Image Generation from Contextually-Contradictory Prompts

This demo accompanies our paper on Image Generation from Contextually-Contradictory Prompts. The source code is available on GitHub. Our SAP (Stage Aware Prompting) method supports multiple diffusion models and can be paired with various large language models (LLMs). This interface allows you to generate images using:

• FLUX.dev: Baseline image generation using the unmodified FLUX model.
• SAP with zephyr-7b-beta: SAP applied to FLUX with zephyr-7b-beta as the LLM.
• SAP with GPT-4o: SAP applied to FLUX with GPT-4o as the LLM (requires an OpenAI API key).

For best results, we recommend using SAP with GPT-4o, which delivers the best implementation of our method.

Note: When using SAP with zephyr-7b-beta, the model may take a few seconds to load on the first run, as the LLM is initialized. Subsequent generations will be faster.

Click a row to compare FLUX vs SAP