ToMe-SD¶
ToMe-SD is a token merging method for diffusion inference. The paper cites it as an early example showing that redundant diffusion tokens can be merged during inference with limited quality loss.
Role in Sol-Engine¶
ToMe-SD is part of the token-reduction design space. It is a candidate when merging similar tokens is safer than dropping tokens outright.
Search knobs¶
- merge ratio.
- layer placement.
- timestep schedule.
- similarity metric.
Validation¶
Token merging can blur fine detail. Inspect texture, object boundaries, and motion consistency.