In the landscape of modern machine learning, frozen pre-trained models provide stability and efficiency but often underperform on specific tasks due to mismatched data distributions. This paper introduces the Whisperer, a novel visual prompting framework that learns diffusion-based preprocessors to adapt inputs in pixel space, effectively "whispering" enhancements to frozen downstream models like EasyOCR. By framing the process as behavioral cloning of stochastically discovered improvement policies, our method achieves an 8% absolute (10.6% relative) reduction in Character Error Rate (CER) on a challenging dataset of 300k degraded synthetic text images, surpassing hand-engineered baselines such as CLAHE. The key innovation is a four-stage training curriculum that uses behavioral cloning to amplify "lucky" improvements discovered through the stochastic exploration of a partially trained diffusion model. This approach is highly sample-efficient and avoids the pitfalls of traditional reinforcement learning. Crucially, we frame this not as naive reinforcement learning, but as behavioral cloning of an exploration policy: we stochastically sample intermediate diffusion outputs, select those that improve CER by chance, and then train the model to reproduce them. This bootstrapping curriculum (4 stages over 60 GPU-hours) amplifies random successes into a systematic strategy. In summary, by whispering to the frozen OCR through its inputs, we improve an imperfect classifier without touching its weights.
Figures & tables
Filter Name
Mean CER
Mean Confidence
Description and Rationale
Original
0.7724
0.32
Unprocessed degraded images; baseline for all comparisons.
CLAHE4
0.7142
0.33
Contrast Limited Adaptive Histogram Equalization;
CLAHE8
0.7281
0.31
Higher clip limit (8); trades detail for broader equalization.
CLAHE2
0.7323
0.35
Lower clip (2); preserves more local contrast.
MorphOpen3×3
0.7624
0.33
Morphological opening; removes small noise but can erode text edges.
Gamma 1.5
0.7689
0.31
Brightens mid-tones; useful for low-light degradations.
Table 1: Performance of various hand-engineered image filters on a frozen EasyOCR model, sorted by Mean Character Error Rate (CER). The results demonstrate a clear performance plateau, with clahe_4 being the best-performing filter but failing to break the 0.71 CER barrier.
Figure 1: Inference-time architecture of the Whisperer: a frozen perceptual encoder conditions a U-Net that produces clamped pixel-space updates applied iteratively to the degraded input.
Figure 2: Training curriculum (four-stage pipeline) used to bootstrap the diffusion-based visual prompt for a frozen OCR model.
Method
Mean CER
Confidence
Δ vs Original
Original
0.7724
0.32
—
CLAHE 4.0 (Best Filter)
0.7142
0.33
-5.8%
Ours (Full Curriculum)
0.6905
0.37
-8.2%
Table 2: Breaking the Preprocessing Wall
Figure 3: Qualitative comparison of classical preprocessing filters. Each panel shows the original input and the corresponding visual outputs after applying different filters (e.g., CLAHE, unsharp masking, and gamma correction).
Vision-Language Models (VLMs) have shown strong promise on Optical Character Recognition (OCR), yet the sheer number of visual tokens required to encode dense documents incurs prohibitive inference cost. Existing pruning methods rely on physical eviction, e.g., permanently discarding visual tokens during the prefill stage. While effective for natural images, this strategy fundamentally breaks down on OCR, where virtually every visual token may correspond to a character or structural element, and any irreversible loss leads to catastrophic accuracy degradation. We observe that, although document images appear globally dense and seemingly unprunable, the model's attention to them is in fact temporally sparse: at each decoding step it concentrates on a small region that shifts gradually across steps, much as a human reader fixates on successive words rather than perceiving an entire page at once. Motivated by this Dynamic Visual Fixation phenomenon, we recast the intractable global pruning problem as a tractable local, dynamic one and propose FastOCR, a training-free framework with two complementary modules. Specifically, Focal-Guided Pruning identifies a small set of focal layers and selects the most task-relevant visual tokens from them at each step, while Cross-Step Fixation Reuse exploits the gradual shift of fixation to warm-start each step from the previous one. By dynamically adjusting which tokens are attended rather than evicting any from the cache, FastOCR avoids permanent information loss. Extensive experiments show that FastOCR serves as a plug-and-play acceleration module, generalizing consistently across five VLMs of varying sizes and architectures. On Qwen2.5-VL, FastOCR retains 98% of the unpruned model's accuracy while attending to only 5% of the visual tokens per decoding step, reducing attention latency by 3.0×.
Optical Character Recognition (OCR) is a fundamental task for digitizing information, serving as a critical bridge between visual data and textual understanding. While modern Vision-Language Models (VLM) have achieved high accuracy in this domain, they predominantly rely on autoregressive decoding, which becomes computationally expensive and slow for long documents as it requires a sequential forward pass for every generated token. We identify a key opportunity to overcome this bottleneck: unlike open-ended generation, OCR is a highly deterministic task where the visual input strictly dictates a unique output sequence, theoretically enabling efficient, parallel decoding via diffusion models. However, we show that existing masked diffusion models fail to harness this potential; those introduce structural instabilities that are benign in flexible tasks, like captioning, but catastrophic for the rigid, exact-match requirements of OCR. To bridge this gap, we introduce DODO, the first VLM to utilize block discrete diffusion and unlock its speedup potential for OCR. By decomposing generation into blocks, DODO mitigates the synchronization errors of global diffusion. Empirically, our method achieves near state-of-the-art accuracy while enabling up to 5x faster inference compared to autoregressive baselines.
Sean Man, Gilad Deutch, Roy Ganz +4
Technion - Israel Institute of Technology, Haifa, Israel. · Amazon Web Services.
Optical character recognition remains critical infrastructure for document digitization, yet state-of-the-art performance is often restricted to well-resourced institutions by prohibitive computational barriers. End-to-end transformer architectures achieve strong accuracy but demand hundreds of GPU hours for domain adaptation, limiting accessibility for practitioners and digital humanities scholars. We present a modular detection-and-correction framework that achieves near-SOTA accuracy with single-GPU training. Our approach decouples lightweight visual character detection (domain-agnostic) from domain-specific linguistic correction using pretrained sequence models including T5, ByT5, and BART. By training the correctors entirely on synthetic noise, we enable annotation-free domain adaptation without requiring labeled target images. Evaluating across modern clean handwriting, cursive script, and historical documents, we identify a critical "Pareto frontier" in architecture selection: T5-Base excels on modern text with standard vocabulary, whereas ByT5-Base dominates on historical documents by reconstructing archaic spellings at the byte level. Our results demonstrate that this decoupled paradigm matches end-to-end transformer accuracy while reducing compute by approximately 95%, establishing a viable, resource-efficient alternative to monolithic OCR architectures.
Arundhathi Dev, Justin Zhan
Department of Computer Science, University of Cincinnati, Cincinnati, USA