CLEAR: Unlocking Generative Potential for Degraded Image Understanding in Unified Multimodal Models

Xiangzhao Hao; Zefeng Zhang; Zhenyu Zhang; Linhao Yu; Yao Chen; Yiqian Zhang; Haiyun Guo; Shuohuan Wang

cs.CVApr 6, 2026

CLEAR: Unlocking Generative Potential for Degraded Image Understanding in Unified Multimodal Models

Authors: Xiangzhao Hao, Zefeng Zhang, Zhenyu Zhang, Linhao Yu, Yao Chen, Yiqian Zhang, Haiyun Guo, Shuohuan Wang,

Organizations: Institute of Automation, Chinese Academy of Sciences · Baidu Inc.

Abstract

Image degradation from blur, noise, compression, and poor illumination severely undermines multimodal understanding in real-world settings. Unified multimodal models that combine understanding and generation within a single architecture are a natural fit for this challenge, as their generative pathway can model the fine-grained visual structure that degradation destroys. Yet these models fail to leverage their own generative capacity on degraded inputs. We trace this disconnect to two compounding factors: existing training regimes never ask the model to invoke generation during reasoning, and the standard decode-reencode pathway does not support effective joint optimization. We present CLEAR, a framework that connects the two capabilities through three progressive steps: (1) supervised fine-tuning on a degradation-aware dataset to establish the generate-then-answer reasoning pattern; (2) a Latent Representation Bridge that replaces the decode-reencode detour with a direct, optimizable connection between generation and reasoning; (3) Interleaved GRPO, a reinforcement learning method that jointly optimizes text reasoning and visual generation under answer-correctness rewards. We construct MMD-Bench, covering three degradation severity levels across six standard multimodal benchmarks. Experiments show that CLEAR substantially improves robustness on degraded inputs while preserving clean-image performance. Our analysis further reveals that removing pixel-level reconstruction supervision leads to intermediate visual states with higher perceptual quality, suggesting that task-driven optimization and visual quality are naturally aligned.

Figures & tables

Figure 1: Top: average scores of commercial and open-source multimodal models on clean versus degraded inputs from MMD-Bench across six benchmarks. All models show substantial performance drops under degradation. Bottom: comparison between existing multimodal models and CLEAR on a degraded image.

Figure 2: Overview of CLEAR. Stage 1 (top) performs supervised fine-tuning to establish the generate-then-answer reasoning pattern and warm-start the Latent Representation Bridge, with both VAE latent and ViT re-encoded features injected during this stage. Stage 2 (bottom) applies Interleaved GRPO, where text tokens are optimized with GRPO and the denoising step with Flow-GRPO, sharing the same group-relative advantage from answer-correctness rewards. The ViT path is removed in Stage 2, making the bridge the sole connection between generation and reasoning.

Figure 3: Left: the standard decode-reencode path in existing unified models. The generated VAE latent must be decoded into pixels and re-encoded through the ViT before it can enter the reasoning context. Right: the Latent Representation Bridge in CLEAR. The generated VAE latent is directly concatenated into the reasoning context, eliminating the decode-reencode bottleneck and providing an effective optimization route from answer correctness back to generation.

	MMD-Bench (Hard)
Method	MMBench	MM-Vet	MMVP	CV-Bench	MMStar	RealWorldQA	R-Bench-Dis	AVG
Closed-source models †
GPT-4o-mini	67.02	50.91	64.00	59.87	45.93	58.95	61.21	58.27
GPT-4.1-mini	76.08	51.88	71.00	74.96	60.73	72.41	72.52	68.51
Gemini-2.5-Flash	79.33	66.55	72.33	76.01	62.00	69.15	72.72	71.16
Open-source unified models

Table 1: Main results under Hard degradation. R-Bench-Dis is an existing degraded-image benchmark; the remaining six are from MMD-Bench. Best in bold , second best underlined . † Closed-source results are included as reference points and are not directly comparable due to differences in model scale and training data.

Method	Clean	Hard	Drop ( $\downarrow$ )
Bagel	66.86	59.57	7.29
CLEAR-SFT	69.34	63.04	6.30
CLEAR-RL	70.27	64.96	5.31

Table 2: Robustness analysis. Clean and Hard scores are averaged over the six MMD-Bench benchmarks. Drop = Clean

-

Hard.

Figure 4: Qualitative examples of CLEAR’s adaptive reasoning. Left: on a mildly noisy image, the model skips generation and answers directly. Right: on a severely blurred image, the model triggers generation to recover obscured details before answering.

