What's this about?
Qwen released a weight-available image model whose 7B visual component handles generation, editing, native transparency, local masks and up to ten references in one checkpoint.
What exactly is native transparency?
The model generates an alpha channel with RGB pixels, so a subject can arrive ready for layered design rather than relying on background removal. Good examples do not establish reliability on arbitrary prompts.
What's the community saying?
HN 519 points, 157 comments at fetch time.
- Users value 7B size, RGBA output and strong small-text rendering.
- Several call the research-only commercial restriction an “open-source” licence trap.
- Others ask for failure cases rather than curated vendor examples.