short review
Last night, TheLastBen released the lora Filmic.
https://huggingface.co/TheLastBen/Filmic
https://civitai.com/models/295249
The stats for the lora: 2 likes on HuggingFace (1 from me) and 1 review on Civit (you get it).
At the same time, the lora seems like a game-changer that could potentially bankrupt millions of startups invested in video models. I don't know if it happened by accident or not, but the lora makes video generation
- more consistent (better than TemporalNet but not yet AnimateDiff)
- changes the composition in the frame (makes it more cinematic)
- adds niceties like first-person view.
I went to the cinema to see "The Master and Margarita" (recommend) and there's a first-person flight scene. I was watching and thinking about how to replicate this in text-to-video. Well, or at least throw a character off a roof, that's also not bad.
The lora has a basic, eye-catching style. Not enough details. It washes the picture into a realistic style. But you can play with the weight and there are plenty of lora
for adding details/style.
I'm very curious how he trained it. It seems to be on sequential frames from video (but that's not precise). Maybe it happened by accident, it happens, lol. And if so, you could try taking blockbuster trailers (I'm not a lawyer, but trailers usually have a special license) and train it better.
In general, I tried a video with this lore last night, and the composition changes even at a low weight.
Either I'm missing something, or everyone's a little blind. Or I'm overreacting and mistaking the improvement in consistency for something else. Hehe.