Hugging Face's logo

Datasets:
Ouzhang
/
T-Stitch

Modalities:
Image
Formats:
imagefolder
Size:
1K - 10K
ArXiv:
Libraries:
Datasets
Dataset card Data Studio Files
xet
 Community
T-Stitch / CDF_EXPERIMENTS.md
Ouzhang's picture
Ouzhang
Add files using upload-large-folder tool
e5371e6 verified
preview code
|
Raw
History Blame Contribute Delete
47.7 kB

CDF / CST Experiment Protocol

This document defines the experiments for Curricular Diffusion Finetuning (CDF) on top of the existing T-Stitch codebase. Run commands from the repository root unless a command explicitly changes directory.

Large runs require the original checkpoints, ImageNet or COCO data, ADM-style reference .npz files, and GPUs. Paths in angle brackets are user-provided paths.

0. Experiment Runbook

This section is the execution checklist. Each experiment has a short code name that is used as the tmux session name, log directory, output directory, and table method key.

0.1 Experiment Table

Code Stage Meaning Train? Primary output
smoke_cst_fa sanity 20-step LDM CST+FA smoke test; verifies data, LR scaling, boundary routing, and checkpoint writing. yes logs/cdf/ldm/smoke_cst_fa
nft baseline / tuning reference No-training LDM T-Stitch baseline on ImageNet for ratios 0.0,0.1,...,1.0; gives the FID/IS reference before any tuning. no outputs/cdf/ldm/nft
lr_sweep_cst_cosine tuning Sweep CST cosine learning rate over 3e-6,1e-5,3e-5,1e-4 for 20k steps. yes logs/cdf/ldm/lr_sweep_cst_cosine
lr_sweep_cst_cosine_eval tuning eval Evaluate LR sweep on key ratios with 1000 samples. no outputs/cdf/ldm/lr_sweep_cst_cosine
fa_sweep_cst_cosine tuning Sweep FA weights after selecting BEST_LR. yes logs/cdf/ldm/fa_sweep_cst_cosine
fa_sweep_cst_cosine_eval tuning eval Evaluate FA sweep on key ratios with 1000 samples. no outputs/cdf/ldm/fa_sweep_cst_cosine
schedule_ablation tuning / ablation Compare curriculum schedules after selecting BEST_LR. yes logs/cdf/ldm/schedule_ablation
schedule_ablation_eval tuning eval Evaluate schedule ablation on key ratios or full ratios. no outputs/cdf/ldm/schedule_ablation
cst_cosine main LDM Full 200k-step CST with cosine curriculum, no FA. yes logs/cdf/ldm/cst_cosine
cst_cosine_eval main eval Evaluate cst_cosine on all ratios with 5000 samples. no outputs/cdf/ldm/cst_cosine
cst_cosine_fa main LDM Full 200k-step CST + FA with best FA weight. yes logs/cdf/ldm/cst_cosine_fa
cst_cosine_fa_eval main eval Evaluate cst_cosine_fa on all ratios with 5000 samples. no outputs/cdf/ldm/cst_cosine_fa
frft baseline Fixed-ratio fine-tuning, one junior checkpoint per ratio. yes logs/cdf/ldm/frft
frft_eval baseline eval Cross-ratio evaluation of all FRFT checkpoints. no outputs/cdf/ldm/frft
senior_only baseline Freeze junior and train senior only. yes logs/cdf/ldm/senior_only
senior_only_eval baseline eval Evaluate senior-only checkpoints. no outputs/cdf/ldm/senior_only
fmgt baseline Full model group training; jointly tune junior and senior. yes logs/cdf/ldm/fmgt
fmgt_eval baseline eval Evaluate FMGT checkpoints. no outputs/cdf/ldm/fmgt
dit_nft DiT DiT-S/2 + DiT-XL/2 no-finetuning T-Stitch baseline. no outputs/cdf/dit/dit_nft
dit_cst_cosine DiT DiT CST training with cosine curriculum. yes logs/cdf/dit/dit_cst_cosine
dit_cst_cosine_eval DiT eval Evaluate DiT CST on all ratios. no outputs/cdf/dit/dit_cst_cosine
dit_cst_cosine_fa DiT DiT CST + FA training. yes logs/cdf/dit/dit_cst_cosine_fa
dit_cst_cosine_fa_eval DiT eval Evaluate DiT CST+FA on all ratios. no outputs/cdf/dit/dit_cst_cosine_fa
dit_sampler_ablation DiT ablation Evaluate DiT CST under DDPM, DDIM, and DPM-Solver++. no outputs/cdf/dit/dit_sampler_ablation_*
sd_nft SD / COCO SD v1.4 + BK-SDM Tiny no-training T-Stitch on COCO. no outputs/cdf/sd/nft_fixed
sd_deepcache SD / COCO SD NFT plus DeepCache. no outputs/cdf/sd/deepcache_i3_fixed
sd_tome SD / COCO SD NFT plus ToMe token merging. no outputs/cdf/sd/tome_05_fixed
sdxl_cst_cosine optional SDXL Implemented SDXL/SSD-1B CDF training; not the SD v1.4 Table 2 setup. yes logs/cdf/sd/sdxl_cst_cosine

Recommended order for finding the best training parameters:

smoke_cst_fa -> nft -> lr_sweep_cst_cosine -> lr_sweep_cst_cosine_eval -> fa_sweep_cst_cosine -> fa_sweep_cst_cosine_eval -> schedule_ablation -> schedule_ablation_eval

Only after selecting BEST_LR, BEST_FA, and BEST_SCHEDULE, run the full 200k-step main experiments.

