The new cache validation in the sd3 branch only checks the key name. When multiresolution is disabled (like with sd and sdxl training), the key is latents and the size of the tensor is not checked. If the training resolution or bucketing changes, the cache should not pass validation.
The new cache validation in the
sd3branch only checks the key name. When multiresolution is disabled (like with sd and sdxl training), the key islatentsand the size of the tensor is not checked. If the training resolution or bucketing changes, the cache should not pass validation.https://github.com/kohya-ss/sd-scripts/blob/123474d784096a2a3063d1ff91ef1957b41ccc72/library/strategy_base.py#L446-L451