Neural Style Transfer at Production Resolution: What Actually Works

Neural style transfer got a lot of attention when it first appeared — and then mostly disappeared from serious VFX conversation because it did not work at production resolution or speed. We revisited it with modern diffusion-based approaches and found the picture has changed significantly.

The resolution problem

Global style models trained at 512px learn to redistribute energy across frequency bands in ways that look plausible at small scale and catastrophic at large scale. We solved this with a tiling approach using overlapping context windows, combined with a face-preservation network that locks high-frequency detail in detected face regions before the style pass runs.

The temporal coherence problem

Frame-by-frame style transfer produces temporal noise immediately visible on motion. We addressed this with optical-flow-guided temporal smoothing applied to the style weight maps rather than output pixels — significantly more stable than post-hoc smoothing without introducing blur from mixing frames.