Paired videos are too expensive—C2R# uses image pairs to synthesize pseudo data
Converting rough 3D and anime videos into live-action texture usually requires a large amount of paired video. C2R# instead uses image pairs to synthesize pseudo-paired data, and has already been used at Roblox RDC. In the comments, some question the quality of the pseudo data, while others ask for open source.
What C2R# aims to solve is the high cost of paired videos. C2R, released last year, used a video generation model to turn rough 3D blockouts into photorealistic videos. When put into a game engine, the problem becomes stability and controllability. Yi Zhou mentions in the tweet that even models like Seedance 2.5 often require multiple attempts. Real applications need a large number of paired coarse-model videos and real videos, and collection costs are high.
C2R#'s approach is to change the data entry point. It first trains a lightweight Real-to-Coarse model using only image pairs. Applying it frame by frame to real videos synthesizes pseudo-coarse-model videos, yielding a large number of pseudo-paired videos. These data are then used to train a large Coarse-to-Real video generation model. The entire pipeline does not require real paired coarse-model videos and real videos.
This technology has already been used at Roblox RDC. It can also handle anime-style videos, provided there are 10K image pairs.
The video shows the comparison directly: on the left are Game Input and Game C2R# Result; on the right are Anime Input and Anime C2R# Result. Game scenes, indoor dances, and animated characters are all pushed toward a more live-action look.
Project page: https://gonzalognogales.github.io/coarse2real-sharp/
Paper: https://arxiv.org/abs/2610.28954
Original tweet: https://x.com/Papagina_Yi/status/2116710992155717699
In the comments, some users argue that if the real images in the dataset only use real images and synthetic data is only used for pairing, quality would be better. This question points directly at the core of the method: whether the distribution of pseudo-paired videos can cover the gap between real coarse models and real videos. Others ask when the model or data will be released; the original post gives no timeline.
C2R#'s highlight is its data production pipeline. Using image pairs to create video training data changes the cost structure. The risk lies here too: once the pseudo-paired distribution is biased, the model will learn a bias. Beyond the results, that comment about data quality is worth continuing to watch.
发布时间: 2026-11-01 17:29