Research map / Image generation / Fast diffusion sampling

From Denoising to Refining: A Corrective Framework for Vision-Language Diffusion Model

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Masked diffusion language model decoding · 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 ReDiff, a vision-language discrete diffusion model that learns to revise its own draft errors via two-stage training (synthetic-error revision, then online self-correction using o4-mini expert revisions), enabling stable few-step parallel captioning.

Problem

Discrete diffusion vision-language models suffer error cascades under parallel decoding because train-inference discrepancy makes early token errors pollute context; already-unmasked tokens are fixed, causing syntactic errors and hallucinations.

Approach

Reframe mask-pred generation from passive denoising to active refining: train the model to revise synthetic errors, then iteratively fine-tune it on its own drafts paired with expert (o4-mini) corrections, and at inference re-predict previously unmasked tokens at each step.

Key results

• On CapMAS at 1 token/step, ReDiff reaches CLAIR 76.74 vs. LLaDA-V 65.54; at 8 tokens/step CLAIR is 67.44 vs. 46.38 for mask-pred training.
• ReDiff achieves CAPTURE 61.88 on DetailCaps-4870 at 1 token/step, above Qwen2.5-VL-7B (60.61), and drops only 0.65 points at 4 tokens/step.
• Stage II online self-correction alone (Base + Stage II) gives larger gains than Stage I alone in the ablation (e.g., CLAIR 73.02 vs 71.31 at 1 token/step).

Surprising findings

• A single round of online self-correction gave the most gains; a second round yielded no significant further improvement.
• Factuality scores for the mask-trained baseline stay stable at fewer steps, which the authors attribute to fewer extractable items rather than true stability.

Limitations

• Stage II depends on a proprietary expert model (o4-mini) for revisions, and the GPT-4o judge is also proprietary.
• Evaluation is limited to detailed image captioning; generalization to other generative tasks is claimed but not shown.
• Gains on Factuality at low step counts are partly attributed to metric artifacts by the authors.

Experimental setup

Datasets: LLaVA-1.5 caption data ShareGPT4V ViCrit

Benchmarks: CapMAS CapArena DetailCaps-4870

Metrics: CLAIR Coverage Factuality CapArena-Auto win-rate score CAPTURE

Baselines: LLaDA-V LaViDa MMaDA FUDOKI LLaVA-1.5-7B InternVL-2.5-7B Qwen2.5-VL-7B Mask-pred training

Builds on: LLaDA-V mask-predict diffusion D3PM ViCrit

Closest papers by meaning

• Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
• Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.89)
• Support Before Frequency in Discrete Diffusion (NeurIPS 2026 · similarity 0.89)
• From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
• CORE: Context-Robust Remasking for Diffusion Language Models (ICML 2026 · similarity 0.89)
• Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models (ICML 2026 · similarity 0.88)