Sampling resume behavior:

  • LDM evaluation skips a ratio if samples/seed-...-ratio-<r>.npz already exists.
  • If sampling was interrupted before .npz creation, rerun the same command; existing PNG batches are skipped and missing images are generated.
  • If .npz exists but metrics-ratio-<r>.txt is missing, rerun the same command; sampling is skipped and only the evaluator runs.
  • If you want a completely fresh run, remove the corresponding output directory first.

Runtime logs:

  • run_cdf_experiments.sh writes stdout/stderr to run_logs/<command>/<timestamp>.log and still prints to tmux.
  • Set CDF_RUN_LOG=/path/to/file.log to choose a specific log path.
  • Set CDF_DISABLE_RUN_LOG=1 to disable automatic tee logging.

Evaluator device and metric resume:

  • The runner tries the TensorFlow FID/IS evaluator on GPU first, then falls back to CPU if TensorFlow CUDA fails.
  • Set CDF_EVALUATOR_USE_GPU=1 to force GPU evaluator and fail if it cannot start.
  • Set CDF_EVALUATOR_USE_GPU=0 to force CPU evaluator.
  • Existing metric files are skipped only when they contain both FID: and Inception Score:. Partial crashed logs are recomputed.

0.2 Tuning Commands 

# LDM Smoke CST+FA
tmux new -s smoke_cst_fa
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
bash run_cdf_experiments.sh ldm-train-cst-fa \
    --data-root "$LDM_DATA" \
    --schedule cosine \
    --ratio-start 0.1 \
    --ratio-end 0.2 \
    --fa-weight 0.1 \
    --fa-boundary-width 10 \
    --logdir logs/cdf/ldm/smoke_cst_fa \
    --max-steps 20 \
    --batch-size 2 \
    --num-workers 0 \
    --learning-rate 1e-5 \
    --scale-lr false \
    --gpus 0

# LDM NFT Baseline Before Tuning
tmux new -s nft
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-nft \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/nft \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# LDM LR Sweep CST Cosine
tmux new -s lr_sweep_cst_cosine
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
for LR in 3e-6 1e-5 3e-5 1e-4; do
    bash run_cdf_experiments.sh ldm-train-cst \
        --data-root "$LDM_DATA" \
        --schedule cosine \
        --ratio-start 0.1 \
        --ratio-end 0.9 \
        --logdir "logs/cdf/ldm/lr_sweep_cst_cosine/lr_${LR}" \
        --max-steps 20000 \
        --batch-size 64 \
        --num-workers 0 \
        --learning-rate "$LR" \
        --scale-lr false \
        --gpus 0
done

# LDM LR Sweep CST Cosine Eval
tmux new -s lr_sweep_cst_cosine_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export TUNE_RATIOS=0.0,0.3,0.5,0.7,0.9,1.0
for LR in 3e-6 1e-5 3e-5 1e-4; do
    bash run_cdf_experiments.sh ldm-eval-cst \
        --checkpoint-dir "logs/cdf/ldm/lr_sweep_cst_cosine/lr_${LR}" \
        --eval-ratios "$TUNE_RATIOS" \
        --num-fid-samples 1000 \
        --num-sampling-steps 100 \
        --cfg-scale 3.0 \
        --ref-batch "$REF" \
        --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
        --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
        --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
        --output-dir "outputs/cdf/ldm/lr_sweep_cst_cosine/lr_${LR}" \
        --nproc-per-node 1 \
        --per-proc-batch-size 8
done

# LDM FA Sweep CST Cosine
tmux new -s fa_sweep_cst_cosine
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
for FA in 0.03 0.1 0.3; do
    bash run_cdf_experiments.sh ldm-train-cst-fa \
        --data-root "$LDM_DATA" \
        --schedule cosine \
        --ratio-start 0.1 \
        --ratio-end 0.9 \
        --fa-weight "$FA" \
        --fa-boundary-width 10 \
        --logdir "logs/cdf/ldm/fa_sweep_cst_cosine/fa_${FA}" \
        --max-steps 20000 \
        --batch-size 64 \
        --num-workers 0 \
        --learning-rate "$BEST_LR" \
        --scale-lr false \
        --gpus 0
done

# LDM FA Sweep CST Cosine Eval
tmux new -s fa_sweep_cst_cosine_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export TUNE_RATIOS=0.0,0.3,0.5,0.7,0.9,1.0
export BEST_LR=1e-5
for FA in 0.03 0.1 0.3; do
    bash run_cdf_experiments.sh ldm-eval-cst-fa \
        --checkpoint-dir "logs/cdf/ldm/fa_sweep_cst_cosine/fa_${FA}" \
        --eval-ratios "$TUNE_RATIOS" \
        --num-fid-samples 1000 \
        --num-sampling-steps 100 \
        --cfg-scale 3.0 \
        --ref-batch "$REF" \
        --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
        --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
        --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
        --output-dir "outputs/cdf/ldm/fa_sweep_cst_cosine/fa_${FA}" \
        --nproc-per-node 1 \
        --per-proc-batch-size 8
done

