{"id":"8a7714f7-c3d9-46a0-ac98-72973f845a66","arxiv_id":"2505.05853","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A screen-and-natural-image codec that losslessly encodes OCR text and uses a diffusion renderer with three levels of conditioning achieves state-of-the-art perceptual quality and high text accuracy at 0.005 to 0.05 bpp.","lead":"PICD is a learned image codec that compresses text and image content separately, then uses a diffusion model to render the decoded text glyph onto the decoded image, aiming for accurate text and natural-looking pictures at very low bitrates. It reports top perceptual quality and the best text accuracy among perceptual screen-content codecs on SCI1K and SIQAD, while also matching perceptual codecs on natural images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Text accuracy metric is circular: the same OCR engine defines the lossless text stream and scores the result, so reported BD-TEXT may measure OCR self-consistency rather than true character preservation.","rationale":"The reader identified OCR dependence as the weakest assumption, and the manuscript's own Section B.6 confirms it. I agree, with a sharper formulation: the issue is not just that text accuracy is bounded by OCR accuracy; the evaluation metric is circular, because the OCR engine that produces the ground-truth text stream is the same engine used to measure success. This makes the reported BD-TEXT numbers insufficient to support the headline claim that PICD preserves text. Other concerns (the entropy argument in Section 3.1, the comparison with direct text rendering in Table 3) are real but secondary: the entropy argument's H(Z|Y)=0 step is not justified, and direct text rendering achieves higher text accuracy in Table 3, but the empirical perceptual-quality advantage is still credible. The central empirical claim stands or falls on whether text accuracy is measured against true text, so the most load-bearing check is an independent text-fidelity evaluation. Because the reader's CONDITIONAL verdict already requires addressing OCR dependence, the verdict should remain unchanged.","tokens_in":14686,"tokens_out":5456,"duration_ms":57085,"concrete_test":"Recompute text accuracy on SCI1K using ground-truth text labels (or a held-out OCR engine not used in the pipeline, e.g., a commercial OCR such as Google Cloud Vision) and compare the decoded text with the actual text in the source images. Concretely, compute the BD-TEXT and word-level character error rate for PICD using this independent ground truth. If the independent text accuracy drops materially below the reported 0.107 BD-TEXT gain over MLIC, or if a sample of 100 source images shows Tesseract source transcription mismatch >5% against a human transcription, then the headline text-accuracy claim is an artifact of the circular metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PICD's defining advantage is 'high text accuracy' on screen content. In the proposed pipeline, the text ground truth is Z = h(X), where h is the Tesseract OCR engine (Section 3.2). Z is compressed losslessly, and the diffusion renderer is trained and guided so that the OCR output on the reconstruction matches Z. The reported text accuracy is then the Jaccard similarity between OCR outputs on the source and the reconstruction. This makes the metric circular: any character Tesseract misreads on the source is treated as ground truth, and the system is explicitly optimized to reproduce that misreading, including via PARSeq-based instance guidance in Section A.3. If Tesseract reads 'a' as 'c' on the source, a reconstruction that crisply renders 'c' scores perfect text accuracy even though the true text is wrong. The authors state this themselves in Section B.6: 'Our text rendering fails if the OCR algorithm fails.' Consequently, the central claim 'high text accuracy' is bounded by Tesseract's accuracy and, more importantly, the reported BD-TEXT values do not establish preservation of the actual characters. The lossless text stream is lossless only with respect to h(X), not with respect to the image's semantic text content. For screen images with unusual fonts, low resolution, or non-Latin scripts, the headline result could be high while true text fidelity is low.