V1: Video Lora is trained 48 stills dumped from a Car Chase scene as part of a 90's Japanese Action Movie. (image based Hunyuan Video Trainer was used)
Prompt:
"xjxcar style video of custom cars, racing on tokyo highways at night, JDM style found footage"
V2 was trained on Video clips, not images, to show the difference (same dataset)
Prompt:
"xjx style video of custom cars, racing on tokyo highways at night, JDM style found footage"
Description
trained video dataset version with replicate
FAQ
Comments (6)
It is trained on one of the "Shuto Kôsoku toraiaru" aka "Shuto Kousoku Trial" aka "Megalopolis Expressway Trial" film series scenes?
I'm unsure which movie exactly, but this is from the same production team, cinematography and locations. The chase in my dataset featured a white maxda rx7 (FC) and a yellow skyline R32.
Really cool. I assume a drift LoRA is in the works?
yes.
I'm working on a few Drift themed Lora, with success on our side, there will be more published very soon
Good job here. v2 IMO looks more solid (maybe was obvious but it's interesting to see a comparison between two). Curious about the captions on your dataset, do you only use the same caption for all scenes? or different for each? Now you need v3 = v2 + v1 images hehe. Cool stuff man!
for V2 I used "XJX video of custom cars [ qwen VL caption ] JDM found footage in the style of XJX"
as a rough blueprint for this one. it seems to be almost exactly right with the caption example in description.
for v1 I had to use a simple catch all prompt which was applied to all the images.
I think that image training to hyvideo is more accurate to the ground truth, but video clip based training gets all those small motions correct.
Importantly i train with 125 frame segment clips and i drop boring or confusing segments from the dataset before training. Basically split the video into segments of 125 frames, then use Uniform frame sampling with 8 frames on that parameter.
I might retrain with fully manual captions, but this will take time to get done.
If you liked this one i have a new one with more varied video data set, training on H100 (for 13hrs)
if we are lucky it will be ready tomorrow
thanks for asking :)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.