Understanding the Impact of Negative Prompts: When and How Do They Take Effect?


📝 - Synthical

The Dynamics of Negative Prompts in AI: A Comprehensive Study by: Yuanhao Ban UCLA, Ruochen Wang UCLA, Tianyi Zhou UMD, Minhao Cheng PSU, Boqing Gong, Cho-Jui Hsieh UCLAE

This study addresses the gap in understanding the impact of negative prompts in AI diffusion models. By focusing on the dynamics of diffusion steps, the research aims to answer the question: "When and how do negative prompts take effect?". The investigation categorizes the mechanism of negative prompts into two primary tasks: noun-based removal and adjective-based alteration.

The role of prompts in AI diffusion models is crucial for guiding the generation process. Negative prompts, which instruct the model to avoid generating certain features, have been less studied compared to their positive counterparts. This study provides a detailed analysis of negative prompts, identifying the critical steps at which they begin to influence the image generation process.


Critical Steps for Negative Prompts

Noun-Based Removal: The influence of noun-based negative prompts peaks at the 5th diffusion step. At this critical step, negative prompts initially generate a target object at a specific location within the image. This neutralizes the positive noise through a subtractive process, effectively erasing the object. However, introducing a negative prompt in the early stages paradoxically results in the generation of the specified object. Therefore, the optimal timing for introducing these prompts is after the critical step.

Adjective-Based Alteration: The influence of adjective-based negative prompts peaks around the 10th diffusion step. During the initial stages, the absence of the object leads to a subdued response. Between the 5th and 10th steps, as the object becomes clearer, the negative prompt accurately focuses on the intended area and maintains its influence.

Cross-Attention Dynamics

At the peak around the 5th step for noun-based prompts, the negative prompt attempts to generate objects in the middle of the image, regardless of the positive prompt's context. As this process approaches its peak, the negative prompt begins to assimilate layout cues from its positive counterpart, trying to remove the object. This represents the zenith of its influence.

For adjective-based prompts, during the peak around the 10th step, the negative prompt maintains its influence on the intended area, accurately targeting the object as it becomes clear.

The study highlights the paradoxical effect of introducing negative prompts in the early stages of diffusion, leading to the unintended generation of the specified object. This finding suggests that the timing of negative prompt introduction is crucial for achieving the desired outcome.

Reverse Activation Phenomenon

A significant phenomenon observed in the study is Reverse Activation. This occurs when a negative prompt, introduced early in the diffusion process, unexpectedly leads to the generation of the specified object within the context of that negative prompt. To explain this, researchers borrowed the concept of the energy function from Energy-Based Models to represent data distribution.

Real-world distributions often feature elements like clear blue skies or uniform backgrounds, alongside distinct objects such as the Eiffel Tower. These elements typically possess low energy scores, making the model inclined to generate them. The energy function is designed to assign lower energy levels to more 'likely' or 'natural' images according to the model’s training data, and higher energy levels to less likely ones.

A positive difference indicates that the presence of the negative prompt effectively induces the inclusion of this component in the positive noise. The presence of a negative prompt promotes the formation of the object within the positive noise. Without the negative prompt, implicit guidance is insufficient to generate the intended object. The application of a negative prompt intensifies the distribution guidance towards the object, preventing it from materializing.

As a result, negative prompts typically do not attend to the correct place until step 5, well after the application of positive prompts. The use of negative prompts in the initial steps can significantly skew the diffusion process, potentially altering the background.


  1. Do not step less than 10th times, going beyond 25th times does not make the difference for negative prompting.

  2. Negative prompts could enhance your positive prompts, depending on how well the model and LoRA have learn their keywords, so they could be understood as an extension of their counterparts.

  3. Weighting-up negative keywords may cause reverse activation, breaking up your image, try keeping the ratio influence of all your LoRAs and models equals.