UCGM
The model was presented in the paper Unified Continuous Generative Models.
Model source
- Kind
- Other
- File size
- 1.50 KB–2.84 GB
- Version
- vRevision 7039016
- Updated
- Sep 15, 2026
- Downloads
- 0
- Source
- Hugging Face
Source description
Unified Continuous Generative Models
The model was presented in the paper Unified Continuous Generative Models.
Paper Abstract
Recent advances in continuous generative models, including multi-step approaches like diffusion and flow-matching (typically requiring 8-1000 sampling steps) and few-step methods such as consistency models (typically 1-8 steps), have demonstrated impressive generative performance. However, existing work often treats these approaches as distinct paradigms, resulting in separate training and sampling methodologies. We introduce a unified framework for training, sampling, and analyzing these models. Our implementation, the Unified Continuous Generative Models Trainer and Sampler (UCGM-{T,S}), achieves state-of-the-art (SOTA) performance. For example, on ImageNet 256x256 using a 675M diffusion transformer, UCGM-T trains a multi-step model achieving 1.30 FID in 20 steps and a few-step model reaching 1.42 FID in just 2 steps. Additionally, applying UCGM-S to a pre-trained model (previously 1.26 FID at 250 steps) improves performance to 1.06 FID in only 40 steps. Code is available at: https://github.com/LINs-lab/UCGM.
Code
The code for this model is available on Github: https://github.com/LINs-lab/UCGM
Sources
1 source
Sources
1 sourceVerified Sep 15
Model artifacts
18 artifacts
Model artifacts
18 artifactsbuffers/vaes/stat/e2evavae_f16d32_256.pt
pt · 1.75 KB · SHA-256 bf237612d372…e186 · Hugging Face
Downloadbuffers/vaes/stat/edm2_sdvae_f8c4_512.pt
pt · 1.50 KB · SHA-256 79acb6778ab6…6cbc · Hugging Face
Downloadbuffers/vaes/stat/vavae_f16d32_256.pt
pt · 1.75 KB · SHA-256 fb96ba21f7f4…5ee7 · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k256_ddt_xl_sdvae_relinear/checkpoints/0000000.pt
pt · 2.51 GB · SHA-256 b5c9bba15f6f…2339 · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k256_lightningdit_xl_vavae_relinear/checkpoints/0000000.pt
pt · 2.52 GB · SHA-256 414e8389a374…1def · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k256_sit_xl_repae_linear/checkpoints/0000000.pt
pt · 2.52 GB · SHA-256 f24157ac9c7e…850f · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k512_ddt_xl_sdvae_relinear/checkpoints/0000000.pt
pt · 2.51 GB · SHA-256 9bc4a665acf1…41a7 · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k512_edm2_l_sdvae_edm/checkpoints/0000000.pt
pt · 1.45 GB · SHA-256 c6a85db972d3…8783 · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k512_edm2_s_sdvae_edm/checkpoints/0000000.pt
pt · 535 MB · SHA-256 515bf0a91213…c653 · Hugging Face
Downloadoutputs/sampling_multi_steps/in1k512_edm2_xxl_sdvae_edm/checkpoints/0000000.pt
pt · 2.84 GB · SHA-256 f00c1d4c8d11…4b85 · Hugging Face
Downloadoutputs/tuning_few_steps/in1k256_ddt_xl_sdvae_relinear/checkpoints/0000800.pt
pt · 2.51 GB · SHA-256 70e2743dac87…81ee · Hugging Face
Downloadoutputs/tuning_few_steps/in1k256_lightningdit_xl_vavae_relinear/checkpoints/0000800.pt
pt · 2.52 GB · SHA-256 572550a9c03f…f2eb · Hugging Face
Downloadoutputs/tuning_few_steps/in1k256_sit_xl_repae_linear/checkpoints/0000500.pt
pt · 2.52 GB · SHA-256 b281c2069903…949e · Hugging Face
Downloadoutputs/tuning_few_steps/in1k256_sit_xl_sdvae_linear/checkpoints/0000800.pt
pt · 2.52 GB · SHA-256 e5caa5cd4b7d…0e5f · Hugging Face
DownloadSource excerpts
2 excerpts
Source excerpts
2 excerpts--- license: apache-2.0 pipeline_tag: unconditional-image-generation library_name: diffusers --- # Unified Continuous Generative Models The model was presented in the paper [Unified Continuous Generative Models](https://huggingface.co/papers/2505.07447). # Paper Abstract Recent advances in continuous generative models, including multi-step approaches like diffusion and flow-matching (typically requiring 8-1000 sampling steps) and few-step methods such as consistency models (typically 1-8 steps), have demonstrated impressive generative performance. However, existing work often treats these approaches as distinct paradigms, resulting in separate training and sampling methodologies. We introduce a unified framework for training, sampling, and analyzing these models. Our implementation, the Unified Continuous Generative Models Trainer and Sampler (UCGM-{T,S}), achieves state-of-the-art (SOTA) performance. For example, on ImageNet 256x256 using a 675M diffusion transformer, UCGM-T trains a multi-step model achieving 1.30 FID in 20 steps and a few-step model reaching 1.42 FID in just 2 steps. Additionally, applying UCGM-S to a pre-trained model (previously 1.26 FID at 250 steps) improves performance to 1.06 FID in only 40 steps. Code is available at: https://github.com/LINs-lab/UCGM. # Code The code for this model is available on Github: https://github.com/LINs-lab/UCGM
Source context: 0 downloads · 6 likes · Pipeline unconditional-image-generation · Library diffusers · Repo sp12138sp/UCGM
