Comfyui-Flow-Upscaler Supercharges Flux.2 Upscaling To 8K In 25 Seconds

The Comfyui-flow-upscaler node set brings a single-step latent upscaling method to ComfyUI, using a compact rectified flow model to enlarge Flux.2 images. It can upscale a 512x512 image to 1024x1024 in just 8 milliseconds inside the latent space. The entire pipeline, including generation, cascade upscaling to 8K, and decoding, finishes in about 25 seconds on an RTX 5090 GPU.

Developer TensorForger created this open-source tool to solve a speed problem in Flux-based image workflows. The project started as part of a real-time Flux pipeline called FluxRT, where low latency was critical. After seeing how well flow distillation worked, TensorForger released these dedicated ComfyUI nodes for the community.

Single-step upscaling trained with flux.2 distillation

Artists and developers who frequently upscale Flux.2 generations will benefit from the dramatically shorter wait times. The tool handles images up to 8K without relying on attention layers, so it scales smoothly on hardware with limited memory. Quick turnaround makes it suitable for iterative design work and batch processing.

Developer notes about the training strategy

The upscaler was not trained from scratch to do rectified flow. Instead, it was distilled from the much larger Flux.2-klein-4B teacher using 20,000 diverse images, so the small model learns to mimic strong image semantics. Direct latent-space downscaling was avoided during conditioning preparation because it introduced artifacts, so the team downscaled in pixel space before re-encoding. The U‑Net architecture uses SDXL-style ResNet blocks with FiLM conditioning and skips attention entirely, which keeps compute demands linear and makes 8K generation practical.

"It is a bit different from other latent upscalers because the upscaler itself is a small (59M params) rectified flow model that samples upscaled latents in just one denoising step." — Source: Reddit