# LDM Schedule Ablation
tmux new -s schedule_ablation
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
bash run_cdf_experiments.sh ldm-train-cst-ablation \
    --schedule cosine,cst_v1,cst_v2 \
    --ratio-start 0.1 \
    --ratio-end 0.9 \
    --data-root "$LDM_DATA" \
    --logdir logs/cdf/ldm/schedule_ablation \
    --max-steps 20000 \
    --batch-size 64 \
    --num-workers 0 \
    --learning-rate "$BEST_LR" \
    --scale-lr false \
    --gpus 0

# LDM Schedule Ablation Eval
tmux new -s schedule_ablation_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export TUNE_RATIOS=0.0,0.3,0.5,0.7,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-cst-ablation \
    --checkpoint-dir logs/cdf/ldm/schedule_ablation \
    --eval-ratios "$TUNE_RATIOS" \
    --num-fid-samples 1000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/schedule_ablation \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

0.3 Main LDM Commands 

# LDM NFT
tmux new -s nft
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-nft \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/nft \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# LDM CST Cosine
tmux new -s cst_cosine
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
export BEST_SCHEDULE=cosine
bash run_cdf_experiments.sh ldm-train-cst \
    --data-root "$LDM_DATA" \
    --schedule "$BEST_SCHEDULE" \
    --ratio-start 0.1 \
    --ratio-end 0.9 \
    --logdir logs/cdf/ldm/cst_cosine \
    --max-steps 200000 \
    --batch-size 64 \
    --num-workers 0 \
    --learning-rate "$BEST_LR" \
    --scale-lr false \
    --gpus 0

# LDM CST Cosine Eval
tmux new -s cst_cosine_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-cst \
    --checkpoint-dir logs/cdf/ldm/cst_cosine \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/cst_cosine \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# LDM CST Cosine FA
tmux new -s cst_cosine_fa
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
export BEST_SCHEDULE=cosine
export BEST_FA=0.1
bash run_cdf_experiments.sh ldm-train-cst-fa \
    --data-root "$LDM_DATA" \
    --schedule "$BEST_SCHEDULE" \
    --ratio-start 0.1 \
    --ratio-end 0.9 \
    --fa-weight "$BEST_FA" \
    --fa-boundary-width 10 \
    --logdir logs/cdf/ldm/cst_cosine_fa \
    --max-steps 200000 \
    --batch-size 64 \
    --num-workers 0 \
    --learning-rate "$BEST_LR" \
    --scale-lr false \
    --gpus 0

# LDM CST Cosine FA Eval
tmux new -s cst_cosine_fa_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-cst-fa \
    --checkpoint-dir logs/cdf/ldm/cst_cosine_fa \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/cst_cosine_fa \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# LDM FRFT
tmux new -s frft
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
bash run_cdf_experiments.sh ldm-train-frft \
    --ratios 0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9 \
    --data-root "$LDM_DATA" \
    --logdir logs/cdf/ldm/frft \
    --max-steps 200000 \
    --batch-size 64 \
    --num-workers 0 \
    --learning-rate "$BEST_LR" \
    --scale-lr false \
    --gpus 0

# LDM FRFT Eval
tmux new -s frft_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-frft \
    --checkpoint-dir logs/cdf/ldm/frft \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/frft \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# LDM Senior Only
tmux new -s senior_only
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
bash run_cdf_experiments.sh ldm-train-senior-only \
    --ratios 0.1,0.3,0.5,0.7,0.9 \
    --data-root "$LDM_DATA" \
    --logdir logs/cdf/ldm/senior_only \
    --max-steps 200000 \
    --batch-size 64 \
    --num-workers 0 \
    --learning-rate "$BEST_LR" \
    --scale-lr false \
    --gpus 0

# LDM Senior Only Eval
tmux new -s senior_only_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-senior-only \
    --checkpoint-dir logs/cdf/ldm/senior_only \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/senior_only \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# LDM FMGT
tmux new -s fmgt
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export LDM_DATA=$PREP/ldm
export BEST_LR=1e-5
bash run_cdf_experiments.sh ldm-train-fmgt \
    --ratios 0.1,0.3,0.5,0.7,0.9 \
    --data-root "$LDM_DATA" \
    --logdir logs/cdf/ldm/fmgt \
    --max-steps 200000 \
    --batch-size 64 \
    --num-workers 0 \
    --learning-rate "$BEST_LR" \
    --scale-lr false \
    --gpus 0

# LDM FMGT Eval
tmux new -s fmgt_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh ldm-eval-fmgt \
    --checkpoint-dir logs/cdf/ldm/fmgt \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 3.0 \
    --ref-batch "$REF" \
    --ldm-s-path "$REPO/ldm/pretrained_models/ldm_s.ckpt" \
    --ldm-path "$REPO/ldm/pretrained_models/ldm.ckpt" \
    --ae-ckpt "$REPO/ldm/pretrained_models/vq-f8/model.ckpt" \
    --output-dir outputs/cdf/ldm/fmgt \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

