Study
Generative AI · Text-to-image
How text steers a diffusion model: conditioning, classifier-free guidance, and the guidance scale trade-off, shown with a small conditional diffusion model on 2-D data where "prompts" are cluster labels.
- Classifier-free guidance runs the model with and without the prompt and pushes the result further toward the prompted direction.
- In my toy, a moderate scale was enough; pushing guidance much higher distorted the samples instead of improving them.