Syncing generative visuals to BPM without losing the "soul"

Been experimenting with Runway and Stable Video Diffusion for my latest ambient track, but hitting a wall with timing. The AI generates beautiful abstract flows that match the mood, but when I cut to beat markers, everything feels either too mechanical or completely chaotic. Tried analyzing audio spectrums and keyframing opacity based on frequency ranges, which helps, but there's this weird middle ground where the footage wants to breathe slowly while the percussion wants sharp cuts. How are you handling the marriage between generative randomness and musical structure? Do you let the AI lead and compose music around the visuals instead? Are you rendering at 24fps and retiming in post, or generating at exact BPM tempo from the start? Does anyone else feel like over-editing kills that ethereal quality AI footage naturally has? This feels like uncharted territory.