0.4 DiT Commands 

# DiT NFT
tmux new -s dit_nft
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export MODEL_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/models
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export VAE_MODEL=$MODEL_ROOT/sd-vae-ft-ema
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh dit-eval-nft \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 1.5 \
    --solver ddim \
    --ref-batch "$REF" \
    --small-model DiT-S/2 \
    --large-model DiT-XL/2 \
    --small-ckpt "$REPO/dit/pretrained_models/dit_s_256.pt" \
    --large-ckpt "$REPO/dit/pretrained_models/dit_xl_256.pt" \
    --vae-model "$VAE_MODEL" \
    --output-dir outputs/cdf/dit/dit_nft \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# DiT CST Cosine
tmux new -s dit_cst_cosine
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export DIT_DATA=$PREP/imagefolder/train
bash run_cdf_experiments.sh dit-train-cst \
    --data-path "$DIT_DATA" \
    --small-model DiT-S/2 \
    --large-model DiT-XL/2 \
    --small-ckpt "$REPO/dit/pretrained_models/dit_s_256.pt" \
    --large-ckpt "$REPO/dit/pretrained_models/dit_xl_256.pt" \
    --schedule cosine \
    --ratio-start 0.1 \
    --ratio-end 0.9 \
    --results-dir logs/cdf/dit/dit_cst_cosine \
    --epochs 1400 \
    --global-batch-size 256 \
    --learning-rate 1e-5 \
    --nproc-per-node 1

# DiT CST Cosine Eval
tmux new -s dit_cst_cosine_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export MODEL_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/models
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export VAE_MODEL=$MODEL_ROOT/sd-vae-ft-ema
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh dit-eval-cst \
    --checkpoint-dir logs/cdf/dit/dit_cst_cosine \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 1.5 \
    --solver ddim \
    --ref-batch "$REF" \
    --small-model DiT-S/2 \
    --large-model DiT-XL/2 \
    --small-ckpt "$REPO/dit/pretrained_models/dit_s_256.pt" \
    --large-ckpt "$REPO/dit/pretrained_models/dit_xl_256.pt" \
    --vae-model "$VAE_MODEL" \
    --output-dir outputs/cdf/dit/dit_cst_cosine \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# DiT CST Cosine FA
tmux new -s dit_cst_cosine_fa
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export PREP=$DATA_ROOT/imagenet-1k/cdf_prepared
export DIT_DATA=$PREP/imagefolder/train
bash run_cdf_experiments.sh dit-train-cst-fa \
    --data-path "$DIT_DATA" \
    --small-model DiT-S/2 \
    --large-model DiT-XL/2 \
    --small-ckpt "$REPO/dit/pretrained_models/dit_s_256.pt" \
    --large-ckpt "$REPO/dit/pretrained_models/dit_xl_256.pt" \
    --schedule cosine \
    --ratio-start 0.1 \
    --ratio-end 0.9 \
    --fa-weight 0.1 \
    --fa-boundary-width 10 \
    --results-dir logs/cdf/dit/dit_cst_cosine_fa \
    --epochs 1400 \
    --global-batch-size 256 \
    --learning-rate 1e-5 \
    --nproc-per-node 1

# DiT CST Cosine FA Eval
tmux new -s dit_cst_cosine_fa_eval
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export MODEL_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/models
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export VAE_MODEL=$MODEL_ROOT/sd-vae-ft-ema
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh dit-eval-cst-fa \
    --checkpoint-dir logs/cdf/dit/dit_cst_cosine_fa \
    --eval-ratios "$RATIOS" \
    --num-fid-samples 5000 \
    --num-sampling-steps 100 \
    --cfg-scale 1.5 \
    --solver ddim \
    --ref-batch "$REF" \
    --small-model DiT-S/2 \
    --large-model DiT-XL/2 \
    --small-ckpt "$REPO/dit/pretrained_models/dit_s_256.pt" \
    --large-ckpt "$REPO/dit/pretrained_models/dit_xl_256.pt" \
    --vae-model "$VAE_MODEL" \
    --output-dir outputs/cdf/dit/dit_cst_cosine_fa \
    --nproc-per-node 1 \
    --per-proc-batch-size 8

# DiT Sampler Ablation
tmux new -s dit_sampler_ablation
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export MODEL_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/models
export REF=$REPO/assets/fid_stats/VIRTUAL_imagenet256_labeled.npz
export VAE_MODEL=$MODEL_ROOT/sd-vae-ft-ema
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
for solver in ddpm ddim dpm-solver++; do
    bash run_cdf_experiments.sh dit-eval-cst \
        --checkpoint-dir logs/cdf/dit/dit_cst_cosine \
        --eval-ratios "$RATIOS" \
        --num-fid-samples 5000 \
        --num-sampling-steps 50 \
        --cfg-scale 1.5 \
        --solver "$solver" \
        --ref-batch "$REF" \
        --small-model DiT-S/2 \
        --large-model DiT-XL/2 \
        --small-ckpt "$REPO/dit/pretrained_models/dit_s_256.pt" \
        --large-ckpt "$REPO/dit/pretrained_models/dit_xl_256.pt" \
        --vae-model "$VAE_MODEL" \
        --output-dir "outputs/cdf/dit/dit_sampler_ablation_${solver}_50" \
        --nproc-per-node 1 \
        --per-proc-batch-size 8
done

0.5 SD / COCO Commands

The repository currently has COCO evaluation for SD v1.4 + BK-SDM Tiny, DeepCache, and ToMe. The training entry sd-train-cst exists for SDXL/SSD-1B in sd/train_sdxl_tstitch.py; it is not the SD v1.4 + BK-SDM Tiny Table 2 training setup.

