File size: 2,169 Bytes
83a780e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
# Code and Checkpoint Usage

## Code Snapshot

The code archive contains two trees:

- `TwoFrame`: experiment orchestration, evaluation, data-engine scripts, and local project code.
- `FastVideo`: the training stack used for this Wan2.2 fine-tune.

FastVideo matters for this run. The training log references FastVideo modules such as `fastvideo_args.py`, `training_pipeline.py`, and `training_utils.py`. The relevant dirty FastVideo changes are recorded in:

- `metadata/FastVideo.git_status.txt`
- `metadata/FastVideo.uncommitted.diff`
- `metadata/FastVideo.untracked_files.txt`

TwoFrame status is recorded similarly in `metadata/TwoFrame.*`.

## Environment

The UCSF environment used during training was:

```bash
source ~/.twoframe_env.sh
conda activate /scratch/user/yuhwang/envs/twoframe
export PYTHONPATH=/scratch/user/yuhwang/code/FastVideo:/scratch/user/yuhwang/code/TwoFrame:$PYTHONPATH
```

`~/.twoframe_env.sh` points caches and temporary directories to scratch.

## Checkpoints

Use the EMA checkpoint for inference-style evaluation unless you explicitly want non-EMA weights:

- EMA: `checkpoints/ema_checkpoint-8000/diffusion_pytorch_model.safetensors`
- non-EMA: `checkpoints/checkpoint-8000/transformer/diffusion_pytorch_model.safetensors`

Use the distributed checkpoint only if resuming training:

- `checkpoints/checkpoint-8000/distributed_checkpoint/`

## Base Model

The run loaded base model components from:

`/scratch/user/yuhwang/model/Wan2.2-TI2V-5B-Diffusers-merged`

That directory is symlink-composed. Its symlink map is recorded in `metadata/base_model_symlinks.txt`. The base model binaries are not duplicated in this package; this package contains the fine-tuned transformer checkpoint and reproducibility assets.

## Training Command Reconstruction

The exact shell launcher was not preserved in the final run directory. The effective training configuration is captured from the log in `metadata/train_args.json`. Use that file as the source of truth for reconstructing a resume or reproduction launch.

The original output directory was:

`/scratch/user/yuhwang/artifacts/twoframe/pants_wan22_finetune/pants_b16_9k_20260519_215532`