The prevailing wisdom in AI image generation suggests that the 'negative prompt' field is the primary lever for controlling output artifacts. However, a new technical deep-dive published on DEV.to by Aveed Walikhan challenges this assumption, demonstrating that positive description is a superior mechanism for most tasks on PixAI's latest anime model, Tsubaki.3. The analysis breaks down three distinct control leversβ€”dedicated negative fields, in-prompt exclusions, and positive descriptionsβ€”and tests them against specific generation tasks.

The Three Control Levers

The study isolates three mechanisms available to users: Lever 1 is the standard negative field, often pre-populated with terms like 'lowres' or 'bad anatomy'. Lever 2 involves writing exclusions directly into the main prompt body, such as 'no props on the face'. Lever 3 relies on positive description, where the user specifies the desired state (e.g., 'a plain flat pastel lavender background') rather than negating the unwanted one. All tests were conducted on Tsubaki.3 with the Prompt Helper disabled to ensure exact parsing of the user's input.

Negative Fields Fail on Surface Details

When testing against text appearing on surfaces, such as signs in a ramen shop scene, the negative field proved ineffective. Adding terms like 'text, letters, kanji, writing, signage, logo' to the negative field did not prevent Japanese characters from appearing on the cloth curtain. The model ignored the negation, though it did improve steam effects. This indicates that while negative fields can handle diffuse scene content, they struggle with specific surface-level details where the model associates text with the object's style.

Positive Descriptions Dominate the Matrix

Lever 3, the positive description, consistently outperformed or matched the other levers. By describing lanterns as 'plain red paper lanterns with no markings' and the curtain as 'solid dark blue cloth with no pattern', the model successfully eliminated all text from the surfaces. This method worked because it left the model with nothing to render on those specific objects. In contrast, the negative field attempted to suppress an output, while the positive description forced the rendering of a different, clean object.

The Limit of Prompt Engineering

Not all artifacts are controllable via prompt wording. In a test involving cyberpunk-style character portraits, mechanical pieces on the character's cheeks persisted across all three levers. Whether using the negative field, an in-prompt exclusion, or a positive background description, the 'style-bound' details remained. This suggests that certain model biases are baked into the style representation and cannot be overridden by simple instruction changes. For these cases, the author recommends using an image edit tool to modify specific attributes on an already approved image.

Key Takeaways

  • Positive descriptions are more reliable than negative prompts for removing text or clutter from specific surfaces.
  • The negative field remains useful for broad scene cleanup, such as removing background buildings or crowds.
  • Style-bound artifacts, like cyberpunk cheek pieces, often resist all prompt-based suppression methods.
  • Image editing tools are the only reliable mechanism for fixing single attributes on an otherwise perfect generation.

The Bottom Line

Stop treating the negative prompt box as a magic eraser; it is a blunt instrument for diffuse noise, while positive description is the precision tool for structural control.