>>3023733I used WAN2.1 successfully on my single 3090 in ComfyUI, but haven't been able to get a 2.2 workflow to work yet. Like you said the smaller 5B model should definitely work, and even the larger ones seem like they should since it's split into high and low noise channels and those can independently fit in 24GB VRAM, but for some reason I'm still having issues.
I plan on trying again with
https://civitai.com/models/1981116?modelVersionId=2388548 when that becomes available in a few days.
>>3023731"Masterpiece" and "very awa" are similar in NoobAI models but have different effects since the algorithms for how they were trained are different.
Pic related is a quick example showing the effects of the positive and negative tags. It shows what the top quality and top aesthetic tags do independently, and then shows what the list of tags I currently use for positives and negatives all do together.
You can see the negative tags have some effect, the positive tags have a more significant effect, and using everything together really does improve the quality (or at least the level of detail since quality is subjective). Using the quality tags especially seems to emphasize making gens more colorful and including "detailed background" like in the final column gives a real background.
Also a fun trick is you can reverse the positives and negatives if you want to deliberately make something look bad.
Full res comparison at:
https://files.catbox.moe/ul8v0u.png 