Optimizing Diffusion Models with Emotional Vocabulary Prompts

Описание и применение

Исследование показывает, что использование эмоциональной лексики в промптах для модели LTX-2 позволяет сократить количество шагов диффузии на 20% без потери качества изображения. Это снижает затраты на вычисления и электроэнергию.

Задача

Standard literal prompts in diffusion models require more compute steps (e.g., 30 steps) to achieve high perceptual quality, increasing GPU time, electricity costs, and latency.

Результат

Replacing literal prompt descriptions with emotional vocabulary reduces diffusion steps by 20% without compromising visual quality, leading to significant compute cost savings.

Рабочий процесс

• Identify literal motion descriptors in the prompt (e.g., 'head tilt', 'blinks', 'smiles').
• Use the Neuromorphic Prompt Translator (or manual mapping) to convert literal descriptors into emotional/narrative equivalents (e.g., 'head tilt' -> 'contemplative pause').
• Configure the diffusion model (LTX-2) with the new emotional prompt.
• Reduce the number of diffusion steps by approximately 20% (e.g., from 30 to 24).
• Validate output quality using LPIPS against a high-step reference to ensure no perceptible loss.
• Iterate on prompt vocabulary if quality drops below acceptable thresholds.

Доказательства

emotional vocabulary in prompts achieves the same perceptual quality as literal descriptions using 20% fewer diffusion steps.

Источник

LPIPS = 0.011 ± 0.005, n=15, p < 10⁻¹⁹

Источник