# SD NFT
tmux new -s sd_nft
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export COCO_PROMPTS=$DATA_ROOT/coco/annotations/captions_val2014.json
export COCO_REF=$REPO/assets/fid_stats/coco_val2014_256.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh sd-eval-nft \
    --pipeline sd \
    --prompts "$COCO_PROMPTS" \
    --ref-batch "$COCO_REF" \
    --output-dir outputs/cdf/sd/nft_fixed \
    --limit 5000 \
    --ratios "$RATIOS" \
    --height 256 \
    --width 256 \
    --steps 50 \
    --guidance-scale 7.5

# SD DeepCache
tmux new -s sd_deepcache
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export COCO_PROMPTS=$DATA_ROOT/coco/annotations/captions_val2014.json
export COCO_REF=$REPO/assets/fid_stats/coco_val2014_256.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh sd-eval-deepcache \
    --pipeline sd \
    --prompts "$COCO_PROMPTS" \
    --ref-batch "$COCO_REF" \
    --output-dir outputs/cdf/sd/deepcache_i3_fixed \
    --limit 5000 \
    --ratios "$RATIOS" \
    --height 256 \
    --width 256 \
    --steps 50 \
    --guidance-scale 7.5 \
    --deepcache-interval 3

# SD ToMe
tmux new -s sd_tome
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export REPO=$PWD
export DATA_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/datas
export COCO_PROMPTS=$DATA_ROOT/coco/annotations/captions_val2014.json
export COCO_REF=$REPO/assets/fid_stats/coco_val2014_256.npz
export RATIOS=0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0
bash run_cdf_experiments.sh sd-eval-cst \
    --pipeline sd \
    --prompts "$COCO_PROMPTS" \
    --ref-batch "$COCO_REF" \
    --output-dir outputs/cdf/sd/tome_05_fixed \
    --limit 5000 \
    --ratios "$RATIOS" \
    --height 256 \
    --width 256 \
    --steps 50 \
    --guidance-scale 7.5 \
    --tome-ratio 0.5

# Optional SDXL CST Cosine Training
tmux new -s sdxl_cst_cosine
cd /inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/T-Stitch
export MODEL_ROOT=/inspire/hdd/project/intelligentcreativedesign/dangshengqi-253114050252/z-anna/models
bash run_cdf_experiments.sh sd-train-cst \
    --pretrained_model_name_or_path "$MODEL_ROOT/stable-diffusion-xl-base-1.0" \
    --small_model_name_or_path "$MODEL_ROOT/SSD-1B" \
    --train_data_dir <image_text_train_dir> \
    --metadata_file <metadata.jsonl> \
    --output_dir logs/cdf/sd/sdxl_cst_cosine \
    --resolution 1024 \
    --ratio 0.9 \
    --ratio_schedule cosine \
    --ratio_start 0.1 \
    --ratio_schedule_steps 100000 \
    --train_mode junior \
    --max_train_steps 100000 \
    --train_batch_size 1 \
    --learning_rate 1e-5

1. Notation

Let T be the full DDPM training horizon. In the main ImageNet experiments, T = 1000.

Let r in [0, 1] be the junior ratio, i.e. the fraction of denoising responsibility assigned to the junior model. CDF uses the same coarse-to-fine inference direction as T-Stitch:

t_zeta(r) = floor((1 - r) * T)
junior interval = {t | t_zeta(r) <= t < T}
senior interval = {t | 0 <= t < t_zeta(r)}

With T = 1000, r = 0.1 gives t_zeta = 900, while r = 0.9 gives t_zeta = 100. This is the exact schedule expected by the paper text: training moves from short junior responsibility to long junior responsibility.

During inference, sampling starts from high-noise timesteps and moves toward low-noise timesteps. The junior model handles early high-noise/coarse steps, and the senior model handles late low-noise/refinement steps.

2. CDF Inference

Given a junior denoiser eps_j and senior denoiser eps_s, CDF inference selects the model by timestep:

eps_CDF(x_t, t, y; r) =
  eps_j(x_t, t, y), if t >= t_zeta(r)
  eps_s(x_t, t, y), if t <  t_zeta(r)

Meaning:

  • r = 0.0: pure senior model.
  • r = 1.0: pure junior model.
  • 0.1 <= r <= 0.9: collaborative CDF/T-Stitch trajectory.
  • NFT is exactly this inference rule without any CDF fine-tuning.

Default evaluation ratios:

0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0

3. Training Objectives

3.1 Standard Diffusion Loss

All model families keep their native diffusion objective. For LDM and DiT ImageNet experiments, this is epsilon prediction:

x_t = sqrt(alpha_bar_t) * x_0 + sqrt(1 - alpha_bar_t) * eps
L_diff = E || eps - eps_theta(x_t, t, y) ||_2^2

For SD/SDXL, the same rule applies to the pipeline's native prediction type:

target =
  eps, if prediction_type == epsilon
  v,   if prediction_type == v_prediction

3.2 Fixed-Ratio Fine-Tuning

FRFT trains the junior model for one fixed ratio r0. The senior model is frozen.

t_zeta = floor((1 - r0) * T)
t ~ Uniform({t_zeta, ..., T - 1})
L_FRFT(r0) = E || eps - eps_j(x_t, t, y) ||_2^2

Meaning:

  • One checkpoint is trained for one ratio.
  • To support 9 ratios, train 9 junior checkpoints.
  • This tests whether ratio-specific specialization generalizes.

