Research map / Image generation / Autoregressive and latent generation

DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling

NeurIPS 2025 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DiCo, a fully convolutional diffusion backbone (1x1 pointwise + 3x3 depthwise convs with compact channel attention) that matches or beats DiT and linear-attention/Mamba diffusion models on ImageNet at lower compute.

Problem

Diffusion Transformers are strong but computationally expensive due to quadratic self-attention, and recent linear/RNN-style alternatives (Mamba, linear attention) do not yield much runtime gain at high resolution. The paper asks whether a hardware-efficient ConvNet backbone can match DiT generation

Approach

Analysis of pretrained DiTs shows attention is mostly local; replacing self-attention with convolutions causes channel redundancy (many inactive channels), which a compact channel attention (GAP + 1x1 conv + sigmoid gating) fixes, yielding a U-shaped pure-ConvNet diffusion model.

Key results

• DiCo-XL with CFG reaches FID 2.05 at 256x256 and 2.53 at 512x512, with 2.7x and 3.1x speedup over DiT-XL/2 respectively.
• DiCo-H (1B parameters) reaches FID 1.90 on ImageNet 256x256 without extra modern tricks.
• Removing compact channel attention raises FID from 49.97 to 54.78 on the small model (ablation), i.e. CCA gives about 4.81 points improvement.

Measured effects reported in the paper

Surprising findings

• Naive replacement of self-attention with convolution in DiT yields many inactive channels, causing substantial channel redundancy and degraded FID.
• Pretrained DiT attention maps are predominantly local, with only a few neighboring tokens contributing significantly to a given anchor token.

Limitations

• Scaling limited to 1B parameters due to compute; larger models not explored.
• Text-to-image experiments are small-scale, trained from scratch on MS-COCO rather than large-scale T2I.

Experimental setup

Datasets: ImageNet-1K MS-COCO

Benchmarks: ImageNet 256x256 ImageNet 512x512 ImageNet 1024x1024 MS-COCO

Metrics: FID IS Precision Recall sFID Gflops throughput

Baselines: DiT DiG DiC DiS DiM DiffuSSM DiMSUM U-ViT FlowDCN LightningDiT MMDiT+REPA

Builds on: DiT ConvNeXt iDDPM classifier-free guidance

Closest papers by meaning

• DiC: Rethinking Conv3x3 Designs in Diffusion Models (CVPR 2025 · similarity 0.92)
• CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up (NeurIPS 2025 · similarity 0.92)
• SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.90)
• DiG: Scalable and Efficient Diffusion Models with Gated Linear Attention (CVPR 2025 · similarity 0.90)
• Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
• Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.89)