run_examples.sh

#!/usr/bin/env bash
set -euo pipefail

ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"

usage() {
  cat <<'EOF'
T-Stitch command helper.

This script groups the main runnable entrypoints in this repository.
Most commands assume you are using a GPU environment and have already
downloaded the required model weights or datasets.

Usage:
  ./run_examples.sh <command> [extra args...]

Commands:
  help
    Print this help.

  install-basic
    Install the root-level Python dependencies from requirements.txt.

  gradio
    Launch the Gradio demo for SD 1.x, SDXL, and SDXL + LCM examples.

  sd-demo
    Run the Stable Diffusion T-Stitch demo at the repo root.

  sdxl-demo
    Run the SDXL T-Stitch ratio sweep demo.

  sdxl-canny
    Run the SDXL + ControlNet canny demo.
    Use -- --image /path/to/image.jpg to override the synthetic fallback image.

  sdxl-depth
    Run the SDXL + ControlNet depth demo.
    Use -- --image /path/to/image.jpg to override the synthetic fallback image.

  sdxl-pose
    Run the SDXL + ControlNet openpose demo.
    Note: this downloads a reference image from Hugging Face.

  sdxl-lcm
    Run the SDXL + LCM-LoRA T-Stitch demo.

  train-sdxl -- <train args...>
    Launch SDXL T-Stitch training with accelerate.
    Example:
      ./run_examples.sh train-sdxl -- \
        --train_data_dir /path/to/images \
        --metadata_file /path/to/metadata.jsonl \
        --ratio 0.3

  dit-sample
    Run the DiT T-Stitch sampling example from dit/sample_t_stitch.py.

  dit-sample-all
    Run the DiT all-tradeoffs sampling example.

  dit-train -- <train args...>
    Launch DiT T-Stitch training from dit/train_t_stitch.py.
    Example:
      ./run_examples.sh dit-train -- \
        --data-path /path/to/imagenet_train \
        --ratio 0.3 \
        --ratio-schedule fixed \
        --image-size 256 \
        --global-batch-size 4 \
        --epochs 1

  dit-fid
    Run distributed DiT T-Stitch sample generation for FID evaluation.
    Extra args are passed to dit/sample_ddp_t_stitch.py.

  dit-fid-xl-b
    Generate the Figure 5 DiT-B/XL ratio sweep.

  dit-fid-b-s
    Generate the Figure 5 DiT-S/B ratio sweep.

  dit-fid-three
    Generate the Figure 6 DiT-S/B/XL three-model sweep.

  dit-fid-dpm
    Generate Figure 8-style DiT-S/XL samples with DPM-Solver++.

  sd-coco-generate -- --prompts /path/to/captions_val2014.json --output-dir outputs/sd_coco
    Generate SD/BK-SDM T-Stitch images for MS-COCO prompt evaluation.

  sd-pack -- --image-dir outputs/sd_coco/ratio-0.3 --output outputs/sd_coco/ratio-0.3.npz
    Pack generated images into an ADM-style .npz file for FID/IS.

  clip-score -- --metadata outputs/sd_coco/ratio-0.3/metadata.jsonl
    Compute CLIP cosine score for generated prompt-image pairs.

  sdxl-prompts -- --prompts prompts.txt --output-dir outputs/sdxl_prompts
    Generate SDXL/SSD-1B prompt sweeps for quantitative or qualitative evaluation.

  sdxl-controlnet -- --control canny --prompts prompts.txt --output-dir outputs/controlnet_canny
    Generate SDXL ControlNet T-Stitch prompt sweeps.

  ldm-sample
    Run the LDM T-Stitch sampling example.

  ldm-sample-all
    Run the LDM all-ratios sampling example.

  ldm-train
    Print a commented LDM training template.

  ldm-fid
    Run distributed LDM T-Stitch sample generation for FID evaluation.
EOF
}

require_passthrough_args() {
  if [[ $# -eq 0 ]]; then
    echo "This command needs extra arguments after --." >&2
    exit 1
  fi
}

cmd="${1:-help}"
if [[ $# -gt 0 ]]; then
  shift
fi

case "$cmd" in
  help|-h|--help)
    usage
    ;;

  install-basic)
    cd "$ROOT_DIR"
    pip install -r requirements.txt
    ;;

  gradio)
    cd "$ROOT_DIR"
    python sd/gradio_demo.py
    ;;

  sd-demo)
    cd "$ROOT_DIR"
    python sd/sd_demo.py
    ;;

