A new set of Flow Upscaler nodes for ComfyUI has been introduced, providing extremely fast image upscaling within the Flux.2 model's latent space. By utilizing a lightweight Rectified Flow architecture with 59 million parameters, the resolution enhancement process is completed in only one denoising step.
What Happened
Developers have released the Flow Upscaler ComfyUI nodes, optimized for working with Flux.2. The 59-million-parameter model allows upscaling from 512x512 to 1024x1024 resolution in just 8 ms on an RTX 5090 GPU. A full generation and subsequent upscale to 8K resolution takes approximately 25 seconds.
Context
The technology is based on the Flow Distillation method, which allows reducing the upscaling process to a single denoising step without significant loss of quality. The architecture is specifically optimized for the Flux.2 latent space, ensuring seamless integration into ComfyUI workflows.
Why It Matters for the Industry
The emergence of such compact and fast models paves the way for high-performance, real-time generative content creation. This significantly lowers computational resource requirements and allows for a reduction in TCO (Total Cost of Ownership) when scaling production services that utilize image generation.
Why It Matters for Users
ComfyUI users can now almost instantaneously preview compositions in high resolution and create ultra-high-quality images (up to 8K) locally on consumer hardware, without needing to rely on powerful GPU farms.
Sources
Author
Look at AI, Editorial Team