Qwen-Image-2.1 generates transparency natively with a 4-channel VAE

Qwen-Image-2.1 arrived on September 20, 2026, and the release materials suggest the interesting part is not a new benchmark score but where transparency sits in the model. Its VAE config lists "in_channels": 4 and "out_channels": 4, and its latent representation is 64 channels with 16× spatial compression. The body also describes a Mixed-Granularity Attention design in which text and condition-image tokens take a fixed t=0 while the generation target uses the current timestep, which is what makes it possible to cache reference-side computation across the default 40 inference steps.

The article frames this against a different tradition of transparency work. A mask-estimation model such as BiRefNet estimates a foreground mask and applies it to the original pixels with putalpha, so the decoded RGB survives. A 4-channel VAE instead generates color and alpha together, which is useful when the asset does not exist yet but is not the same as faithfully preserving an existing photo. The author also notes a limit that applies to mask work too: photographs can already contain background color mixed into the outline, so preserving RGB exactly and removing color bleed at the edges cannot both be demanded unconditionally.

Whether this reshapes asset pipelines is likely to hinge on licensing and memory rather than image quality. The released weights fall under the Qwen Research License, with commercial use requiring a separate license, and the 7B figure applies to the generation transformer alone, since a Qwen3-VL 8B text encoder and the VAE load alongside it. Teams weighing a switch would need to judge peak memory at their target resolution, and the body cautions that no speed multiple from the caching design can be stated without measurement.