NOTE: FUN LORA does not work on I2V!This Lora is my experiment on Kohya Musubi Tuner's newest FUN model training capability. As a side bonus I also have a I2V model, T2V hasn't converged for me so need more time there. This will also be the last couple models I plan to train and publish on this platform before I learn something cool and new. Getting content out has been extremely hard for me as multiple submissions have been subject to review for days and it's exhaustive.So what does this Lora do if it's not T2V? It basically stabilizes the character's face. My experiment was to concat multiple videos generated using controlnet to form a longer dance video, and without this lora the faces would morph mid flight. The experiment was semi-successful, because I didn't use last frame clip extend, there's still some "jumps" every 5 seconds in terms of some details, lighting, and hair style, but when we have access to infinite gen duration controlnet models ( I'm dreaming about Skyreel FUN or Framepack WAN FUN now :P), this issue will hopefully go away!To recover my FUN workflow, look for the file that doesn't have audio.The lora is trained by 30+ quality 3D clips including in game footage and NSFW community contents, so it does have access to nudity, some nsfw motions and expressions, feel free to play around and discover yourself what it can do.

Tifa WAN 2.1 FUN controlnet + I2V + T2V

Preview, still in training but it already works well

tifaface-final-000003.safetensors

Files with SHA256: 7068a862e447d37ca0e6415280a51dc7fc65b3aaee8e801ca1308bc1805be89b