A LoRA that softens the harsh upscale Qwen Image 2.1 does by itself: the result is a little softer, still detailed, makes up less, and stays true to your picture.

Qwen Image 2.1 can upscale a small or worn picture without this LoRA, but it overdoes it. Skin turns gritty, pores and freckles show up that were not there, faces look older, and the picture shifts a few pixels. With this LoRA the result stays on your picture.

Blurry photos need a lower strength. If your picture is out of focus, shaky, or a frame from a video, set the LoRA to 0.5. At 1.0 it stays true to the blur and the result comes back soft. A small or compressed picture that is in focus is fine at 1.0.

No trigger word. Start at strength 1.0, enlarge your picture to about 2 megapixels, give it as image 1, and ask:

Restore the photo: remove blur, noise and compression artifacts and recover sharp, natural detail, keeping the same composition, people and colors.

For scratched or faded prints:

Restore the old photo: remove scratches, dust, fading and grain and recover sharp, natural detail and color, keeping the same composition and people.

On an old print it takes out scratches, dust and grain and brings back most of the color.

Why use it

Measured on 37 small pictures it never trained on, each compared with the full-size picture it was shrunk from. Same seed and request, 8 steps with Viggle Turbo:

• The picture stays in place. Without the LoRA the result sits more than 2 px off. With it: about 0.3 px.

• The face stays the same person. Face match 0.79 without, 0.88 with (on 29 of them; closer on every one).

• The colors stay. Color and brightness shift 1.9 % without, 0.9 % with.

• It makes up less. Plain Qwen draws more grain than the full-size picture has. The LoRA draws a little less than it has, so results are slightly soft.

The pictures on this page show it: the small input, the result without the LoRA (amber) and with it (teal), and a section enlarged underneath.

How to use it

The workflow on this page has it set up: load your image and press Run. One pass at about 2 megapixels, with a Fast switch (8 steps with Viggle Turbo, or 25 steps without it). It is the second file of this version, and it is inside every still picture here: drop one on ComfyUI. It uses my AusBoss nodes.

For a result over 2 megapixels, use my Qwen Image 2.1 Tiled Upscale workflow. It redraws the picture in tiles with this LoRA.

In your own Qwen Image 2.1 edit graph:

1. A LoRA loader right after the model loader, strength 1.0 to start. Strength works as a dial: 0.5 for a blurry photo, 0.75 for a little more of Qwen's own texture, 1.0 to stay truest to your picture.

2. Resize your picture to about 2 MP, both sides divisible by 32.

3. Text Encode Qwen Image 2.1: the resized picture as image_1, resolution 0, the request as the prompt.

4. KSampler on the encoder's latent output: CFG 1, euler / simple, denoise 1. 8 steps with Viggle Turbo at 1.0, or 25 steps without it.

5. VAE Decode, then Split Image with Alpha (the Qwen 2.1 VAE decodes RGBA).

Limits

• A blurry photo comes back soft at 1.0. Lower the strength to 0.5. On 26 real low-resolution photos, the 6 blurriest needed that.

• It is a redraw, not a recovery. Fine detail is made up to fit, and a face stays the same person, not the same pixels.

• Plain Qwen 2.1 draws more texture. This LoRA trades some of that for staying true to the picture.

• Small faces are redrawn. A face about 100 px tall in your picture can come back as a slightly different person.

• Trained at 1 to 2 megapixels. Keep each pass at about 2 MP; for a bigger result, redraw the picture in tiles (my Qwen Image 2.1 Tiled Upscale workflow does that).

• Small lettering can come back misspelled.

• One pass is enough. A second pass over the result at the same size came back crunchy.

• Don't stack it with my Consistency LoRA.

Training

• ai-toolkit (qwen_image_2) on the Comfy-Org INT8 convrot base. Rank 32, alpha 32, AdamW8bit, learning rate 1e-4, batch 1, 1,500 steps.

• 1,209 pairs: a clean picture, and the same picture damaged the way pictures get damaged (shrunk for the web, saved and re-saved as JPEG, blur and noise, phone smoothing, old prints, video stills).

The full numbers, the step 1,250 file and the notes on how it was made are on the Hugging Face model card.

Qwen Image 2.1 is by Alibaba Qwen, under the Qwen Research License: non-commercial use only.

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