Trained: musubi-tuner, 30 images
Caption: adjusted caption from Qwen2.5-VL-3B (includes described background)
lr:(4e-4,flow_shift7.0,timestep_boundary 885)
epoch:100(60epoch when Wan2.2 lora start from Wan2.1 lora weights)
train_step:3000(1800)
lora_strength:0.8
アクションは強調promptによって3.0～5.0に調整する必要があります。またはモーションloraを併用ください。
lightx2v Wan2.2 v1.1との併用を推奨します。
Action prompt should be adjusted to 3.0 to 5.0 depending on emphasis prompt. Or motion lora use together.
It is recommended to use with lightx2v Wan2.2 v1.1 lora.
・Wan2.2 T2V LoRA作成メモ
https://civarchive.com/models/1522679/igarashi-rika-or-wan21-and-22-14b-t2v
・Wan2.2 vid2vid サンプル
参考動画はComfyUI/input に置く。
Description
FAQ
Comments (7)
My friend, the size of these lora's are nuts... These are cell-shaded images. Even Lossless quality at 4K on 30 cell-shade images would not be anywhere near the GB's these LoRA's are. Good luck with them...
While I agree that the way these are packaged is terrible to use, the loras themselves aren't that big.
The file seems that large because each one is a zip that contains 10's of loras in each, I think some have up to 40 or more....
I'm not sure if it's worth the trouble for me though - as in order to use them properly I need to try to get a preview image for each, and well.... that's a ridiculous amount of work for something I can probably create with many other loras.
Zipping up the models is a bad idea. They're unusable without decompression (which online generators don't do), you need to download gbs for potentially just one file, you cannot update without compressing again.
If you want a themed group, you should instead use civ's collection feature on the individual models you uploaded (I'm guessing these are the separate ones already uploaded).
There are 72 LoRAs that have been posted, and 139 LoRAs that have not been posted.
Unposted LoRAs cannot be grouped.
Of course, LoRAs in Wan2.2 reuse the training data and prompts from Wan2.1, so it is not as much of a burden, but posting LoRAs still counts as a burden.
@kihh I understand, citrus once told me what a drag civ's uploader is like for mass upload (crashes and parallel upload limit were the biggest pain points, I think).
Just expressing my thoughts on the matter and providing a suggestion; by all means, don't feel obliged to do something that causes you great discomfort.
By the way, I think it's currently broken, but narugo posted this auto upload code back in the day, perhaps it can be refurbished. https://github.com/narugo1992/civitai_client
('fcourse maybe you already have one. but mentioning just in case.)
Understandable that with this many loras (thanks btw!) zipping them is pretty well essential - but do you think you could consider zipping them with at least one preview file of the same name for each?
that would make it all so much easier!!
I've posted a zip of sample output, although the lora file names don't match.
I wanted to post it because some of the output samples get hidden by civitai when combined.