3.3 Curricular Stitching Training

CST trains one junior model over a moving ratio distribution. The senior model is frozen.

Let s = global_step / total_steps. Let r_min = 0.1, r_max = 0.9.

r(s) = r_min + (r_max - r_min) * g(s)
t_zeta(s) = floor((1 - r(s)) * T)
t ~ Uniform({t_zeta(s), ..., T - 1})
L_CST = E || eps - eps_j(x_t, t, y) ||_2^2

Curriculum variants:

linear:     g(s) = s
cosine:     g(s) = (1 - cos(pi * s)) / 2
cst_v1:     g(s) = sqrt(s)     # large-ratio-biased, harder ratios appear earlier
cst_v2:     g(s) = s^2         # small-ratio-biased, easier ratios last longer
staircase:  g(s) = floor(K * s) / K, default K = 8

Meaning:

  • CST turns ratio into a training variable.
  • The final junior checkpoint should work for all evaluation ratios.
  • cst_v1 and cst_v2 are the main ablation schedules.

3.4 Feature Alignment at the Stitching Boundary

Feature Alignment (FA) aligns junior and senior representations near the handoff boundary.

For a boundary width delta, sample:

t_b ~ Uniform({max(0, t_zeta - delta), ..., min(T - 1, t_zeta + delta)})

Let h_j^ell and h_s^ell be hidden features from the same layer family ell in junior and senior. Because junior and senior can have different channel widths, the alignment uses a channel-invariant feature signature:

phi(h) = normalize(concat(mean_channel(h), std_channel(h)))
L_FA = E || phi(h_j^ell) - stopgrad(phi(h_s^ell)) ||_2^2
L_total = L_CST + lambda_FA * L_FA

Default alignment locations:

LDM / SD: U-Net middle block and last decoder/up block signatures.
DiT: final transformer block token signature.
Fallback: denoiser output signature if a hook location is unavailable.

Meaning:

  • FA does not train the senior model in the default CDF setting.
  • FA only acts around the transition boundary, not across every timestep.
  • The signature avoids requiring equal hidden dimensions across model scales.

3.5 Senior-Only Fine-Tuning

This baseline freezes the junior model and only trains the senior model on the senior interval.

t_zeta = floor((1 - r0) * T)
t ~ Uniform({0, ..., t_zeta - 1})
L_senior_only(r0) = E || eps - eps_s(x_t, t, y) ||_2^2

Meaning:

  • Tests whether improving the large/refinement part alone explains gains.
  • This is not the default CDF deployment strategy.

3.6 Full Model Group Training

FMGT jointly trains junior and senior in their assigned intervals.

t_zeta = floor((1 - r0) * T)
t_j ~ Uniform({t_zeta, ..., T - 1})
t_s ~ Uniform({0, ..., t_zeta - 1})
L_FMGT = E || eps - eps_j(x_tj, t_j, y) ||_2^2
       + E || eps - eps_s(x_ts, t_s, y) ||_2^2

Meaning:

  • This is the high-cost joint-training baseline.
  • It is less deployment-friendly because both checkpoints are updated.

4. Main LDM Experiments

Main setting:

Dataset: ImageNet 256x256
Junior: LDM-S
Senior: LDM
Batch size: 64
Learning rate: 1e-5
DDPM training timesteps: T = 1000
Evaluation samples: 5000
Evaluation metrics: FID, IS, latency, speedup
Sampler: DDIM 100 steps unless specified
CFG scale: 3.0
Hardware report target: one RTX 4090 or the exact GPU used

Resume behavior:

  • LDM training writes checkpoints/last.ckpt with model, optimizer, scheduler, epoch, global step, and callback state.
  • Resume one interrupted run at a time. For fixed-ratio jobs, pass a single ratio matching the checkpoint directory.
  • Resume checks the checkpoint optimizer LR against the current effective LR and fails by default on mismatch.
bash run_cdf_experiments.sh ldm-train-fmgt \
  --ratios 0.1 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/fmgt \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5 \
  --resume logs/cdf/ldm/fmgt/ratio-0.1/<run_name>/checkpoints/last.ckpt

Only use this when intentionally changing LR after resume:

--allow-lr-mismatch true

4.1 NFT / T-Stitch Baseline

No training. Directly evaluate the existing pretrained junior and senior checkpoints.

bash run_cdf_experiments.sh ldm-eval-nft \
  --ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --num-fid-samples 5000 \
  --num-sampling-steps 100 \
  --cfg-scale 3.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/nft

Expected output:

outputs/cdf/ldm/nft/samples/ratio-0.5.npz
outputs/cdf/ldm/nft/metrics.csv

4.2 FRFT Baseline

Train one junior checkpoint per ratio.

bash run_cdf_experiments.sh ldm-train-frft \
  --ratios 0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/frft \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5

Evaluate all FRFT checkpoints on all ratios to measure cross-ratio generalization:

bash run_cdf_experiments.sh ldm-eval-frft \
  --checkpoint-dir logs/cdf/ldm/frft \
  --eval-ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/frft