Configuration	Clean	Hard
Bagel (base)	66.86	59.57
+ SFT	69.34	63.04
+ SFT + Dec-reenc + GRPO	70.14	63.72
+ SFT + Bridge (w/o GRPO)	69.51	63.11
+ SFT + Bridge + GRPO	70.27	64.96

Table 3: Component ablation averaged over six MMD-Bench benchmarks. “Dec-reenc” replaces the bridge with the standard decode-reencode path during GRPO.

State	BRISQUE $\downarrow$	NIQE $\downarrow$	MUSIQ $\uparrow$	Hard AVG $\uparrow$
SFT state	43.73	5.32	42.63	63.04
RL state	41.53	4.93	45.74	64.96

Table 4: No-reference perceptual quality and reasoning accuracy of intermediate visual states. BRISQUE and NIQE (lower is better) measure distortion; MUSIQ (higher is better) measures overall quality.

Figure 5: Generation triggering rate (bars, left axis) and total inference time (line, right axis) across degradation severity levels for each benchmark.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Property	R-Bench	MMD-Bench
Clean reference available	No	Yes
Severity levels	Single	Three (Low/Mid/Hard)
Structured categorization	No	Four categories
# base benchmarks	1	6
Capability diagnosis	Combined score	Per-benchmark

Appendix

Table 5: Comparison between R-Bench and MMD-Bench.

Benchmark	# Samples	Primary Focus	Evaluation
MMBench	2,974	Fine-grained multi-ability assessment	Accuracy (CircularEval)
MM-Vet	218	Integrated VL capability evaluation	GPT-4 scoring
MMVP	300	CLIP-blind visual perception	Accuracy
CV-Bench	2,638	Vision-centric spatial understanding	Accuracy
MMStar	1,500	Vision-indispensable reasoning	Accuracy
RealWorldQA	764	Real-world spatial comprehension	Accuracy

Appendix

Table 6: Base benchmarks used in MMD-Bench.

Category	Corruption Types	Real-world Source
Capture	lens_blur, lens_flare, motion_blur,	Camera hardware and
	dirty_lens, hsv_saturation	shooting conditions
Transmission	jpeg_compression, block_exchange,	Lossy compression and
	mean_shift, scan_lines	bandwidth limitations
Environmental	dark_illumination, atmospheric_turbulence,	Adverse lighting and
	gaussian_noise, color_diffusion	atmospheric conditions

Appendix

Table 7: The 16 corruption types in MMD-Bench, organized into four categories that reflect distinct real-world degradation sources.

Figure 6: Visualization of all 16 corruption types at three severity levels. Each row shows one corruption type applied to the same source image at Low (left), Mid (center), and Hard (right) severity.

Property	Value
Total samples	24,886
Direct-answer samples	12,267
Generate-then-answer samples	12,619
Average trace length (direct)	606
Average trace length (generate)	1080
corruption types used	16

Appendix

Table 8: SFT dataset statistics.

Level	MMBench	MM-Vet	MMVP	CV-Bench	MMStar	RealWorldQA	AVG
Clean	80.03	61.19	77.00	76.14	65.86	61.43	70.27
Low	79.41	57.93	75.33	75.86	64.40	63.39	69.39
Mid	78.48	55.27	75.00	75.17	64.60	61.69	68.37
Hard	72.52	51.97	71.33	72.25	60.67	61.05	64.97

Appendix

Table 9: CLEAR-RL results across degradation severity levels on each MMD-Bench benchmark. AVG is computed over the six benchmarks.

Category	Corruption	Bagel	CLEAR-RL	$\Delta$
Capture	lens_blur	57.82	64.48	+6.66
	lens_flare	59.23	64.85	+5.62
	motion_blur	56.35	63.52	+7.17
	dirty_lens	58.94	64.25	+5.31
	hsv_saturation	59.51	64.63	+5.12
	Category avg.	58.37	64.35	+5.98

Appendix

Table 10: Per-corruption accuracy under Hard degradation, averaged over six MMD-Bench benchmarks. Corruptions are grouped by category.

\Delta

shows the improvement of CLEAR-RL over Bagel. Category averages are shown in italics.

Benchmark	High	Mid	Low
MMBench	1h 35m	1h 01m	54m
MM-Vet	9m	7m	7m
MMVP	8m	6m	5m
CV-Bench	31m	19m	15m
MMStar	43m	31m	28m
RealWorldQA	27m	17m	13m

Appendix

Table 11: CLEAR-RL evaluation time across degradation levels on each benchmark. R-Bench-Dis is evaluated only once as it contains pre-degraded images at a single severity level.

