Reproduce “AlphaAnything: Extending Pretrained RGB Generators to RGBA” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Qijie Wang, Zhixin Xu, Zhe Li, Haowei Zhu, Honghao Qiao, Bo Peng, Dawei Li, Bin Wang
AlphaAnything extends pretrained RGB generators to produce RGBA assets for text-prompted generation and image-prompted extraction. It adds an opacity branch that shares the RGB latent and uses alpha supervision to train the generator, without RGBA reconstruction pretraining. On 735 RGBA-Image examples, Qwen Image Edit reaches 0.9223 Soft IoU versus 0.6881 for OmniAlpha; the result is limited to tested tasks and data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.