Render farm is a bunch of networked servers dedicated to render image in a distributed or parallel fashion. The capacity of in-house render farm has an upper bound but a rendering workload in real world tends to highly fluctuate. Thus execution of overloaded jobs on a remote cloud in a seamless way is very attractive to build cost-effective rendering infrastructures. However, extending a render farm from a cluster to a public cloud is not straightforward. The most critical obstacle is data synchronization between render farms. A simple approach to overcome this issue is that rendering data are uploaded to or downloaded from public cloud. This kind of configuration is used by most cloud-based rendering services. Despite its simplicity, data synchronization by an explicit data transferring imposes a heavy burden on a user and incurs unnecessary data transmission as well as startup latency. A more elaborated data synchronization technique armed with an on-demand fetch and cooperative caching can ease the burden and improve a rendering throughput by exploiting rendering workload characteristics. Our proposed revision-aware caching is that rendering data identifiable with additional revisions are cached on cloud and are fetched from on-premises revision-aware repository or a cache pool of peer cloud node. We proved the performance of our scheme via experiments based on synthetic workload and real workload.