$\lambda$	Clean	Hard
0.0	69.85	63.52
0.1	70.06	64.31
0.3 (default)	70.27	64.97
0.5	70.18	64.72
1.0	69.71	63.89

Appendix

Table 12: Effect of Flow-GRPO loss weight

\lambda

on accuracy (6-bench average).

$w_{\text{acc}}$	$w_{\text{fmt}}$	$w_{\text{dec}}$	Clean	Hard
1.0	0.0	0.0	69.92	64.18
0.85	0.1	0.05	70.13	64.55
0.75	0.1	0.15 (default)	70.27	64.97
0.65	0.1	0.25	70.08	64.61
0.60	0.15	0.25	69.83	64.24

Appendix

Table 13: Effect of reward weight configurations on accuracy (6-bench average).

Steps	Hard	Denoising Time
10	63.68	1.8s
20	64.53	3.5s
30 (default)	64.97	5.2s
50	65.04	8.7s

Appendix

Table 14: Effect of denoising steps on accuracy (6-bench Hard average) and per-sample denoising time (measured on samples that trigger generation).

Generated ( $g$ )	Correct ( $c$ )	$R_{\text{dec}}$	Interpretation
1	1	$+1$	Generation helped
0	0	$-1$	Should have generated
1	0	$0$	Neutral
0	1	$0$	Neutral

Appendix

Table 15: Decision reward

R_{\text{dec}}

for the four possible outcomes.

Figure 13: Additional qualitative examples across different degradation types.

Unified Multimodal Models

Unified multimodal models combine multiple data types in shared systems for cross-modal understanding, generation, reasoning, and interaction.

See all 290 papers

Reinforcing Multimodal Reasoning Against Visual Degradation
May 10, 2026Rui Liu, Dian Yu, Haolin Liu +6VLM Robustness Image Corruption Robustness
Tencent Hunyuan · University of Maryland, College Park · University of Virginia +1
Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation
Mar 13, 2026Yichen Zhang, Da Peng, Zonghao Guo +19Unified Multimodal Models Multimodal Generation
Tsinghua University · Xi’an Jiaotong University · University of Chinese Academy of Sciences
Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards
Jun 25, 2026Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar +5Unified Multimodal Models Multimodal Generation
1Mohamed bin Zayed University of Artificial Intelligence · 4Linköping University · 3Australian National University +1

May 10, 2026J/K move · Enter open · S save

Reinforcing Multimodal Reasoning Against Visual Degradation

Rui Liu, Dian Yu, Haolin Liu, Yucheng Shi+5

Tencent Hunyuan · University of Maryland, College Park · University of Virginia+1

VLM Robustness Image Corruption Robustness Multimodal Large Language Models VLM Reasoning Multimodal Foundation Models RL for VLMs

Open paper AskPDF arXiv

Reinforcement Learning has significantly advanced the reasoning capabilities of Multimodal Large Language Models (MLLMs), yet the resulting policies remain brittle against real-world visual degradations such as blur, compression artifacts, and low-resolution scans. Prior robustness techniques from vision and deep RL rely on static data augmentation or value-based regularization, neither of which transfers cleanly to critic-free RL fine-tuning of autoregressive MLLMs. Reinforcing reasoning against such corruptions is non-trivial: naively injecting degraded views during rollout induces reward poisoning, where perceptual occlusions trigger hallucinated trajectories and destabilize optimization. We propose ROMA, an RL fine-tuning framework that modifies the optimization dynamics to reinforce reasoning against visual degradation while preserving clean-input performance. A dual-forward-pass strategy uses teacher forcing to evaluate corrupted views against clean-image trajectories, avoiding new rollouts on degraded inputs. For distributional consistency, we apply a token-level surrogate KL penalty against the worst-case augmentation; to prevent policy collapse under regularization, an auxiliary policy gradient loss anchored to clean-image advantages preserves a reliable reward signal; and to avoid systematically incorrect invariance, correctness-conditioned regularization restricts enforcement to successful trajectories. On Qwen3-VL 4B/8B across seven multimodal reasoning benchmarks, our method improves robustness by +2.4% on seen and +2.3% on unseen corruptions over GRPO while matching clean accuracy.

Abstract

Unified Multimodal Models

Explore similar work

Reinforcing Multimodal Reasoning Against Visual Degradation

Cheers: Decoupling Patch Details from Semantic Representations Enables Unified Multimodal Comprehension and Generation

Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards