Introducing
ECAD
Evolutionary Caching to Accelerate Your Off‑the‑Shelf Diffusion Model
Anirud AggarwalAbhinav ShrivastavaMatthew Gwilliam
ICLR 2026  ·  UNIVERSITY OF MARYLAND
Off-the-shelf · No retraining
A new Pareto frontier for diffusion caching.
Speed up
1.97–2.90×
faster than no caching
Better quality, faster
0.00↓ COCO FID
at 2.58× vs. prior SOTA 2.35×
How it works
ECAD discovers when, where, and what to cache.
1 · Caching schedulesEach schedule is a binary tensor over timestep, layer, and component (self‑attention, cross‑attention, feed‑forward). Over 21680 possibilities.
How it works
ECAD discovers when, where, and what to cache.
2 · Generate imagesEvery schedule in the population renders images using the calibration prompts, caching features where specificed by the schedule.
How it works
ECAD discovers when, where, and what to cache.
3 · Score images + measure speedEach image is scored for quality (ImageReward), and each schedule’s compute cost (MACs) is measured.
How it works
ECAD discovers when, where, and what to cache.
4 · Genetic algorithmNSGA‑II mates and mutates the best schedules into the next generation.
How it works
ECAD discovers when, where, and what to cache.
↻ repeat for many generations. No gradients or weight updates.
Evolution in action
Higher quality at fixed compute.
At a fixed 2.2 TMACs budget, ECAD finds schedules achieving better image quality than the unaccelerated model.
Generation
0
ImageReward
0.75
Speed up · fixed
2.60×
Evolution in action
Lower compute at fixed quality.
At baseline quality, ECAD reduces compute from 5.7 to 1.5 TMACs.
Generation
0
ImageReward · fixed
0.90
Speed up
1.15×
Quality over generations
Faster and cleaner every generation.
Gen 0
Gen 50
Gen 150
Gen 300
Gen 575
Baseline 1×
One method, many models
Works on any DiT.
PixArt-α
0.00×
PixArt-Sigma sample
PixArt-Σ
0.00×
FLUX.1-dev
0.00×
Transfers across resolutions
0.00×
Schedules found at 256² run at 1024² with no re-search (FLUX.1-dev).
Plug and play
One schedule search.
Reusable forever.
0.00×
on FLUX.1-dev — zero retraining
Composes with
Distillation Quantization Pruning torch.compile
Competitive schedules in ~50 generations
~44 GPU-hours of search (PixArt-α)
Code + learned schedules released
ECAD
PixArt-αPixArt-ΣFLUX.1-dev
Paper · Code · Interactive demo
research.aniaggarwal.com/ecad