ECT: Consistency Models Made Easy

Pytorch implementation for Easy Consistency Tuning (ECT).

ECT unlocks state-of-the-art (SoTA) few-step generative abilities through a simple yet principled approach. With minimal tuning costs, ECT demonstrates promising early results and scales with training FLOPs and model sizes.

Try your own Consistency Models! You only need to fine-tune a bit. :D

Introduction

This repository is organized in a multi-branch structure, with each branch offering a minimal implementation for a specific purpose. The current branches support the following training protocols:

• main: ECT on CIFAR-10. Best for understanding CMs and fast prototyping.
• amp: Mixed Precision training with GradScaler on CIFAR-10.
• imgnet: ECT ImageNet 64x64.

⭐ Update ⭐

Baking more in the oven. 🙃

• 2024.10.12 - Add ECT code for ImgNet 64x64. Switch to the imgnet branch: git checkout imgnet.
• 2024.09.23 - Add Gradscaler for Mixed Precision Training. To use mixed precision with GradScaler, switch to the amp branch: git checkout amp.
• 2024.04.27 - Upgrade environment to Pytorch 2.3.0.
• 2024.04.12 - ECMs can now surpass SoTA GANs using 1 model step and SoTA Diffusion Models using 2 model steps on CIFAR10. Checkpoints available.

Environment

You can run the following command to set up the Python environment through conda. Pytorch 2.3.0 and Python 3.9.18 will be installed.

conda env create -f env.yml

Datasets

Prepare the dataset in the EDM's format. See a reference here.

Training

Run the following command to tune your SoTA 2-step ECM and match Consistency Distillation (CD) within 1 A100 GPU hour.

bash run_ecm_1hour.sh 1 <PORT> --desc bs128.1hour

Run the following command to run ECT at batch size 128 and 200k iterations. NGPUs=2/4 is recommended.

bash run_ecm.sh <NGPUs> <PORT> --desc bs128.200k

Replace NGPUs and PORT with the number of GPUs used for training and the port number for DDP sync.

Half Precision Training

In this branch, we enable fp16 training with AMP GradScaler for more stable training dynamics. To enable fp16 and GradScaler, add the following arguments to your script:

bash run_ecm_1hour.sh 1 <PORT> --desc bs128.1hour --fp16=True --enable_amp=True

For more information, please refer to this PR. Full support for Automatic Mixed Precision (AMP) will be added later.

Evaluation

Run the following command to calculate FID of a pretrained checkpoint.

bash eval_ecm.sh <NGPUs> <PORT> --resume <CKPT_PATH> 

Fixed-seed evaluation

Role D sampling uses seeds 0-63 and verifies that work-group sizes 8 and 16 produce pixel-identical results for both NFE=1 and NFE=2. It also repeats each configuration to verify deterministic output:

See docs/EVALUATION_PROTOCOL.md for the complete protocol, metadata requirements, checkpoint-isolated output layout, and metric boundary.

bash scripts/sample_checkpoint.sh <CKPT_PATH> \
  --outdir /mnt/ect_project/evaluations \
  --seeds 0-63 --nfe 1 2 --mid-t 0.821 \
  --work-group-size 8 --verify-work-group-size 16 \
  --precision fp32

The output is isolated under <outdir>/<checkpoint-stem>-<sha256-prefix>/ and contains one 8x8 grid per NFE, metadata.json, sha256_manifest.txt, and individual seed images. Keep the individual PNG files under /mnt; only commit the grids, metadata, manifest, scripts, and tests.

The unified metric entry point supports explicit one-step or two-step evaluation through --nfe=1 or --nfe=2:

bash scripts/evaluate_checkpoint.sh 1 <PORT> <CKPT_PATH> \
  --outdir ct-evals --data datasets/cifar10-32x32.zip \
  --nfe=2 --mid_t=0.821 --metrics=fid50k_full

The frozen three-training-seed final comparison uses explicit per-sample seeds, KID-5k as the primary proxy, FID-5k as an auxiliary proxy, and a method-blinded A/B ballot. See docs/FINAL_PERFORMANCE_EVALUATION.md. These 5k-sample results are not standard FID-50k benchmarks.

Generative Performance

FID Evaluation

Taking the models trained by ECT as ECM, we compare ECMs' unconditional image generation capabilities with SoTA generative models on the CIFAR10 dataset, including popular diffusion models w/ advanced samplers, diffusion distillations, and consistency models on the CIFAR10 dataset.

Method	FID	NFE	Model	Params	Batch Size	Schedule
Score SDE	2.38	2000	NCSN++	56.4M	128	~1600k
Score SDE-deep	2.20	2000	NCSN++ (2 $\times$ depth)	> 100M	128	~1600k
EDM	2.01	35	DDPM++	56.4M	512	400k
PD	8.34	1	DDPM++	56.4M	512	800k
Diff-Instruct	4.53	1	DDPM++	56.4M	512	800k
CD (LPIPS)	3.55	1	NCSN++	56.4M	512	800k
CD (LPIPS)	2.93	2	NCSN++	56.4M	512	800k
iCT-deep	2.51	1	NCSN++ (2 $\times$ depth)	> 100M	1024	400k
iCT-deep	2.24	2	NCSN++ (2 $\times$ depth)	> 100M	1024	400k
ECM (100k)	4.54	1	DDPM++	55.7M	128	100k
ECM (200k)	3.86	1	DDPM++	55.7M	128	200k
ECM (400k)	3.60	1	DDPM++	55.7M	128	400k
ECM (100k)	2.20	2	DDPM++	55.7M	128	100k
ECM (200k)	2.15	2	DDPM++	55.7M	128	200k
ECM (400k)	2.11	2	DDPM++	55.7M	128	400k

$\mathrm{FD}_{\text{DINOv2}}$ Evaluation

Since DINOv2 could produce evaluation better aligned with human vision, we evaluate the image fidelity using Fréchet Distance in the latent space of SoTA open-source representation model DINOv2, denoted as $\mathrm{FD}_{\text{DINOv2}}$.

---

Using dgm-eval, we have $\mathrm{FD}_{\text{DINOv2}}$ against SoTA Diffusion Models and GANs.

Method	$\mathrm{FD}_{\text{DINOv2}}$	NFE
EDM	145.20	35
StyleGAN-XL	204.60	1
ECM	198.51	1
ECM	128.63	2

Without combining with other generative mechanisms like GANs or diffusion distillation like Score Distillation, ECT is capable of generating high-quality samples much faster than SoTA diffusion models and much better than SoTA GANs SoTA Diffusion Models and GANs.

Checkpoints

• CIFAR10 $\mathrm{FD}_{\text{DINOv2}}$ checkpoint.

Contact

Feel free to drop me an email at zhengyanggeng@gmail.com if you have additional questions or are interested in collaboration. You can find me on Twitter or WeChat.

Citation

@article{ect,
  title={Consistency Models Made Easy},
  author={Geng, Zhengyang and Pokle, Ashwini and Luo, William and Lin, Justin and Kolter, J Zico},
  journal={arXiv preprint arXiv:2406.14548},
  year={2024}
}