| --- |
| license: cc-by-nc-sa-4.0 |
| pipeline_tag: text-to-image |
| tags: |
| - pytorch |
| - image-generation |
| - class-conditional |
| - diffusion-distillation |
| - imagenet |
| --- |
| |
| # Consistent Distribution Matching for Data-Free Diffusion Distillation |
|
|
| This is the **official PyTorch** checkpoint repository for [Consistent Distribution Matching for Data-Free Diffusion Distillation](https://arxiv.org/abs/2610.09221). |
|
|
| Project page: https://consistentdmd.github.io/ |
|
|
| Code: https://github.com/felix-yuxiang/CDMD |
|
|
| This release contains one **LightningDiT-teacher CDMD student for class-conditional ImageNet 256×256 generation**. The same student produces the project page's **2.04 / 1.62 / 1.37 FID** results with **1 / 2 / 4 Euler steps**. Each step is one network evaluation; inference uses no classifier-free guidance pair or teacher model. |
|
|
| ## Checkpoint and statistics |
|
|
| | File | Size | Description | |
| |---|---:|---| |
| | [consistentdmd-ldit-best.pt](consistentdmd-ldit-best.pt) | 2.73 GB | Reporting PostHocEMA index **1**, `sigma_rel=0.2`. Compact FP32 student weights and portable metadata, compatible with the public code. | |
| | [latents_stats.pt](latents_stats.pt) | 1.8 KB | Matching VA-VAE latent mean and standard deviation, required for image generation. | |
| | [fid-50k-256.npz](fid-50k-256.npz) | 33.6 MB | ImageNet reference statistics for FID evaluation with the pinned metric implementation. | |
|
|
| The statistics for this model are `.pt` and `.npz` files; no statistics `.pkl` is required. The compact checkpoint contains no optimizer, teacher, or training history and cannot resume training. Its implementation configuration is `LightningDiT-XL/1`, `dmf_depth=10`, with 32-channel 16×16 VA-VAE latents and 1,000 classes. |
|
|
| The required **VA-VAE decoder** is downloaded separately by the code's asset script. The two statistics files are unmodified copies from [Peterande/FACM at revision b7b5fee6d3f43347883d0fe6f9f357e3068375b8](https://huggingface.co/Peterande/FACM/tree/b7b5fee6d3f43347883d0fe6f9f357e3068375b8). [assets.json](assets.json) records their upstream revision, sizes, and SHA-256 hashes, together with the decoder and training-only teacher identities. |
|
|
| ## Download and sample |
|
|
| Set up the environment from the [code repository](https://github.com/felix-yuxiang/CDMD#-environment). Run the following commands from its root: |
|
|
| ```bash |
| hf download fyxfelixfu/consistentDMD \ |
| consistentdmd-ldit-best.pt latents_stats.pt fid-50k-256.npz \ |
| --local-dir cache |
| |
| # Downloads the VA-VAE and verifies the matching inference assets. |
| python scripts/download_assets.py |
| |
| # Generate 64 images with one network evaluation per image. |
| bash scripts/test.sh --sampling-steps 1 --num-samples 64 --skip-metrics \ |
| --output-dir samples/preview |
| ``` |
|
|
| Use `--sampling-steps 2` or `--sampling-steps 4` for the other project-page settings. The public sampler also supports 8, 16, 32, and 64 Euler steps. Inference requires a BF16-capable NVIDIA GPU. The decoder runs in FP32 and the student runs in BF16, while checkpoint tensors remain FP32. |
|
|
| ## Reproduce FID |
|
|
| Use the public repository's pinned requirements, including the [LTH14 torch-fidelity fork](https://github.com/LTH14/torch-fidelity/tree/bfe72b995473e4fee9351705d7b005f17442a4ca). The recorded setup uses two A100 80GB GPUs, seed 0, 50,000 images, and batch size 32 per GPU: |
|
|
| ```bash |
| torchrun --standalone --nproc_per_node=2 test.py \ |
| --ckpt-path cache/consistentdmd-ldit-best.pt \ |
| --sampling-steps 1 --num-samples 50000 \ |
| --batch-size 32 --decode-batch-size 32 \ |
| --output-dir samples/euler-1 |
| ``` |
|
|
| Repeat with `--sampling-steps 2` and `4`, using a fresh output directory for each run. |
|
|
| | Euler steps / NFE | Recorded FID-50K ↓ | |
| |---:|---:| |
| | 1 | **2.0358** | |
| | 2 | **1.6185** | |
| | 4 | **1.3711** | |
| | 8 | 1.2764 | |
|
|
| [results.json](results.json) contains the evaluation protocol and results. |
|
|
| ## Integrity and provenance |
|
|
| The release was checked against the frozen reporting EMA: **all 419 state tensors are exactly equal**, with no precision conversion. The checkpoint SHA-256 is: |
|
|
| ```text |
| 0eddca248af6cfaa21e5b9b32d0c02329923e5066d1bb2b140fa6827c061c1d4 |
| ``` |
|
|
| See [consistentdmd-ldit-best.pt.json](consistentdmd-ldit-best.pt.json) for model and EMA metadata, [consistentdmd-ldit-best.pt.sha256](consistentdmd-ldit-best.pt.sha256) for its checksum, and [SHA256SUMS](SHA256SUMS) for all release files. This is an exact copy of the validated export under the filename expected by the public code. |
|
|
| ## License and acknowledgements |
|
|
| The CDMD checkpoint follows the project's [CC BY-NC-SA 4.0 license](LICENSE.md). The architecture and decoder build on Lightning-DiT and DiT. Upstream assets retain their respective terms; see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md). |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{fu2026cdmd, |
| title = {Consistent Distribution Matching for Data-Free Diffusion Distillation}, |
| author = {Fu, Yuxiang and Yan, Qi and Wu, Zike and Zhang, Yongxing and |
| Abolmaesumi, Purang and Wang, Lele and Liao, Renjie}, |
| journal = {arXiv preprint arXiv:2610.09221}, |
| year = {2026}, |
| url = {https://arxiv.org/abs/2610.09221} |
| } |
| ``` |
