A video renderer follows new scene formats in four steps

When a scene’s layout, motion, and camera are already specified, can one renderer turn different kinds of structural input into realistic video without a separate model for each format? Structure alone does not supply the appearance that makes the output look like a finished video.

D SolarRenderer adapts a frozen video model to follow varied structural inputs, then trains a separate adapter to render in four steps.

D SolarRenderer adapts a frozen video model so it can turn structural videos into realistic footage, then trains a separate adapter for four-step rendering.

To test how well it follows structure, Gemini 3.8 Flash compares pairs of renderings on visual quality, structural adherence, first-frame preservation, and overall preference.

On the generalization test, D SolarRenderer ties LynnReal-Omni in overall Elo rating and rates higher for structural adherence. On familiar formats, its overall and structure ratings are higher on the in-distribution test.

These aggregate pairwise judgments support the conclusion that the renderer can follow varied structure at four steps, including tested formats absent from training.

The generalization test covers 208 cases drawn from 57 structural videos and has no ground-truth videos for direct comparison. On some individual unfamiliar formats, the 50-step version was preferred over the four-step model.