Expected output:

outputs/cdf/ldm/frft/cross_ratio_metrics.csv

4.3 Senior-Only Baseline

Freeze junior, train senior on the senior interval for each fixed ratio.

bash run_cdf_experiments.sh ldm-train-senior-only \
  --ratios 0.1,0.3,0.5,0.7,0.9 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/senior_only \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh ldm-eval-senior-only \
  --checkpoint-dir logs/cdf/ldm/senior_only \
  --eval-ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/senior_only

4.4 FMGT Baseline

Jointly train junior and senior.

bash run_cdf_experiments.sh ldm-train-fmgt \
  --ratios 0.1,0.3,0.5,0.7,0.9 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/fmgt \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh ldm-eval-fmgt \
  --checkpoint-dir logs/cdf/ldm/fmgt \
  --eval-ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/fmgt

4.5 CST Main Run

Train one junior checkpoint with a progressive curriculum.

bash run_cdf_experiments.sh ldm-train-cst \
  --schedule cosine \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/cst_cosine \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5

Evaluate the one CST checkpoint at all ratios:

bash run_cdf_experiments.sh ldm-eval-cst \
  --checkpoint logs/cdf/ldm/cst_cosine/checkpoints/last.ckpt \
  --eval-ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/cst_cosine

4.6 CST Schedule Ablation

Run several curriculum schedules with all other settings fixed.

bash run_cdf_experiments.sh ldm-train-cst-ablation \
  --schedule linear,cosine,cst_v1,cst_v2,staircase \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/schedule_ablation \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh ldm-eval-cst-ablation \
  --checkpoint-dir logs/cdf/ldm/schedule_ablation \
  --eval-ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/schedule_ablation

4.7 CST + Feature Alignment

Train one junior checkpoint with curriculum and boundary feature alignment.

bash run_cdf_experiments.sh ldm-train-cst-fa \
  --schedule cosine \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --fa-weight 0.1 \
  --fa-boundary-width 10 \
  --data-root <imagenet_train> \
  --logdir logs/cdf/ldm/cst_cosine_fa \
  --max-steps 200000 \
  --batch-size 64 \
  --num-workers 0 \
  --scale-lr false \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh ldm-eval-cst-fa \
  --checkpoint logs/cdf/ldm/cst_cosine_fa/checkpoints/last.ckpt \
  --eval-ratios 0.0,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9,1.0 \
  --ref-batch <adm_imagenet_ref.npz> \
  --output-dir outputs/cdf/ldm/cst_cosine_fa

5. DiT Experiments

Purpose:

  • Provide quantitative evidence beyond LDM.
  • Explicitly connect NFT to T-Stitch under the same DiT samplers.
  • Report FID, IS, latency, and speedup for all ratios.

Default DiT setting:

Dataset: ImageNet 256x256
Junior: DiT-S/2
Senior: DiT-XL/2
Optional medium: DiT-B/2
Sampler: DDIM 100, plus DDPM and DPM-Solver++ ablations
CFG scale: 1.5
Evaluation samples: 5000

5.1 DiT NFT / T-Stitch 

bash run_cdf_experiments.sh dit-eval-nft \
  --small-model DiT-S/2 \
  --large-model DiT-XL/2 \
  --solver ddim \
  --num-sampling-steps 100 \
  --cfg-scale 1.5 \
  --num-fid-samples 5000 \
  --ref-batch <adm_imagenet_ref.npz> \
  --vae-model <local_sd_vae_ft_ema_or_hf_id> \
  --output-dir outputs/cdf/dit/nft_ddim100

5.2 DiT FRFT 

bash run_cdf_experiments.sh dit-train-frft \
  --ratios 0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9 \
  --data-path <imagenet_train> \
  --small-model DiT-S/2 \
  --large-model DiT-XL/2 \
  --logdir logs/cdf/dit/frft \
  --epochs 1400 \
  --global-batch-size 256 \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh dit-eval-frft \
  --checkpoint-dir logs/cdf/dit/frft \
  --solver ddim \
  --num-sampling-steps 100 \
  --ref-batch <adm_imagenet_ref.npz> \
  --vae-model <local_sd_vae_ft_ema_or_hf_id> \
  --output-dir outputs/cdf/dit/frft

5.3 DiT CST and CST + FA 

bash run_cdf_experiments.sh dit-train-cst \
  --schedule cosine \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --data-path <imagenet_train> \
  --small-model DiT-S/2 \
  --large-model DiT-XL/2 \
  --logdir logs/cdf/dit/cst_cosine \
  --epochs 1400 \
  --global-batch-size 256 \
  --learning-rate 1e-5

bash run_cdf_experiments.sh dit-train-cst-fa \
  --schedule cosine \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --fa-weight 0.1 \
  --fa-boundary-width 10 \
  --data-path <imagenet_train> \
  --small-model DiT-S/2 \
  --large-model DiT-XL/2 \
  --logdir logs/cdf/dit/cst_cosine_fa \
  --epochs 1400 \
  --global-batch-size 256 \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh dit-eval-cst \
  --checkpoint logs/cdf/dit/cst_cosine/checkpoints/last.pt \
  --solver ddim \
  --num-sampling-steps 100 \
  --ref-batch <adm_imagenet_ref.npz> \
  --vae-model <local_sd_vae_ft_ema_or_hf_id> \
  --output-dir outputs/cdf/dit/cst_cosine

