Research map / Image generation / Autoregressive and latent generation

MMAR: Towards Lossless Multi-Modal Auto-Regressive Probabilistic Modeling

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Latent diffusion with learned tokenizers · 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 MMAR, a continuous-token masked autoregressive framework with a lightweight diffusion head that jointly models image understanding and generation, using v-prediction for low-precision stability and a two-stage training schedule.

Problem

Joint image-text probabilistic models lose image information for understanding tasks, either through image discretization (VQ tokens) or through diffusion denoising steps that only expose noise-dependent representations.

Approach

Model image patches as continuous tokens within an autoregressive LLM using a masked-AR image expert (PLoRA) and a diffusion MLP head per patch, so every image hidden state is optimized for both P(I) and P(T|I); the numerical error of epsilon-prediction in low precision is removed by deriving v-prediction parameterization.

Key results

• MMAR-7B reaches 70.45 MMBench, 1486.9 MME-P, 84.02 POPE, and 48.25 AVE@18 at 256x256 resolution, above Show-o, VILA-U, and EMU-3 on the reported metrics.
• MMAR-7B achieves FID-30K of 22.9 on MSCOCO, versus 16.8 for Transfusion (as cited) and 9.24 for Show-o.
• Replacing v-prediction with epsilon-prediction in MMAR-0.5B raises FID-30K from 36.6 to 61.53 and lowers AVE@18 from 34.56 to 32.21.

Measured effects reported in the paper

Surprising findings

• The transfusion-like reproduction underperforms even the VQ-based variant, which the authors attribute to image information loss under limited training epochs.
• Generated images after stage 1 showed holes concentrated in the last 30% of the generation sequence, traced to the narrow [0.7, 1.0] mask ratio range.

Limitations

• Experiments are at limited resolution (256x256), data scale, and model size, and do not use more advanced continuous tokenizers, according to the authors.
• Generation quality lags behind dedicated generative models; MMAR-7B FID of 15.6 on MJHQ-30K and GenEval 0.51 are below several baselines.
• Hardware and compute budget are not reported.

Experimental setup

Datasets: Capfusion-120M CC12M LAION-aesthetics-12M JourneyDB LLaVA-v1.5-mix-665K MSCOCO 30k MJHQ-30K

Benchmarks: MMBench MME-P POPE SEED MM-Vet MSCOCO FID-30K MJHQ-30K FID GenEval

Metrics: average over 18 understanding benchmarks (AVE@18) FID GenEval overall

Baselines: LLaVA-1.5 EMU-2 SEED-X Chameleon-7B Show-o VILA-U EMU-3 Transfusion

Builds on: MAR Transfusion PLoRA DDIM v-prediction

Closest papers by meaning

• Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.90)
• Diffusion Models for Multi-Task Generative Modeling (ICLR 2024 · similarity 0.88)
• Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.87)
• Bifrost-1: Bridging Multimodal LLMs and Diffusion Models with Patch-level CLIP Latents (NeurIPS 2025 · similarity 0.87)
• Towards General Modality Translation with Contrastive and Predictive Latent Diffusion Bridge (NeurIPS 2025 · similarity 0.87)
• MMaDA: Multimodal Large Diffusion Language Models (NeurIPS 2025 · similarity 0.87)