retowyss/PromptBridge-0.6b-Alpha

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kPublished:Jan 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

retowyss/PromptBridge-0.6b-Alpha is a specialized Qwen3-0.6B model fine-tuned for bidirectional prompt transformation for text-to-image generation. It can expand brief keywords into detailed prompts or compress lengthy prompts into concise keywords or single sentences. This model is optimized for generating variations of image prompts, specifically for single, adult human(oid) subjects.

Loading preview...

PromptBridge-0.6b-Alpha: Bidirectional Prompt Transformation

PromptBridge-0.6b-Alpha is a specialized Qwen3-0.6B model developed by Reto Wyss, designed for transforming text-to-image prompts. Its core functionality revolves around two key operations:

Key Capabilities

  • Prompt Expansion: Converts short keywords or single sentences into highly detailed prompts suitable for image generation models.
  • Prompt Compression: Distills extensive image generation prompts into succinct keywords or single sentences.
  • Prompt Variation: Enables the generation of prompt variations by chaining expansion and compression operations.

Training and Specialization

The model was trained on approximately 300,000 synthetic prompt pairs (300M tokens), achieving a final perplexity of 3.43. It is exclusively trained on prompts featuring a single, adult human(oid) subject, making it highly specialized for this domain. The training data's content rating ranges from PG to R, with explicit (X-rated) content and prompts containing minors or brand references filtered out.

Usage Considerations

PromptBridge-0.6b-Alpha requires specific system prompts for its operations: "Expand the prompt.", "Compress the prompt into one sentence.", or "Compress the prompt into keyword format.". It is not designed for general instruction-following, reasoning, or chat functions. Users should be aware of potential token length limitations and that its performance is optimized solely for human subject descriptions.