MearcAll papers

arXiv 2205.11487 · cs.CV cs.LG

Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding

Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, and 8 more

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 recordedres_25fb02d9
Imagen achieves a new state-of-the-art FID score of 7.27 on the COCO dataset, without ever training on COCO, and human raters find Imagen samples to be on par with the … (excerpt)

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

No reproduction recordedres_554f506e
Our zero-shot FID score is also better than state-of-the-art models trained on COCO, e.g., Make-A-Scene [22] (at 7.6).

Introduction · page 3 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_13419719
We introduce DrawBench, a new structured suite of text prompts for text-to-image evaluation.

Introduction · page 3 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_88ea6b9e
We achieve a new state-of-the-art COCO FID of 7.27.

Introduction · page 3 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_cef3494c
We introduce DrawBench, a new comprehensive and challenging evaluation benchmark for the text-to-image task.

Introduction · page 3 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_d1aa9a94
On DrawBench human evaluation, we find Imagen to outperform all other work, including the concurrent work of DALL-E 2 [54].

Introduction · page 3 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_76e9ddce
We corroborate the results of recent text-guided diffusion work [16, 41, 54] and find that increasing the classifier-free guidance weight improves image-text alignment, but damages image fidelity producing highly saturated … (excerpt)

Large guidance weight samplers · page 4 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_9dc6eb58
Dynamic thresholding: We introduce a new dynamic thresholding method: at each sampling step we set s to a certain percentile absolute pixel value in ˆxt0, and if s > 1, … (excerpt)

Large guidance weight samplers · page 4 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_f83506ec
Furthermore, we found Layer Normalization [ 2] for text embeddings in the attention and pooling layers to help considerably improve performance.

Neural network architecture · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_3258866a
We make several modifications to this U-Net model for improving memory efficiency, inference time and convergence speed (our variant is 2-3x faster in steps/second over the U-Net used in [40, … (excerpt)

Neural network architecture · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_06311ace
For photorealism, Imagen achieves 39.2% preference rate indicating high image quality generation.

Results on COCO · page 7 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_7d1d2891
Scaling text encoder size is more important than U-Net size.While scaling the size of the diffusion model U-Net improves sample quality, we found scaling the text encoder size to be … (excerpt)

Analysis of Imagen · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_75728e61
We show that dynamic thresholding results in samples with significantly better photorealism and alignment with text, over static or no thresholding, especially under the presence of large classifier-free guidance weights … (excerpt)

Analysis of Imagen · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_3f453cd7
We show that training the super-resolution models with noise conditioning augmentation leads to better CLIP and FID scores.

Analysis of Imagen · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_378a3c93
We observe that conditioning over the sequence of text embeddings with cross attention significantly outperforms simple mean or attention based pooling in both sample fidelity as well as image-text alignment.

Analysis of Imagen · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_9baedd9f
B.1 Efficient U-Net We introduce a new architectural variant, which we term Efficient U-Net, for our super-resolution models.

Architecture Details · page 20 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_8b1fa891
We make several key modifications to the U-Net architecture, such as shifting of model parameters from high resolution blocks to low resolution, scaling the skip connections by 1/√2 similar to … (excerpt)

Architecture Details · page 20 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_06d798d3
Since lower resolution blocks typically have many more channels, this allows us to increase the model capacity through more model parameters, without egregious memory and computation costs. • When using … (excerpt)

Architecture Details · page 20 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_51d2d143
We observe that scaling the size of the language model text encoders generally results in better image-text 5https://github.com/openai/CLIP/blob/main/model-card.md 21 Category Description Examples Colors Ability to generate objects “A blue colored … (excerpt)

Pre-trained Text Encoders · page 21 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_cd7e9e51
Furthermore, for all values of aug_level, we observe that FID improves considerably with increasing guidance weight upto around 7− 10.

Pre-trained Text Encoders · page 23 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_646e2c6f
As we scale from 300M parameters to 2B parameters for the U-Net model, we obtain better trade-off curves with increasing model capacity.

Impact of Model Size · page 25 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_e67c3eb4
Imagen outperforms GLIDE on 8 out of 11 categories on image-text alignment, and 10 out of 11 categories on image fidelity.

Comparison to GLIDE and DALL-E 2 · page 27 · 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 (2205.11487) 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.