How mrzhangjl/6127 works
This is a single-file comparative image-generation experiment, rather than a general-purpose application. Its workflow prepares a custom reference set, loads VAR and DiT, generates and evaluates images, then benchmarks and explores model settings to produce tables and figures. The source excerpt is partial and the README is empty, so links are shown only where the excerpt or its explicit comments support them; all repository components use the exact source path shown in the tree.
Main entry points: var_vs_dit_pipeline.py Classes and config
Components 15
Experiment setup
- Classes and config
files-2/files-2/var_vs_dit_pipeline.py - Reference preparation
files-2/files-2/var_vs_dit_pipeline.py - Custom reference images
Generation models
- VAR model loading
files-2/files-2/var_vs_dit_pipeline.py - DiT pipeline loading
files-2/files-2/var_vs_dit_pipeline.py - VAR scale sampler
files-2/files-2/var_vs_dit_pipeline.py - DiT timestep sampler
files-2/files-2/var_vs_dit_pipeline.py - Sample generation
files-2/files-2/var_vs_dit_pipeline.py
Evaluation
- Timing and VRAM benchmark
files-2/files-2/var_vs_dit_pipeline.py - Quality metrics
files-2/files-2/var_vs_dit_pipeline.py - Parameter ablations
files-2/files-2/var_vs_dit_pipeline.py
Results
- Progression and tradeoff figures
files-2/files-2/var_vs_dit_pipeline.py - Tables and image outputs
files-2/files-2/var_vs_dit_pipeline.py
Other components
- Researcher
- Hugging Face Hub
Connections 23
- Researcher → Custom reference images: provides images
- Researcher → Classes and config: sets experiment
- Reference preparation → Custom reference images: normalizes images
- Classes and config → Reference preparation: sets classes
- VAR model loading → Hugging Face Hub: downloads checkpoints
- Classes and config → VAR model loading: sets depth
- Classes and config → DiT pipeline loading: sets model
- VAR model loading → VAR scale sampler: provides model
- DiT pipeline loading → DiT timestep sampler: provides pipeline
- Classes and config → Sample generation: sets sampling options
- Sample generation → VAR scale sampler: calls
- Sample generation → DiT timestep sampler: calls
- Timing and VRAM benchmark → VAR scale sampler: times calls
- Timing and VRAM benchmark → DiT timestep sampler: times calls
- Sample generation → Tables and image outputs: saves samples
- Quality metrics → Custom reference images: compares against
- Quality metrics → Tables and image outputs: writes metrics
- Parameter ablations → Sample generation: varies settings
- Parameter ablations → Quality metrics: evaluates samples
- Timing and VRAM benchmark → Tables and image outputs: writes timings
- Progression and tradeoff figures → VAR scale sampler: captures progression
- Progression and tradeoff figures → DiT timestep sampler: captures progression
- Progression and tradeoff figures → Tables and image outputs: saves figures