Remote VAEs for decoding with Inference Endpoints 🤗
Hugging Face Blog
Hugging Face's Diffusers library now supports remote VAE decoding, allowing users to offload the memory-intensive decoding process to cloud endpoints instead of running it locally on consumer GPUs. On a 1024x1024 resolution with Stable Diffusion v1.5, decoding takes 0.148 seconds on an RTX 4090 versus 0.301 seconds when using local tiled decoding, demonstrating speed advantages alongside reduced local memory consumption. This experimental feature eliminates the need for device transfer overhead or quality-degrading tiling, enabling users to run high-resolution image and video synthesis on GPUs with limited VRAM while maintaining inference speed.
