To rename for future compatibility with transformers
#71
by xiaohei66 - opened
Our vision encoder is a heavily modified version of SigLIP, featuring a dynamic resolution mechanism and 2D RoPE instead of the original’s fixed resolution and learnable absolute position embeddings.
This makes our implementation fundamentally different from the standard SigLIP in libraries like Transformers. To avoid future naming conflicts and confusion, we must move away from the Siglip* name.
xiaohei66 changed pull request title from rename to To rename for future compatibility with transformers
xiaohei66 changed pull request status to open
xiaohei66 changed pull request status to merged