What's Changed
- Support "fixed" HunyuanVideo i2v model (actually a model with a different architecture from the original released version).
- Use fp16 as the default compute dtype for the WAN 2.1 models.
- Support fp8_scaled model files that don't enable the fp8 matrix mult by default.
- fixed: Incorrect guide message for missing frontend. by @ltdrdata in #7105
- Typo in node_typing.py by @JettHu in #7092
- Update frontend to 1.11.8 by @huchenlei in #7119
- Weight Hooks Switching Optimization by @Kosinkadink in #7067
- Fix stable cascade VAE issues with lowvram.
Full Changelog: v0.3.24...v0.3.26