How dreamsartificial648-sketch/loravideotrainer works
LoRA Video Trainer is a local Windows desktop application for preparing video datasets, training Wan 2.1 T2V 1.3B LoRAs, and generating review clips. Its main workflow runs from importing or splitting clips, through captioning and validation, to latent/text caching, training, and test-video generation; the app also supports checkpoint resumption and storyboard organization. The graph reflects the documented stages and sampled app/tool behavior; cache, training, generation, and settings wiring is described by the README but their implementation excerpts were not provided.
Main entry points: app.py Tkinter app · 04_generate_sample.ps1 Sample generation · split_video.py Video splitter · validate_dataset.py Dataset validator
Components 15
Desktop App
- Tkinter app
app.py - Session settings
app.py
Dataset Preparation
- Video clips
- Clip captions
- Video splitter
tools/split_video.py - Dataset validator
tools/validate_dataset.py
Training Pipeline
- Latent caching
scripts/01_cache_video.ps1 - Text caching
scripts/02_cache_text.ps1 - LoRA training
scripts/03_train.ps1 - LoRA checkpoints
- Musubi Tuner
Generation & Review
- Take storyboard
app.py - Sample generation
scripts/04_generate_sample.ps1 - Generated takes
Other components
- Local user
Connections 24
- Local user → Tkinter app: operates
- Tkinter app → Session settings: restores saves
- Tkinter app → Video clips: imports previews
- Tkinter app → Clip captions: edits captions
- Tkinter app → Video splitter: launches splitting
- Video splitter → Video clips: writes clips
- Video splitter → Clip captions: writes captions
- Tkinter app → Dataset validator: launches validation
- Dataset validator → Video clips: inspects videos
- Dataset validator → Clip captions: checks pairs
- Tkinter app → Latent caching: starts stage
- Tkinter app → Text caching: starts stage
- Tkinter app → LoRA training: starts training
- Latent caching → Musubi Tuner: uses trainer engine
- Text caching → Musubi Tuner: uses trainer engine
- LoRA training → Musubi Tuner: runs training
- LoRA training → LoRA checkpoints: saves checkpoints
- Tkinter app → Sample generation: starts generation
- Sample generation → Musubi Tuner: uses trainer engine
- Tkinter app → LoRA checkpoints: selects resume model
- Sample generation → LoRA checkpoints: loads LoRA
- Sample generation → Generated takes: writes takes
- Tkinter app → Take storyboard: organizes takes
- Take storyboard → Generated takes: references takes