Negative Prompt

A negative prompt operates by leveraging classifier-free guidance (CFG). The model calculates two predictions: one conditioned on the positive prompt and one conditioned on an empty or null prompt (unconditional). The final denoising direction is steered away from the unconditional prediction and amplified toward the conditional one. When a negative prompt is provided, it replaces the null condition, effectively defining what to move away from. The guidance scale controls the strength of this steering effect.

• Core Process: final_prediction = conditional_prediction + guidance_scale * (conditional_prediction - negative_prediction)
• Result: The sampling trajectory is pushed away from regions of the data distribution associated with the negative description.

Negative prompts are used to suppress unwanted artifacts, styles, or content that the model might otherwise default to or hallucinate.

• Removing Artifacts: ugly, deformed, disfigured, poor details, bad anatomy
• Enhancing Aesthetic Quality: blurry, grainy, noisy, oversaturated, watermark, signature
• Controlling Style: photorealistic (when aiming for painting), 3d render, cartoon (when aiming for photo)
• Excluding Content: text, people, cars, buildings
• Correcting Composition: extra limbs, extra fingers, mutated hands, poorly drawn face

Example Prompt: "A majestic eagle in flight, photorealistic, detailed plumage" Negative Prompt: "blurry, cartoon, watermark, text, deformed beak, extra wings"

The guidance scale (often denoted as cfg_scale) is a critical hyperparameter that determines how strongly the model adheres to all prompts, both positive and negative.

• Low Guidance Scale (e.g., 1-5): Results in more diverse but potentially less coherent images. The influence of the negative prompt is minimal.
• Optimal Range (e.g., 7-12): Provides a strong balance. The negative prompt effectively suppresses unwanted elements while maintaining natural image variation.
• Very High Guidance Scale (e.g., >15): Can lead to over-saturated, hyper-contrasted, or "overcooked" images. The model becomes overly rigid, sometimes introducing new artifacts as it strenuously avoids the negative concepts. Tuning this value is essential for effective negative prompting.

Advanced syntax allows for fine-grained control over the emphasis of different concepts within a negative prompt.

• Basic Weighting: (ugly:1.3), (deformed:0.8) increases emphasis on "ugly" and decreases it on "deformed".
• Blended Prompts: Concepts can be combined, e.g., blue [red:0.5] tries to avoid a blend of blue and red.
• Attention/Emphasis: Using parentheses () increases attention, square brackets [] decreases it. ((ugly)) applies more negative weight than ugly.
• AND & BREAK Operators: Some implementations support AND to group concepts (ugly AND deformed) and BREAK to separate conceptual groups, providing more discrete control over the negative conditioning.

Negative prompts are powerful but not a panacea and come with specific limitations.

• The Curse of Negation: Overly long or aggressive negative prompts can constrain the model too much, leading to bland, low-diversity outputs or even causing it to fail to generate coherent images.
• Semantic Bleed: Negating a broad concept (e.g., person) can inadvertently remove related desirable elements (e.g., the ambiance of a populated scene).
• Model Dependency: Effectiveness varies significantly between different base models (e.g., SDXL vs. SD 1.5) and fine-tuned checkpoints, as their latent understanding of concepts differs.
• Not a Precision Tool: It is better at suppressing general concepts (blurry) than executing precise spatial edits ("remove the third tree from the left").

An advanced technique that aims to more precisely isolate and remove unwanted concepts without affecting unrelated aspects of the image. Instead of using the standard CFG formula, it projects the negative conditioning vector to be orthogonal (perpendicular) to the positive conditioning direction in the model's latent space.

• Goal: Cancel out only the components of the noise prediction aligned with the negative concept, leaving other creative directions untouched.
• Benefit: Can theoretically provide more surgical removal of attributes (e.g., "watercolor style") while preserving the core content and composition specified by the positive prompt.
• Status: Largely an experimental/research-level technique, not yet standard in mainstream consumer interfaces, but represents the frontier of controlled negation.

The core technique that enables negative prompting. CFG steers the diffusion sampling process by calculating a weighted combination of a conditional and an unconditional model prediction.

• Mechanism: At each denoising step, the model computes both a prediction guided by the text prompt and a prediction for an empty or generic prompt. The final output is pushed further in the direction of the conditional prediction and away from the unconditional one.
• Guidance Scale: A scalar hyperparameter (e.g., 7.5) that controls the strength of this steering. Higher values increase adherence to the positive prompt and the effect of the negative prompt but can reduce sample diversity and sometimes introduce artifacts.

The overarching paradigm of controlling a generative model's output via explicit input signals. A negative prompt is a form of conditional generation.

• Positive Condition: The primary input (e.g., a text prompt) specifying desired attributes.
• Negative Condition: An auxiliary input specifying attributes to avoid or suppress.
• Modalities: Conditions can be text, images (for img2img), segmentation maps, sketches, or class labels. The model learns to map these conditions to specific regions or features in the output data distribution.

The component that transforms natural language prompts (both positive and negative) into a format the diffusion model can use. Its quality is critical for prompt understanding.

• Function: Converts prompt text into a sequence of dense vector representations (embeddings).
• Common Model: CLIP's text encoder is standard in models like Stable Diffusion. It projects the prompt into the same semantic space as images.
• Process: The positive and negative prompts are encoded separately. Their embeddings are used to compute the conditional and unconditional predictions during CFG.

The neural network mechanism that allows the diffusion model's U-Net or DiT to "attend to" the text embeddings during the denoising process.

• Role: Injects conditional information from the text encoder into the visual feature maps of the generative model.
• For Negative Prompts: The model computes cross-attention for both the positive and negative text embeddings. The guidance process then uses the difference between these attended representations to move the image away from concepts activated by the negative prompt.

A critical hyperparameter that acts as a dial controlling the intensity of classifier-free guidance, directly amplifying the effect of both positive and negative prompts.

• Trade-off: Governs the fidelity-diversity trade-off.
  • Low Scale (e.g., 1-3): Output is more diverse and creative but may ignore the prompt.
  • High Scale (e.g., 10-20): Output closely follows the prompt but becomes less diverse and can exhibit saturated colors or unnatural textures.
• Practical Use: A typical starting point for text-to-image is 7.5. The negative prompt's effectiveness is proportional to this scale.

The algorithm that performs the iterative denoising steps. The choice of sampler can influence how effectively negative prompts are applied.

• Deterministic vs. Stochastic:
  • DDIM: A deterministic sampler. Negative prompts have a consistent, predictable effect.
  • Ancestral Samplers (e.g., Euler a): Introduce new noise each step. This can make the effect of a negative prompt slightly less predictable but is often associated with more creative outputs.
• Step Count: More denoising steps generally allow for finer-grained application of guidance, letting the model more gradually move away from negative concepts.