V2:
Trained using 489 clips + 356 images from the show at the native 768x576 resolution. This is my first time doing images + videos together with the Ostris adapter v3 on musubi fork and the result is great! The motion is more alive and fluid than the base model comparisons. I think it makes things more interesting while bringing in the style from the original show. Same triggers as V1.
Images were microbatching with 8; videos up to 124 frames max length thus the low # of steps.
V1:
Trained on 369 screen captures from Golden Boy at native res 768x576.
H3 is wonderful, it knows anime quite well, I will potentially try training it on video clips from the show but it does a pretty good job with just images.
It trains mega fast on images at this res, around 1.1s/it and this lora trained in around 3 hours. I'll do some extra training tomorrow but I am happy enough with the 12k checkpoint.
Trigger:
retro 90s 2d anime style, golden boy style
Not tested with any speed up loras like lightx. All examples are t2v 1mp 20 steps, euler beta57.
Training:
Trained using musubi tuner fork, lr 5e-5, --h3_spatial_density_jitter 0.2 --h3_guidance_distillation_scale 3.5 --h3_base_preservation_loss_weight 0.02 \
Rank 16
Description
Trained using 489 clips + 356 images from the show at the native 768x576 resolution. This is my first time doing images + videos together with the Ostris adapter v3 on musubi fork and the result is great! The motion is more alive and fluid than the base model comparisons. I think it makes things more interesting while bringing in the style from the original show. Same triggers as V1.
Images were microbatching with 8; videos up to 124 frames max length thus the low # of steps.