# Reproducibility Notes

This repository now contains runnable entrypoints for the main T-Stitch experiment families, but full paper reproduction
still requires external pretrained checkpoints, datasets, reference statistics, and some third-party baseline projects.

## DiT ImageNet

Figure 5 pairwise sweeps can be generated with `dit/sample_ddp_t_stitch.py` by selecting the small and large DiT models:

```bash
cd dit
torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py \
  --small-model DiT-S/2 \
  --large-model DiT-XL/2 \
  --solver ddim \
  --num-sampling-steps 100 \
  --num-fid-samples 5000

torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py \
  --small-model DiT-B/2 \
  --large-model DiT-XL/2 \
  --solver ddim \
  --num-sampling-steps 100 \
  --num-fid-samples 5000

torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py \
  --small-model DiT-S/2 \
  --large-model DiT-B/2 \
  --solver ddim \
  --num-sampling-steps 100 \
  --num-fid-samples 5000
```

Figure 6 three-model sweeps are enabled by `--three_combo`. By default it enumerates the 0.1 allocation grid. Use
`--three-fractions S B XL` to run a single allocation.

```bash
torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py \
  --three_combo \
  --solver ddim \
  --num-sampling-steps 100 \
  --num-fid-samples 5000
```

Figure 8 sampler ablations can use `--solver ddpm`, `--solver ddim`, or `--solver dpm-solver++`. The local
DPM-Solver++ implementation supports `--dpm-solver-order 1` and `--dpm-solver-order 2`.

Evaluate generated `.npz` files with:

```bash
python evaluator.py /path/to/adm/ref_batch.npz /path/to/samples.npz
```

## Stable Diffusion and MS-COCO

Generate Table 2-style SD v1.4 / BK-SDM Tiny samples from COCO captions:

```bash
python sd/generate_tstitch_prompts.py \
  --pipeline sd \
  --prompts /path/to/captions_val2014.json \
  --output-dir outputs/sd_coco \
  --height 256 \
  --width 256 \
  --steps 50 \
  --guidance-scale 7.5 \
  --limit 5000
```

Pack each ratio directory for FID/IS:

```bash
python sd/images_to_npz.py \
  --image-dir outputs/sd_coco/ratio-0.3 \
  --output outputs/sd_coco/ratio-0.3.npz
```

Compute raw CLIP cosine score:

```bash
python sd/clip_score.py \
  --metadata outputs/sd_coco/ratio-0.3/metadata.jsonl \
  --output outputs/sd_coco/ratio-0.3-clip.json
```

FID/IS still require a matching MS-COCO reference `.npz`, evaluated with `dit/evaluator.py` or `ldm/evaluator.py`.

## SDXL, ControlNet, LCM, DeepCache, ToMe

SDXL and LCM prompt sweeps use the same prompt generator:

```bash
python sd/generate_tstitch_prompts.py --pipeline sdxl --prompts prompts.txt --output-dir outputs/sdxl
python sd/generate_tstitch_prompts.py --pipeline sdxl-lcm --prompts prompts.txt --output-dir outputs/sdxl_lcm --steps 4 --guidance-scale 0
```

ControlNet prompt sweeps:

```bash
python sd/generate_tstitch_controlnet.py --control canny --prompts prompts.txt --output-dir outputs/controlnet_canny
python sd/generate_tstitch_controlnet.py --control depth --prompts prompts.txt --output-dir outputs/controlnet_depth
```

Optional acceleration baselines are wired into `sd/generate_tstitch_prompts.py`:

```bash
python sd/generate_tstitch_prompts.py --pipeline sd --prompts prompts.txt --output-dir outputs/sd_tome --tome-ratio 0.5
python sd/generate_tstitch_prompts.py --pipeline sd --prompts prompts.txt --output-dir outputs/sd_deepcache --deepcache-interval 3
```

These require external packages:

```bash
pip install tomesd
pip install DeepCache
```

## External Baselines Not Vendored

The paper's SN-Netv2 model-stitching baseline and U-ViT experiments require third-party model definitions, training code,
and checkpoints that are not included in this repository. This repository can now generate T-Stitch sample batches and
metrics for comparison, but it still cannot train or evaluate SN-Netv2/U-ViT baselines without those external projects.