>>97096574>what are y'all thoughts on NAI V4? It seems to get the artist more accurately, but it's so different than how it used to gen it with v3 I am not sure how I feelpros
>The gens no longer suffer a stroke when handling small details like text or forked tongues>Bigger dataset with more recent tags to experiment with>Natural language support actually helps with complex compositions>Character prompts are a really good additioncons
>Most gens have weird artifacts and random blurriness>While it can replicate some artists better, for most the quality is actually worse>Art style mixing in v4.0 is a mess. You have to play around a lot with the weights of every artist tag just to get a decent result, but doing that also increases the chances of adding a pile of artifacts to the gen>Art style mixing in v4.1 is just gone. One artist will always dominate, and trying to mix styles results in a blurry mess>The art style dominance also affects variation. if you specify an artist, you’ll likely get nearly identical gens with only small differences if you don't use wildly different prompts>Character prompts can straight up fry the image>UC tags can straight up fry the image>Quality tags can straight up fry the image>Using natural language for the prompts is broken. Sometimes it works, sometimes it gets ignored, and sometimes the model just prints the whole prompt as a speech bubble>The composition tags are now completely unreliable. You ask for a face close up, you get an ass shot>Even if you use all the new available tools to control the characters' actions, the model still gets things wrong half the timeOverall, it feels like a half baked upgrade. Even ignoring the massive issue with the output quality, the supposed accuracy boost from the new prompt system just isn’t there, and getting consistent results for scenes with more than 1 character is a pain in the ass. It's a shame because the model itself is really powerful, but it clearly wasn't ready for full release, and meanwhile the devs are just throwing random hotfixes at the wall, hoping something sticks, without really knowing how to address the issues people are reporting
