# Flow Upscaler

## Summary

Fast latent upscaler for the `FLUX.2` family. It takes existing FLUX latents and upscales them in latent space, so you can push images much larger without relying only on classic pixel upscalers.

## Model Family

- `flux`

## Status

- `imported-only`

## Source

- Original page: <https://github.com/tensorforger/comfyui-flow-upscaler>
- Workflow file: <https://raw.githubusercontent.com/tensorforger/comfyui-flow-upscaler/main/workflow.json>
- Original author: `TensorForger`
- Date imported: `2026-08-08`

## Developer notes

- The author positions this as a fast latent upscaler trained in `Flux.2` latent space.
- The model is trained for `2x` upscaling, but multiple passes can be chained for much larger outputs.
- The README recommends the tiny `Flux.2` VAE (`taef2`) for `2048x2048` and above because the original VAE becomes impractical at high resolutions.
- The workflow included in the repo is the demo path the author expects people to load first.

## Our notes

- This is one of the most practically useful imports here because it solves a real workflow problem instead of just offering another generation variant.
- It looks especially relevant if you end up using `FLUX.2-klein` or other lighter FLUX-family generation paths and want cleaner high-res finishing.
- It is more specialized than a generic Ultimate SD Upscale graph, which is a good thing if you want something clearly tied to the FLUX latent stack.

## Required custom nodes

- `comfyui-flow-upscaler`
- `rgthree-comfy` for `Image Comparer`

## Built-in / core nodes used

- `Flux2Scheduler`
- `EmptyFlux2LatentImage`
- standard CLIP / UNET / VAE loaders
- sampler selection and advanced sampler nodes

## Notes

- Model weights are separate from this workflow and need to be downloaded from the source repo's linked Hugging Face page.
- Best use: upscale a good FLUX-family image after the composition is already right.