I ENCOURAGE YOU TO MAKE AND POST SOME VIDEOS, TRAINING THIS WAS FAR FROM FREE!
AB TESTS 4 and 8 steps:
Short model description
frieren_000003000.safetensors is a rank-32 motion and native-audio LoRA for
the LTX 2 family, tested with the LTX 2.3 22B distilled image-to-video workflow.
It is intended to animate an already-correct first frame into restrained anime
character performance: breathing, blinking, gaze changes, conversational
reactions, walking, sequential dialogue, and quiet environmental movement.
It works best when the prompt explicitly identifies every visible character, assigns every
character a speech state, and uses Then, responds, and replies to express
speaker order.
This is i2v LoRA. It does not create the initial
Frieren/Fern/Stark image and cannot repair a wrong first-frame identity. Generate
or supply the correct guide image first, then use this LoRA for LTX motion and
native audio.
Recommended settings
Ltx 2.3 dev distilled
8 steps
i2v
1080p (1920*1080)
What it is good at
- Calm anime I2V performance from a strong first-frame guide.
- Frieren, Fern, and Stark reacting naturally in the same shot.
- One visible speaker while the other characters listen silently.
- One, two or three speakers talking in an explicitly ordered sequence. Also by prompting "Fern is speaking off-screen..." you can often make it happen.
- Native LTX ambience and speech generated with the video.
- Subtle breathing, blinking, gaze changes, hair/cloth movement, and wind.
Description
initial commit
FAQ
Comments (8)
cool! how many clips did you train and for how long the quality is crisp
I wouldn't say the quality is crisp haha, except the 6 talking sequences the clips are somewhat cherry-picked. This was trained with batch size 2 for 3000 steps, 5 temporal buckets, 8 resolution buckets, total 300+ clips.
@HSDHC nice that's a lot of clips! must gave taken some time. many thanks!
For some reason, I can't reproduce the anime style. Instead, it generates a generic anime style with characters that only vaguely resemble the description. The strength is set to 1.0 and prompt is copied.
do you use i2v (image to video)? this is an i2v model.
@HSDHC got it, my fault.
@Tagrim you can use this model to make frames https://civitai.red/models/2835233/frieren-anime-screencap
@HSDHC thanks