FLUX 3 Image Klein Upscale | SeedVR2 + Restore LoRA

Load a picture, pick what kind it is, and run. FLUX 3 Image Klein redraws it sharp, one tile at a time.

How to use it

1. Load your image. Optional: crop it with the blue handles.

2. Set Megapixels: the size of the result. 4 is a good start.

3. Pick what you have: Photo, Blurry photo or Generated picture.

4. Run. Slide Before and after to compare. The result is saved as a PNG.

What to pick

• Photo, the default: a small, old or noisy photo. SeedVR2 3B cleans it up and enlarges it first, then FLUX 3 redraws it in tiles of about 2 megapixels. It stays close to your picture.

• Blurry photo: out of focus, or blown up from a small picture. FLUX 3 draws the whole picture again first, so it gets much more detail, but faces can change.

• Generated picture: one an AI model made. 4x-UltraSharpV2 enlarges it first, then the tiles.

• Shrink first, for Photo: turn it on for a big file that is blurry. The picture is made smaller first, so its blur becomes small enough to clean up. Leave it off for a sharp photo. Blurry photo does this by itself.

• Describe your picture is optional: a few words about what is in it, like "a woman with black bangs and dark eye makeup".

The restore LoRA is optional. Without it the workflow still runs. With it, faces come out a little closer to your picture: on my examples, a face-recognition check matched the face to the original more closely on 7 of 8 faces, and the colours were closer on 8 of 8. The difference is small.

How I picked it

I tried about 40 ways to upscale with FLUX 3 Image Klein and scored the results blind. SeedVR2 first, then FLUX 3 tiles, was my pick for small photos. On blurry photos, drawing the whole picture again looked great most often.

What you need

• ComfyUI 0.41.0 or newer. That is the first version with FLUX 3.

• AusBoss nodes 2.13.0 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.

• The FLUX 3 Image Klein files. They are on Hugging Face behind a license click: open black-forest-labs/flux-3-image-klein-comfy, sign in and accept, then these links work.

  • flux3-image-klein-int8-convrot.safetensors → models/diffusion_models · 5.83 GB

  • qwen3vl_4b_int8_convrot.safetensors → models/text_encoders · 4.86 GB

  • flux2-vae.safetensors → models/vae · 336 MB

• Optional, my restore LoRA: flux-3-image-klein-restore.safetensors → models/loras · 161 MB. Without the file the workflow skips it and runs.

• For Photo, SeedVR2 3B:

  • seedvr2_3b_int8_convrot.safetensors → models/diffusion_models · 3.46 GB

  • seedvr2_ema_vae_fp16.safetensors → models/vae · 501 MB

• For Generated picture: 4x-UltraSharpV2.safetensors → models/upscale_models · 140 MB

Tips

• A bigger file is not a sharper picture. A big photo that looks soft was often blown up from a small one. Turn on Shrink first, or pick Blurry photo.

• A face came back different? Photo stays closer to your picture than Blurry photo. Or run it again: the seed changes each run.

• Stubble on a woman's face? Say who is in the picture in Describe your picture, like "a young woman". It can help.

• On an RTX 5090, a 4 megapixel result takes about 13 seconds as a Photo, 18 as a Blurry photo and 24 as a Generated picture, once the models are loaded.

What it does not do

• It does not bring back what the camera saw. It is a redraw: fine detail is made up to fit, and a face stays the same person, not the same pixels.

• A very noisy old webcam photo came out bad in every way I tried with FLUX 3.

• Small faces and very blurry faces can change, most of all with Blurry photo.

• Shadows on a face can come back as stubble.

• Everything was measured on one RTX 5090 with one seed per picture.

About the examples: they are my own pictures. I made each photo small and saved it as a low-quality JPEG first, the way web photos usually are. The blurry ones were made small and blown back up. The workflow got only that file. None of them is in the restore LoRA's training set.

The restore LoRA, its test numbers and its license are on the Hugging Face repo.

On Qwen Image 2.1 instead: Qwen Image 2.1 Tiled Upscale.

Credits

FLUX 3 Image Klein is by Black Forest Labs, under the FLUX Kommunity License: use of the pictures you make is covered by its rules for outputs (section 2), so read it if you plan to use them commercially. The restore LoRA is mine, a modified version of FLUX 3 Image Klein shared under the same license. SeedVR2 is by ByteDance Seed, under the Apache 2.0 license; the ComfyUI files are Comfy-Org's. 4x-UltraSharpV2 is by Kim2091, under CC BY-NC-SA 4.0: non-commercial use only.

v1.0 (2026-10-17): first release.

Questions or bugs: GitHub issues or the comments here. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.