LEAP

Official implementation of LEAP: Accelerating ODE-based Diffusion Sampling with Optimal Linear Leaps.

ACM Multimedia 2026 · Paper · Citation

LEAP accelerates diffusion sampling by alternating learned linear predictions with model evaluations. This repository supports SD3.5-Medium, FLUX.1-dev and TRELLIS.2, with calibration, generation and evaluation tools.

Installation

conda env create -f environment.yml
conda activate LEAP
python -m pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu124
python -m pip install -r requirements.txt

The environment uses Python 3.11, PyTorch 2.5.1 and Diffusers 0.35.1. See installation for model access, evaluation dependencies and cache setup.

Model	Experiment config	512-resolution smoke config
SD3.5-Medium	configs/sd35.yaml	configs/smoke.yaml
FLUX.1-dev	configs/flux.yaml	configs/flux-smoke.yaml

Use --model-path /path/to/snapshot --local-files-only to load local weights, or --offload to reduce GPU memory use.

Quick start

Generate four images with the included SD3.5 / Euler / NFE 10 coefficient bank. Run all commands from the repository root:

leap generate --config configs/sd35.yaml \
  --prompts benchmarks/prompts/drawbench.json --count 4 --seed 10000 \
  --nfe 10 --method leap --bank coefficients/sd35-euler-nfe10.json \
  --output outputs/sd35-demo

Each run saves images and a run.json containing prompts, seeds, model settings, measured NFE and timing.

Run experiments

Calibrate on 32 reference trajectories, then generate paired baseline and LEAP images for 200 DrawBench prompts:

leap benchmark --config configs/sd35.yaml \
  --prompts benchmarks/prompts/drawbench.json --count 200 \
  --nfes 10 15 20 --solvers euler dpmpp \
  --reference-count 32 --reference-seed 0 --seed 10000 \
  --output outputs/sd35-drawbench

Each solver and NFE budget gets its own coefficient bank. Supported solvers are euler, dpmpp, unipc and ipndm. For FLUX, use configs/flux.yaml and a separate output directory; see the FLUX guide. For a quick smoke test, use the corresponding smoke config with --count 2 --reference-count 2 --nfes 10 --solvers euler.

After installing the evaluation dependencies, score every run with HPSv2.1, ImageReward and CLIP:

for solver in euler dpmpp; do
  for nfe in 10 15 20; do
    for method in baseline leap; do
      run="outputs/sd35-drawbench/${solver}-nfe${nfe}/${method}"
      leap score --run "$run" --output "${run}-scores.json"
    done
  done
done

Use a new output directory for each experiment. The benchmark guide covers individual runs, generation shards and metric settings.

Calibrate a coefficient bank

leap calibrate --config configs/sd35.yaml \
  --prompts benchmarks/prompts/drawbench.json --count 32 --seed 0 \
  --nfe 10 --output outputs/sd35-euler-nfe10/bank.json

Pass the resulting file to leap generate --bank. Calibration and generation must use the same model, solver, NFE and configuration. With the default window of 5, NFE 10 produces 15 solver steps; leap plan --nfe 10 displays the prediction/correction schedule.

Documentation

• Environment and deployment
• Benchmark commands and metrics
• FLUX.1-dev experiments
• TRELLIS.2 experiments
• Regression and sampling details
• Tests and validation results
• Acknowledgments and dependencies