Research map / Image generation / Text-to-image diffusion
MMDiff: Extending Diffusion Transformers for Multi-Modal Generation
Yagmur Akarken, Orest Kupyn, Christian Rupprecht
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-timestep aggregation (4 timesteps) over single-timestep t=0 extraction (ablation) | MMDiff (single-stream + concept attention) · PASCAL VOC 2012 | mIoU: 50.20 -> 78.90 (+28.7%) | Improved |
| use single-stream blocks instead of double-stream blocks (t=0) (ablation) | MMDiff · PASCAL VOC 2012 | mIoU: 40.52 -> 50.20 | Improved |
| add concept attention to single-timestep features (ablation) | MMDiff · PASCAL VOC 2012 | mIoU: 50.20 -> 63.77 (+13.57%) | Improved |
| use concept attention instead of averaged cross-attention maps (ablation) | MMDiff (single timestep) · PASCAL VOC 2012 | mIoU: 58.31 -> 63.77 | Improved |
| add cross-attention maps to single-stream t=0 features (ablation) | MMDiff · DUTS | S-measure: 0.890 -> 0.869 | Hurt |
| add DINO-v3 features to multi-timestep FLUX features (ablation) | MMDiff · PASCAL VOC 2012 | mIoU: 78.90 -> 84.95 | Improved |
| add spatially varying aggregation weights to full model (ablation) | MMDiff + DINO-v3 · NYU Depth v2 | AbsRel: 0.130 -> 0.116 | Improved |
| add aggregation to full model (ablation) | MMDiff + DINO-v3 · PASCAL VOC 2012 | mIoU: 78.21 -> 84.95 | Improved |
| choose concept set object, background, detail, edges instead of no concepts (single-stream t=0) (ablation) | MMDiff · PASCAL VOC 2012 | mIoU: 50.2 -> 63.8 | Improved |
| use a larger concept set with 23 class names plus generic concepts (ablation) | MMDiff · PASCAL VOC 2012 | mIoU: 63.3 vs 63.8 | No clear effect |
| replace denoised latent with aggregated MMDiff features | Lyra video-to-3DGS decoder · RealEstate10K | PSNR: 21.37 -> 21.61 | Improved |
| train a unified decoder jointly on VOC, DUTS and NYU (ablation) | MMDiff · NYU Depth v2 | AbsRel: 0.116 -> 0.220 | Hurt |
Later papers that compared with this method
1 later accepted paper reports MMDIFF as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- A Joint Diffusion Model with Pre-Trained Priors for RNA Sequence–Structure Co-Design (ICLR 2026 · better on 9 of 9 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- DatasetDM — DatasetDM: Synthesizing Data with Perception Annotations Using Diffusion Models (NeurIPS 2023) · better on 3 of 3 shared results
- Dataset Diffusion — Dataset Diffusion: Diffusion-based Synthetic Data Generation for Pixel-Level Semantic Segmentation (NeurIPS 2023) · better on 2 of 2 shared results
- Diffusion Hyperfeatures — Diffusion Hyperfeatures: Searching Through Time and Space for Semantic Correspondence (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.90)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.90)
- Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.90)
- Seg4Diff: Unveiling Open-Vocabulary Semantic Segmentation in Text-to-Image Diffusion Transformers (NeurIPS 2025 · similarity 0.89)
- Diffuse Everything: Multimodal Diffusion Models on Arbitrary State Spaces (ICML 2025 · similarity 0.89)
- ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features (ICML 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