5.4 DiT Sampler Ablation 

for solver in ddpm ddim dpm-solver++; do
  bash run_cdf_experiments.sh dit-eval-cst \
    --checkpoint logs/cdf/dit/cst_cosine/checkpoints/last.pt \
    --solver "$solver" \
    --num-sampling-steps 50 \
    --ref-batch <adm_imagenet_ref.npz> \
    --vae-model <local_sd_vae_ft_ema_or_hf_id> \
    --output-dir outputs/cdf/dit/cst_cosine_${solver}_50
done

6. Stable Diffusion / Text-to-Image Experiments

Purpose:

  • Address general text-to-image prompts beyond ImageNet.
  • Report FID, IS, and CLIP score on MS-COCO.
  • Test orthogonality with DeepCache.

Default SD setting:

Dataset/prompts: MS-COCO validation captions
Junior: BK-SDM Tiny
Senior: SD v1.4 or stylized SD
Sampler: PNDM 50 steps
Guidance scale: 7.5
Resolution: 256x256 for Table 2-style metrics
Evaluation samples: 5000

6.1 SD NFT / T-Stitch on COCO 

bash run_cdf_experiments.sh sd-eval-nft \
  --prompts <captions_val2014.json> \
  --ref-batch <coco_ref_batch.npz> \
  --output-dir outputs/cdf/sd/nft \
  --limit 5000

6.2 SD CST / CST + FA 

bash run_cdf_experiments.sh sd-train-cst \
  --schedule cosine \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --train-data-dir <image_text_train_dir> \
  --metadata-file <metadata.jsonl> \
  --output-dir logs/cdf/sd/cst_cosine \
  --max-train-steps 100000 \
  --train-batch-size 1 \
  --learning-rate 1e-5

bash run_cdf_experiments.sh sd-train-cst-fa \
  --schedule cosine \
  --ratio-start 0.1 \
  --ratio-end 0.9 \
  --fa-weight 0.1 \
  --fa-boundary-width 2 \
  --train-data-dir <image_text_train_dir> \
  --metadata-file <metadata.jsonl> \
  --output-dir logs/cdf/sd/cst_cosine_fa \
  --max-train-steps 100000 \
  --train-batch-size 1 \
  --learning-rate 1e-5

Evaluate:

bash run_cdf_experiments.sh sd-eval-cst \
  --checkpoint logs/cdf/sd/cst_cosine \
  --prompts <captions_val2014.json> \
  --ref-batch <coco_ref_batch.npz> \
  --output-dir outputs/cdf/sd/cst_cosine \
  --limit 5000

6.3 SD + DeepCache Orthogonal Experiment

Compare:

SD NFT
SD NFT + DeepCache
SD CDF/CST
SD CDF/CST + DeepCache

Run:

bash run_cdf_experiments.sh sd-eval-deepcache \
  --checkpoint logs/cdf/sd/cst_cosine \
  --prompts <captions_val2014.json> \
  --ref-batch <coco_ref_batch.npz> \
  --deepcache-interval 3 \
  --output-dir outputs/cdf/sd/cst_cosine_deepcache \
  --limit 5000

Metrics:

FID, IS, CLIP score, latency, speedup

7. Data Products for Paper Figures

The final implementation should export the following CSV files.

Figure 1: outputs/cdf/ldm/main_tradeoff.csv
  columns: method, ratio, fid, is, latency_sec, speedup, checkpoint

Figure 3: outputs/cdf/training_cost.csv
  columns: method, num_checkpoints, trainable_params, gpu_days, supported_ratios

Figure 4: outputs/cdf/ldm/cross_ratio_generalization.csv
  columns: train_method, train_ratio, eval_ratio, fid, is

Figure 5: outputs/cdf/ldm/cst_convergence.csv
  columns: epoch_or_step, ratio, fid, is

Figure 6: outputs/cdf/ldm/schedule_fa_ablation.csv
  columns: method, schedule, fa_weight, ratio, fid, is, latency_sec

Figure 7: outputs/cdf/efficiency_by_arch_sampler.csv
  columns: arch, sampler, num_steps, ratio, latency_sec, speedup

Figure 8: outputs/cdf/sd/deepcache_compatibility.csv
  columns: method, ratio, fid, kid, is, clip_score, latency_sec

8. Implementation Checklist

The current repository already has partial T-Stitch training and sampling. The following items must be implemented or verified before the commands above are considered runnable end-to-end:

  • Add explicit train_mode values: junior, senior, joint.
  • Add exact schedule functions: fixed, linear, cosine, cst_v1, cst_v2, staircase.
  • Sample training timesteps from the correct interval for each training mode.
  • Add boundary Feature Alignment with documented layer hooks and output fallback.
  • Add LDM runner commands for NFT, FRFT, senior-only, FMGT, CST, CST+FA.
  • Add all-ratio evaluation that emits .npz samples and metrics.csv.
  • Align DiT training flags with the same CDF names and schedules.
  • Align SD/SDXL training flags with the same CDF names and schedules.
  • Add COCO prompt generation, CLIP score, FID/IS aggregation, and DeepCache toggles to the CDF runner.