TLDR: Researchers propose VL-DNP, a new method for diffusion models that uses Vision-Language Models (VLMs) to dynamically generate negative prompts during image creation. This approach allows for real-time, context-aware content filtering, significantly improving safety (reducing unwanted content) while maintaining image quality, outperforming traditional fixed negative prompting.
Generating images with AI diffusion models has become incredibly powerful, but ensuring these models produce safe and appropriate content remains a significant challenge. One common technique for content filtering is “negative prompting,” where users specify concepts they want the model to avoid. However, this traditional approach has its drawbacks.
Traditional negative prompts are often fixed and predefined. This means they can sometimes over-correct, disrupting the image generation even when unwanted content isn’t present, leading to less accurate or “semantically drifted” images. Furthermore, it’s difficult to anticipate every potential unwanted element that might arise from a given positive prompt, making fixed negative prompts either inefficient or insufficient.
Introducing Vision-Language Guided Dynamic Negative Prompting (VL-DNP)
A new research paper titled “Dynamic VLM-Guided Negative Prompting for Diffusion Models” by Hoyeon Chang, Seungjin Kim, and Yoonseok Choi from KAIST introduces a novel solution: Vision-Language guided Dynamic Negative Prompting (VL-DNP). This innovative method leverages the advanced understanding capabilities of modern open-source Vision-Language Models (VLMs) to act as a dynamic negative prompt generator.
The core idea behind VL-DNP is to detect the emergence of unwanted content during the image generation process itself and then generate targeted negative prompts in real-time to suppress it. Crucially, this framework can be easily integrated into any existing diffusion model that uses Classifier-Free Guidance (CFG) without needing extensive retraining or model modifications.
How VL-DNP Works
The VL-DNP process unfolds in a few key steps:
First, the diffusion model begins its standard denoising process using the positive prompt. At specific, predefined timesteps during this process, the model predicts an intermediate version of the denoised image. This intermediate image, even if still somewhat blurred, is then sent to a VLM. The VLM analyzes this image and, based on its visual and language understanding, generates a contextually appropriate negative prompt. This dynamically generated negative prompt is then fed back into the diffusion model, guiding the subsequent denoising steps away from any detected unsafe or unwanted content. This adaptive approach ensures that the negative guidance is precise and relevant to the evolving image content.
Experimental Validation and Key Findings
To evaluate VL-DNP, the researchers used Stable Diffusion v1.4 as their base diffusion model and Qwen2.5-VL-7B-Instruct as the Vision-Language Model. They tested their approach on various datasets, including COCO-30K for general image generation performance, and specialized datasets like Ring-a-Bell-16, P4D, and Unlearn-diff, which are designed to test content filtering effectiveness against adversarial prompts.
The results demonstrate a significant improvement in the trade-off between safety and image quality. While traditional static negative prompting can reduce unwanted content, it often comes at the cost of degrading the overall image quality and alignment with the positive prompt. VL-DNP, however, manages to drastically cut down on unsafe content (measured by Attack Success Rate and Toxic Rate) while remarkably preserving the image quality and text-image alignment (measured by CLIP score and FID).
For instance, increasing the negative guidance scale with static prompting led to a notable drop in CLIP score and an increase in FID. In contrast, VL-DNP maintained the CLIP score almost constant and kept FID in a much lower range, even at higher negative guidance scales, while still effectively filtering out unwanted content. This dynamic strategy effectively pushes the entire safety-alignment frontier outward, offering a superior balance.
Why Dynamic Prompting is More Effective
The paper highlights two main reasons for VL-DNP’s superior performance:
1. Concept-Specific Negatives: The VLM can propose very specific negative prompts (e.g., “Male breast,” “Buttocks”) instead of generic terms like “nsfw.” This reduces unnecessary suppression of other image elements.
2. Evolving Prompts: As an unwanted artifact disappears from the image during denoising, the VLM can drop that irrelevant negative prompt and target new risks, preventing over-suppression and allowing the image to develop naturally.
Also Read:
- Guiding Text-to-Image Models Towards Safer Content Without Retraining
- Improving Text-to-Image Models with Adaptive Text Embeddings
Future Directions
While VL-DNP offers a significant leap forward, the researchers acknowledge some limitations. The dynamic prompting introduces additional latency due to the VLM queries. Future work could explore mitigating this overhead by caching intermediate predictions, querying the VLM less frequently, or utilizing more lightweight VLMs for real-time deployment. Further research will also focus on jointly scheduling both the strength and content of the guidance for even more optimized results.
For more in-depth technical details, you can read the full research paper here: Dynamic VLM-Guided Negative Prompting for Diffusion Models.
