference images, and generated images do not leave it.
- Choice between multiple image models — Currently supported are Qwen-Image-2512 and black-forest-labs/FLUX.2-dev through the same interface. The active model is configured centrally.
- Preset system for common cases — Presets for mode (resolution and image count) and quality (inference steps and guidance scale) allow a quick start without requiring familiarity with the underlying parameters.
- Manual parameters override presets — Resolution, image count, inference steps, guidance scale, and seed can be set individually through an accordion and complement the presets without hiding them.
- Iterative image editing — A generated image can be adopted as a reference for the next run with a single click, allowing gradual refinement towards a desired result.
- Negative prompts — Unwanted image elements can be excluded explicitly, reducing the share of unusable generations.
- Reproducibility via seed — With a fixed seed, identical inputs yield identical outputs. This supports systematic variation of individual parameters and the reconstruction of earlier results.