Official data processing script
Hello!
LTX already has a native trainer. However, launching a training script is the least friction part. The bottleneck is dataset collection / creation / filtering out.
Would it be possible for you to open-source a script, a pipeline, or an AI agent, which will generate a proper LTX format training dataset (finding the targets, cropping, auto-captioning, etc.) given a video or a collection of videos and the target concept / style?
I think lowering the fine-tuning effort floor will increase the retention as people are getting accustomed to LoRA-less models from competitors and any training hustle is now fully undesirable.
Thank you!
Hi @kabachuha -- did you take a look at the training agent we make available here?
https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-trainer/README.md#-agent-assisted-training
Need to take a look, thanks! Concepts, styles need probably some more instructions on how to cut the particular fragments, and to train them losslessly (resampling to the framerate/duration) to fit the needed SFX or action into the latent frame downsampling window, clean the dataset of unneeded samples, and find the fragment of interest by some algorithm, when the default scene-splitting is not enough.