Questions
#1
by
AlekseyCalvin
- opened
- There is no workflow example provided on the main page here.
- The github workflow may be sufficient for many use cases, but is rather minimal and leaves some questions.
- Such as: What is the role of the provided ipips model for inference? Is co-activating it necessary within an inference implementation? I skimmed the inference code provided on github and did not immediately see any reference to the ipips model. To be fair, I haven't yet read your full paper. But generally speaking, that's not commonly a prerequisite for testing sampling acceleration adapters (at least in the HF orbit).
- Is your acceleration paradigm/technique compatible with DiT-based base models (incl. FLUX models, PixArt, S.D.3.5, etc...)? Is the provided implementation/version?
In directions, here and at GitHub, you invite prospective testers/users to apply it to a "base model", but you do not specify the likely limits of the adapter's current base model compatibility. Having skimmed your paper, I've only seen samples over SDXL-base there (and in relation to other SDXL acceleration frameworks). But this specificity is not made explicit within any summary you provide.