  sdxl-demo)
    cd "$ROOT_DIR"
    python sd/sdxl_demo.py
    ;;

  sdxl-canny)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    python sd/sdxl_canny.py "$@"
    ;;

  sdxl-depth)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    python sd/sdxl_depth.py "$@"
    ;;

  sdxl-pose)
    cd "$ROOT_DIR"
    python sd/sdxl_pose.py
    ;;

  sdxl-lcm)
    cd "$ROOT_DIR"
    python sd/sdxl_lcm_lora.py
    ;;

  train-sdxl)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    accelerate launch sd/train_sdxl_tstitch.py "$@"
    ;;

  dit-sample)
    cd "$ROOT_DIR/dit"
    python sample_t_stitch.py --solver ddim --num-sampling-steps 100 --seed 4 --ratio 0.5
    ;;

  dit-sample-all)
    cd "$ROOT_DIR/dit"
    python sample_t_stitch.py --solver ddim --num-sampling-steps 100 --seed 4 --all_tradeoffs
    ;;

  dit-train)
    cd "$ROOT_DIR/dit"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    torchrun --master_port=29501 --nnodes=1 --nproc_per_node=1 train_t_stitch.py "$@"
    ;;

  dit-fid)
    cd "$ROOT_DIR/dit"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py --num-fid-samples 5000 --solver ddim --num-sampling-steps 100 "$@"
    ;;

  dit-fid-xl-b)
    cd "$ROOT_DIR/dit"
    torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py --num-fid-samples 5000 --solver ddim --num-sampling-steps 100 --small-model DiT-B/2 --large-model DiT-XL/2
    ;;

  dit-fid-b-s)
    cd "$ROOT_DIR/dit"
    torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py --num-fid-samples 5000 --solver ddim --num-sampling-steps 100 --small-model DiT-S/2 --large-model DiT-B/2
    ;;

  dit-fid-three)
    cd "$ROOT_DIR/dit"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py --three_combo --num-fid-samples 5000 --solver ddim --num-sampling-steps 100 "$@"
    ;;

  dit-fid-dpm)
    cd "$ROOT_DIR/dit"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    torchrun --nnodes=1 --nproc_per_node=8 sample_ddp_t_stitch.py --num-fid-samples 5000 --solver dpm-solver++ --num-sampling-steps 50 "$@"
    ;;

  sd-coco-generate)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    python sd/generate_tstitch_prompts.py --pipeline sd --height 256 --width 256 --steps 50 --guidance-scale 7.5 "$@"
    ;;

  sd-pack)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    python sd/images_to_npz.py "$@"
    ;;

  clip-score)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    python sd/clip_score.py "$@"
    ;;

  sdxl-prompts)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    python sd/generate_tstitch_prompts.py --pipeline sdxl "$@"
    ;;

  sdxl-controlnet)
    cd "$ROOT_DIR"
    if [[ "${1:-}" == "--" ]]; then
      shift
    fi
    require_passthrough_args "$@"
    python sd/generate_tstitch_controlnet.py "$@"
    ;;

  ldm-sample)
    cd "$ROOT_DIR/ldm"
    python scripts/sample_imagenet_32_t_stitch.py --ratio 0.5 --sampling-steps 100 --cfg-scale 3.0
    ;;

  ldm-sample-all)
    cd "$ROOT_DIR/ldm"
    python scripts/sample_imagenet_32_t_stitch.py --all_ratios --sampling-steps 100 --cfg-scale 3.0
    ;;

  ldm-train)
    cat <<'EOF'
# LDM T-Stitch training template.
# Run this from the repo root or adapt the paths first.

cd /Users/ouzhang/Desktop/nips/T-Stitch/ldm
python main.py \
  --base configs/latent-diffusion/cin-ldm-vq-f8-t-stitch.yaml \
  -t True \
  --gpus 0, \
  model.params.small_ratio=0.3 \
  model.params.ratio_schedule=fixed \
  model.params.distill_weight=0.0
EOF
    ;;

  ldm-fid)
    cd "$ROOT_DIR/ldm"
    python -m torch.distributed.launch --nproc_per_node=8 --master_port 1236 --use_env scripts/sample_imagenet_ddp_t_stitch.py --num-fid-samples 5000 --num-sampling-steps 100 --cfg-scale 3.0 --ratio 0.5
    ;;

  *)
    echo "Unknown command: $cmd" >&2
    echo >&2
    usage
    exit 1
    ;;
esac