MearcAll papers

arXiv 2105.05233 · cs.LG cs.AI cs.CV

Diffusion Models Beat GANs on Image Synthesis

Prafulla Dhariwal, Alex Nichol

Claims were extracted automatically by reducer-1 on 2026-11-12. The paper's license does not allow republishing it, so long claims are shortened to a quotation. Read the full text on arXiv.

No reproduction recordedclm_3d08a6b5
We show that diffusion models can achieve image sample quality superior to the current state-of-the-art generative models.

Abstract · page 1 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_93211d1b
We achieve an FID of 2.97 on ImageNet 128 ×128, 4.59 on ImageNet 256×256, and 7.72 on ImageNet 512 ×512, and we match BigGAN-deep even with as few as 25 … (excerpt)

Abstract · page 1 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_331d7da3
Finally, we find that classifier guidance combines well with upsampling diffusion models, further improving FID to 3.94 on ImageNet 256×256 and 3.85 on ImageNet 512×512.

Abstract · page 1 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_f7f5df81
With these improvements, we achieve a new state-of-the-art, surpassing GANs on several different metrics and datasets.

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_35d66ff3
In Section 2, we give a brief background of diffusion models based on Ho et al. [ 25] and the improvements from Nichol and Dhariwal [ 43] and Song et … (excerpt)

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_d4773615
In Section 3, we introduce simple architecture improvements that give a substantial boost to FID.

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_4291c8d8
We find that a single hyperparameter, the scale of the classifier gradients, can be tuned to trade off diversity for fidelity, and we can increase this gradient scale factor by … (excerpt)

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_de01cb50
Finally, in Section 5 we show that models with our improved architecture achieve state-of-the-art on unconditional image synthesis tasks, and with classifier guidance achieve state-ofthe-art on conditional image synthesis.

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_f0b54bbf
We also compare our improved models to upsampling stacks, finding that the two approaches give complementary improvements and that combining them gives the best results on ImageNet 256×256 and 512×512.

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_85ab644e
We show the same result on ImageNet 128×128, finding that architecture can indeed give a substantial boost to sample quality on much larger and more diverse datasets at a higher … (excerpt)

Architecture Improvements · page 4 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_055da973
We observe in Figure 2 that while increased depth helps performance, it increases training time and takes longer to reach the same performance as a wider model, so we opt … (excerpt)

Architecture Improvements · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_89a26ac2
Table 2 shows our results, indicating that more heads or fewer channels per head improves FID.

Architecture Improvements · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_2c2a7ccd
In Table 3, we explicitly ablate this choice, and find that the adaptive group normalization layer indeed improved FID.

Adaptive Group Normalization · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_8bbe7201
In the rest of the paper, we use this final improved model architecture as our default: variable width with 2 residual blocks per resolution, multiple heads with 64 channels per … (excerpt)

Adaptive Group Normalization · page 6 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_7c3922d8
Table 4 shows that the sample quality of both unconditional and conditional models can be greatly improved by classifier guidance.

Conditional Sampling for DDIM · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_a58bbb70
Table 4 also shows that classifier guidance improves precision at the cost of recall, thus introducing a trade-off in sample fidelity versus diversity.

Conditional Sampling for DDIM · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_887ecf9e
To evaluate our improved model architecture on unconditional image generation, we train separate diffusion models on three LSUN [ 71] classes: bedroom, horse, and cat.

Results · page 9 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_34e53196
With the improved architecture, we already obtain state-of-the-art image generation on LSUN and ImageNet 64×64.

State-of-the-art Image Synthesis · page 10 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_33a9d30c
For higher resolution ImageNet, we observe that classifier guidance allows our models to substantially outperform the best GANs.

State-of-the-art Image Synthesis · page 10 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_0999f08a
In Table 6, we show that guidance and upsampling improve sample quality along different axes.

Comparison to Upsampling · page 10 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_e6abb2d9
For upsampling, we use the upsampling stack from Nichol and Dhariwal [43] combined with our architecture improvements, which we refer to as ADM-U.

Comparison to Upsampling · page 11 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_46095b1b
We have shown that diffusion models, a class of likelihood-based models with a stationary training objective, can obtain better sample quality than state-of-the-art GANs.

Conclusion · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_c38471b5
Our improved architecture is sufficient to achieve this on unconditional image generation tasks, and our classifier guidance technique allows us to do so on class-conditional tasks.

Conclusion · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_068f92f2
Finally, by combining guidance with upsampling, we can further improve sample quality on high-resolution conditional image synthesis.

Conclusion · page 13 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_13a0086d
In this section, we demonstrate that we can achieve results better than StyleGAN2 and BigGAN-deep with the same or lower compute budget.

Computational Requirements · page 17 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_f90a844d
In this section, we show that conditional sampling can be achieved with a transition operator proportional to pθ(xt|xt+1)pφ(y|xt), where pθ(xt|xt+1) approximates q(xt|xt+1) and pφ(y|xt) approximates the label distribution for a … (excerpt)

Conditional Diffusion Process · page 25 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_eaed9c1f
28 J Using Fewer Sampling Steps on LSUN We initially found that our LSUN models achieved much better results when sampling with 1000 steps rather than 250 steps, contrary to … (excerpt)

Conditional Diffusion Process · page 28 · Automated extraction, not reviewed · 0 reviewed reproductions

Reproduced one of these claims?

Record your result in the Mearc app on your own machine, then use Publish this result with this paper's id (2105.05233) and the claim id shown next to the claim. Mearc creates a file that holds only what you choose to publish. Add it to the public claim data by pull request; a maintainer other than you reviews it before it counts.