Research map / Image generation / Text-to-image diffusion

EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing

ICLR 2025 · Image generation › Text-to-image diffusion · approach: Text-driven synthesis and semantic guidance · 0 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 EC-DiT, a sparse mixture-of-experts diffusion transformer with expert-choice routing for text-to-image generation, scaled up to 97B parameters.

Problem

Scaling diffusion transformers for text-to-image generation beyond current sizes is underexplored, and standard token-choice MoE routing gives uniform per-token compute and needs load-balancing losses.

Approach

Each expert selects the top-C tokens from the whole image sequence (expert-choice routing) using router inputs that fuse text and image information via cross-attention, so compute adapts to image region complexity and denoising stage, with perfect load balance and no auxiliary loss.

Key results

• EC-DiT-M-64E (97.21B total params) reaches 71.68% GenEval overall, above SD3-Large (68.00%) and SD3-Large with DPO (71.00%).
• EC-DiT-3XL-32E reaches 70.91% GenEval with 5.18B activated params, and its inference time is 33.8% of the 8B dense M model.
• EC-DiT-XXL models trained for 200K steps match the training loss of Dense-XXL models trained for about 400K steps.

Measured effects reported in the paper

Surprising findings

• The solely pretrained EC-DiT-M-64E without fine-tuning outperforms SD3-Large even when SD3-Large is fine-tuned with DPO on GenEval.

Limitations

• Training uses a proprietary 1.2B-pair internet dataset, limiting reproducibility.
• Measured inference overhead (20-28%, 23% for M) exceeds the theoretical overhead, and model parallelism differs from the dense baseline's FSDP.
• Evaluation of DSG relies on an internal Gemini judge and weights/code are only promised upon acceptance.

Experimental setup

Datasets: internal ~1.2B text-image pairs from the Internet MS-COCO 256x256 (30K samples for FID/CLIP)

Benchmarks: GenEval DSG MS-COCO FID MS-COCO CLIP Score PartiPrompts

Metrics: GenEval overall score DSG score zero-shot FID CLIP score training loss inference time

Baselines: Dense DiT (XL, XXL, 3XL, M) GShard token-choice top-2 MoE SD3-Large SD3-Large w/ DPO FLUX.1[dev] PixArt-alpha SDXL DALL-E 3 IF-XL

Builds on: expert-choice routing GShard DiT rectified flow Sparsely-gated MoE

Closest papers by meaning

• Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts (ICML 2025 · similarity 0.91)
• Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
• Diff-MoE: Diffusion Transformer with Time-Aware and Space-Adaptive Experts (ICML 2025 · similarity 0.90)
• On the Scalability of Diffusion-based Text-to-Image Generation (CVPR 2024 · similarity 0.90)
• Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis (CVPR 2025 · similarity 0.89)
• Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.89)