Stableboost

Stableboost implements a queue-based image generation pipeline that accepts multiple prompts and generates images in batches, optimizing GPU utilization by processing multiple inference requests sequentially or in parallel depending on available VRAM. The system maintains a job queue that tracks generation status, parameters, and outputs, allowing users to submit dozens or hundreds of prompts and retrieve results asynchronously without blocking the UI.

Stableboost enables systematic exploration of generation parameter space by allowing users to specify ranges or lists for seed, guidance scale, steps, and other Stable Diffusion parameters, then automatically generating images across all combinations or a sampled subset. This creates a structured exploration matrix where each axis represents a parameter variation, helping users understand how each setting affects output quality and style.

Stableboost organizes generated images in an interactive gallery interface with side-by-side comparison, filtering, and tagging capabilities. Users can mark favorite images, group results by prompt or parameters, and export curated subsets. The gallery maintains metadata for each image (generation parameters, timestamp, prompt) enabling retroactive analysis and filtering based on quality or aesthetic criteria.

Stableboost provides live progress indicators for each image in the generation queue, showing step-by-step completion percentage and estimated time remaining. Users can cancel individual generation jobs or the entire queue without losing previously completed images. The system uses WebSocket or polling to update the UI in real-time, and maintains a persistent queue state so users can pause and resume generation sessions.

Stableboost abstracts Stable Diffusion model loading and switching, allowing users to select from multiple installed checkpoints (base models, fine-tuned variants, LoRA adapters) through a UI dropdown without restarting the backend. The system manages model memory efficiently by unloading unused models and caching frequently-used ones, reducing the overhead of switching between different model variants during exploration.

Stableboost supports prompt templates with variable placeholders that can be substituted with lists of values, enabling systematic prompt variation without manual editing. Users can define templates like 'a {style} painting of a {subject}' and provide lists for {style} and {subject}, which generates the Cartesian product of all combinations. This reduces prompt engineering overhead and ensures consistency across variations.

Stableboost provides explicit seed management allowing users to fix seeds for reproducible outputs or randomize them for diversity. Users can specify a seed range, generate images with the same seed across different prompts/parameters to isolate the effect of those changes, or use random seeds for exploration. The system displays the seed used for each image in metadata, enabling retroactive reproduction of specific outputs.

Stableboost supports negative prompts (concepts to avoid) with optional weighting to control their influence on generation. Users can specify multiple negative prompts and adjust their relative strength, allowing fine-grained control over what the model should NOT generate. The system may support syntax for weighted negative prompts (e.g., '(bad quality:0.7), (blurry:0.5)') enabling nuanced exclusion of undesired attributes.