Skip to content

Latest commit

 

History

7 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

Awesome Pixel Diffusion Papers

A curated list of notable papers on pixel-space diffusion models for image and video generation. Papers are sorted by publication year in descending order, focusing on end-to-end pixel diffusion approaches that operate directly in raw pixel space, avoiding latent encodings where possible.

2025

  • PixelDiT: Pixel Diffusion Transformers for Image Generation
    arXiv:2511.20645
    A fully transformer-based model with dual-level design (patch-level and pixel-level DiTs) for end-to-end pixel-space diffusion, achieving strong FID scores on high-resolution images.

  • There is No VAE: End-to-End Pixel-Space Generative Modeling via Self-Supervised Pre-training
    arXiv:2510.12586 A two-stage framework with self-supervised pre-training for pixel-space diffusion and consistency models, achieving SOTA FID (1.58 on ImageNet-256, 2.35 on ImageNet-512) without any VAE or external pre-trained models.

  • Back to Basics: Let Denoising Generative Models Denoise
    arXiv:2511.13720
    Proposes "Just image Transformers" (JiT) that directly predict clean images in pixel space, leveraging the manifold assumption for efficient high-dimensional generation without noise prediction.

  • DeCo: Frequency-Decoupled Pixel Diffusion for End-to-End Image Generation
    arXiv:2511.19365
    Introduces frequency decoupling to enable stable end-to-end pixel diffusion, addressing VAE limitations for high-fidelity synthesis.

  • DiP: Taming Diffusion Models in Pixel Space
    arXiv:2511.18822
    Decouples global structure (via DiT on large patches) and local details (via lightweight head) for efficient pixel-space generation, balancing quality and compute.

  • PixNerd: Pixel Neural Field Diffusion
    arXiv:2507.23268
    Combines pixel-space diffusion with neural fields for continuous image representation and generation.

  • Advancing End-to-End Pixel Space Generative Modeling via Self-Supervised Pre-training
    arXiv:2510.12586
    A two-stage framework with self-supervised distillation to bridge performance gaps in pixel-space diffusion and consistency models.

  • PixelFlow: Pixel-Space Generative Models with Flow
    arXiv:2504.07963
    An end-to-end flow-based framework for direct pixel-space image generation, eliminating VAEs and upsamplers for simplicity and performance.

2024

  • Simpler Diffusion (SiD2): 1.5 FID on ImageNet512 with Pixel-Space Diffusion
    arXiv:2410.19324
    A scalable recipe for pixel-space diffusion using sigmoid loss-weighting and guidance intervals, achieving SOTA on ImageNet at 512x512 resolution.

  • Novel View Synthesis with Pixel-Space Diffusion Models
    arXiv:2411.07765
    Adapts diffusion architectures for end-to-end novel view synthesis directly in pixel space, outperforming prior methods.

  • Edify Image: High-Quality Image Generation with Pixel Space Laplacian Diffusion Models
    arXiv:2411.07126
    Cascaded models using Laplacian diffusion for photorealistic, pixel-accurate generation.

  • Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers
    arXiv:2401.11605
    Linear scaling with pixel count via hourglass transformers for high-res training.

  • Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis
    arXiv:2402.14797
    Extends pixel diffusion to video with scaled spatiotemporal transformers for text-conditioned synthesis.

2023 and Earlier

  • Simple Diffusion: End-to-End Diffusion for High Resolution Images
    arXiv:2301.11093
    Simplifies high-res diffusion training in pixel space without complex modifiers, achieving SOTA FID on ImageNet.

  • Hierarchical Text-Conditional Image Generation with CLIP Latents (GLIDE)
    arXiv:2204.06125
    Two-stage hierarchical model using CLIP latents for text-to-image, influential for later pixel-space extensions.

  • Cascaded Diffusion Models for High Fidelity Image Generation
    arXiv:2106.15282
    Introduces a pipeline of diffusion models generating images at increasing resolutions, starting from a base low-res model followed by super-resolution stages, enabling high-fidelity class-conditional ImageNet generation without auxiliary classifiers.

This list includes the provided papers and expands with recent works from arXiv searches up to December 2025. For updates, check arXiv under "pixel diffusion" or "pixel-space diffusion models." Contributions welcome!

About

No description, website, or topics provided.

Resources

Stars

38 stars

Watchers

2 watching

Forks

Releases

Packages

Contributors