4 ms·
I disagree. They report that their method produces a model that is 6.5 GB from flux (22.7GB). Why wouldn't you want to know how their 6.5GB model compares to o
by djoldman 2y ago
I disagree.
They report that their method produces a model that is 6.5 GB from flux (22.7GB). Why wouldn't you want to know how their 6.5GB model compares to other 6.5GB models?
Regarding aesthetic prefs: it's an open problem what an appropriate metric is for GenAI... LLM arena is widely regarded as a good way to measure LLMs and that's user preferences.
In any case, the authors report LPIPs etc. They could do the same for other small models.
- snovv_crash 2y agoLPIPS and similar don't work if the scene is different, as happens if the random seed doesn't match. This is why they can use it to compare the quantised network, but not against networks with reduced numbers of weights.