Loop Forge
Loop Forge / Projects

23 September 2026

Qwen-Image 2.1, run locally

Qwen-Image 2.1 generates and edits with one 7 billion parameter model, and its autoencoder carries an alpha channel, so transparency comes out of the model rather than a background remover. Everything below ran locally on an RTX 3060, with every prompt exact and every timing measured.

Watch on YouTube

26 prompts

Transparent backgrounds

Eight objects on one alpha channel

The phrasing that cuts cleanly

Editing by circling

Two circles, two changes

A cap and a backpack, in a room full of people

One cup, with its twin left alone

One scene, four edits

The plate

The pail extracted

The pail removed

Put back on the far side

Multiple references

Nine people in one pass

Four panels, two characters

A character turnaround sheet

Twelve labelled expressions

Three garments onto one person

Panoramic generation

A selfie turned into a full 360

A 360 from text alone

Text rendering

A magazine cover over a photograph it did not generate

Seven strings, all correct

Ten strings in Chinese

Resolution and cost

The same prompt at 1.0 MP

The same prompt at 2.0 MP

The same prompt at 3.0 MP

The same prompt at 4.3 MP

Where it struggles

The phrasing that redraws the subject

Three leaves in a canopy

One person in a crowd

The model, and the machine

Transformer
7B single-stream DiT, 32 layers
Text encoder
Qwen3-VL 8B
VAE
64 channel RGBA, 16x spatial compression
Native output
2K, about 4.2 megapixels
Peak VRAM
10.5 to 11.7 GiB at int8
Licence
Qwen Research, non-commercial
Ran on
RTX 3060, 12GB (local)

Materials and links