Guidance scale controls how strongly the denoiser is pushed toward the conditional prediction instead of the unconditional one. Low scale lets the latent drift away from the prompt but keeps motion loose and natural. High scale forces the latent hard toward the prompt, which fixes alignment but over-saturates colors, flattens texture, and freezes motion because the model keeps collapsing toward the same strong conditional direction. There is no setting that maximizes both; the scale is a trade-off between prompt adherence and motion naturalness.
How Text-to-Video AI Works Under the Hood