Research map / Image generation / Text-to-image diffusion
Meissonic: Revitalizing Masked Generative Transformers for Efficient High-Resolution Text-to-Image Synthesis
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.
Meissonic scales non-autoregressive masked image modeling to 1024x1024 text-to-image generation with a 1B-parameter model, using hybrid multi-modal/single-modal transformer blocks, RoPE, masking-rate conditioning, micro-conditions, and feature compression layers.
Problem
Masked image modeling (MIM) text-to-image models were limited to 512x512 resolution and lagged behind diffusion models such as SDXL in image quality and text alignment.
Approach
Combine a multi-modal-then-single-modal transformer (1:2 ratio) with 1D RoPE, a discretized masking-rate sampling condition (1000 levels), micro-conditions (resolution, crop coordinates, human preference score), and 2D conv feature compression/decompression around the backbone to enable efficient 1024x1024 MIM generation without a super-resolution cascade.
Key results
- Meissonic achieves HPS v2.0 averaged score 28.83, above SDXL Base 1.0 (28.25) and other listed models.
- Meissonic reaches GenEval overall 0.54, versus SDXL 0.55 and SD v2.1 0.50.
- On the MPS RealUser-800 benchmark, Meissonic scores 17.34 versus 16.46 for SDXL Base 1.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a mixed multi-modal and single-modal transformer, with about a 1:2 block ratio | Meissonic · Text-to-image generation | performance | Helped |
| Add 1D RoPE for positional encoding | Meissonic · 512×512 text-to-image generation | generative performance | Helped |
| Condition sampling on the masking rate | Meissonic · Text-to-image generation | image detail and overall quality | Helped |
| Add original image resolution, crop coordinates, and human preference score as micro-conditions | Meissonic · High-resolution image generation | stability and image quality | Helped |
| Add feature compression and decompression layers | Meissonic · 1024×1024 image generation | computational efficiency | Helped |
| Fine-tune on high-quality curated data with refined captions | Meissonic · Text-to-image generation | generative capabilities | Helped |
| Add human preference score as a micro-condition during final-stage fine-tuning | Meissonic · High-resolution image generation | performance | Helped |
| Use CLIP text encoding instead of a large text encoder such as T5 | Meissonic · Text-to-image generation | visual quality | No clear effect |
| Add image resolution, crop coordinates, and human preference score as micro-conditions | Meissonic · High-resolution aesthetic image generation | image quality | Helped |
Surprising findings
- Meissonic is trained with a CLIP text encoder instead of T5-XXL, and the authors report this does not diminish visual quality, though letter/word rendering is limited.
- Zero-shot editing with Meissonic-512 via inversion-free masking outperforms SD1.5 and SDXL with Null-Text Inversion on EMU-Edit and MagicBrush CLIP metrics while being much faster.
Limitations
- Text encoder is CLIP rather than a large LLM, which may limit understanding of lengthy prompts and rendering words.
- GenEval results trail SDXL on several categories such as two-object, counting, position and color attribution.
- Training data is partly internal and undisclosed, limiting reproducibility.
Experimental setup
Datasets: LAION-2B SA-10M internal curated high-quality dataset (6M) synthetic image-text pairs (1.2M)
Benchmarks: HPSv2 GenEval MPS (RealUser-800) EMU-Edit MagicBrush HumanEdit
Metrics: HPS v2.0 GenEval overall MPS CLIP-I CLIP-T DINO L1 CLIPdir
Baselines: SDXL DeepFloyd-XL Stable Diffusion v1.5 Stable Diffusion v2.1 DALL-E 2 SDXL Refiner Null-Text Inversion InstructPix2Pix MagicBrush EMU-Edit
Builds on: MaskGIT MUSE Multi-modal Transformer (MMDiT-style) RoPE classifier-free guidance
Closest papers by meaning
- SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers (ICLR 2025 · similarity 0.88)
- Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis (CVPR 2025 · similarity 0.88)
- DreamText: High Fidelity Scene Text Synthesis (CVPR 2025 · similarity 0.87)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.87)
- PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis (ICLR 2024 · similarity 0.87)
- Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · similarity 0.86)
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.