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PICD, a perceptual image codec intended to work for both screen content and natural images. The encoder extracts text information Z from the source using OCR, compresses Z losslessly, and compresses the image with MLIC conditioned on Z. The decoder renders the reconstructed image and the text glyph into a final image using a Stable Diffusion model, with conditioning integrated at three levels: domain-level LoRA fine-tuning on screen images, an adaptor-level hybrid ControlNet/StableSR conditioning module, and instance-level guidance that combines OCR-based and codec-based losses during DDPM sampling. The authors report BD metrics on SCI1K, SIQAD, Kodak, and CLIC, claiming state-of-the-art text accuracy and FID among perceptual codecs on screen images and the lowest FID on natural images, together with ablations on the three conditioning levels and on alternative text-preservation tools.","tokens_in":14930,"tokens_out":5251,"duration_ms":50077,"significance":"If the text-accuracy results were validated against true character-level ground truth, PICD would be a notable contribution: it is one of the first systems to combine a lossless text stream with diffusion-based rendering and to demonstrate competitive perceptual quality on both screen and natural content in a single framework. The three-level conditioning design is well motivated, the ablations are informative, and the comparison set spans six codecs including both MSE-optimized and perceptual baselines. The explicit discussion of failure cases and computational complexity is a strength. However, the central claim of high text accuracy is currently threatened by a circular evaluation protocol, and the formal optimality arguments are conditional on assumptions that the paper itself acknowledges are only approximately satisfied.","major_comments":[{"comment":"The text-accuracy evaluation is circular. The ground truth for the lossless text stream is Z = h(X), where h is the Tesseract OCR engine, and the reported metric is the Jaccard similarity between OCR outputs on the source and the reconstruction. The decoder and the instance-level guidance (Section A.3, using PARSeq) are explicitly optimized to make OCR output on the reconstruction match Z. Consequently, any character that Tesseract misreads on the source is encoded as ground truth, and a reconstruction that crisply reproduces that misread scores as perfectly accurate. The paper's own limitation statement in Section B.6, 'Our text rendering fails if the OCR algorithm fails,' confirms that the reported BD-TEXT values do not establish preservation of the actual characters in the image. Please re-evaluate text accuracy against ground-truth transcriptions or human reading, and report the source OCR accuracy as an upper bound on achievable performance.","section":"Section 3.2, Section 4.1, Eq. (3), Table 2"},{"comment":"The optimality arguments are conditional on definitions rather than substantive. Eq. (5), H(Z|Y)=0, holds because Y is constructed to contain the losslessly encoded Z; the rate identity in Eq. (6) says only that losslessly transmitting Z and conditioning on it costs no more than transmitting Y alone. Eq. (7) asserts p(hat X)=p(X) by integrating p(Xbar,Z)p(X|Xbar,Z), but this requires the diffusion model to sample the exact posterior p(X|Xbar,Z). The paper itself states in Section 3.5 that the conditional distribution 'may not be perfectly trained,' which is why instance-level guidance is introduced. The claims of optimal text preservation and optimal perceptual quality should therefore be softened or supplemented with error bounds rather than presented as proven optimality.","section":"Section 3.1, Eq. (5)-(7)"},{"comment":"The headline claim that PICD 'excels' in text accuracy is nuanced by the paper's own ablation. In Table 3, Direct Text Render achieves text accuracy 0.463 versus Diffusion Render's 0.445 at essentially the same bpp (0.016 versus 0.016). The advantage of diffusion rendering is in perceptual quality (FID 34.77 versus 52.20, CLIP 0.9059 versus 0.8785), not in text accuracy. The text in Section 4.2 should be rephrased to claim state-of-the-art text accuracy among perceptual codecs, not among all text-preservation approaches, unless the small difference in Table 3 is shown to be within noise or the comparison is made at matched operating points.","section":"Section 4.3, Table 3"}],"minor_comments":[{"comment":"The phrase 'compression mains to maintain' should be corrected to 'compression aims to maintain.'","section":"Section 2"},{"comment":"Eq. (8) denotes the OCR model as h(.) in the guidance loss, but Section A.3 explains that Tesseract is non-differentiable and that PARSeq is used instead. Please align the notation and state explicitly in the main text that the differentiable OCR engine replaces h in the guidance term.","section":"Section 3.5 and Section A.3, Eq. (8)"},{"comment":"The text contains a typo: 'Due to the efficient design of PCID' should read 'PICD.'","section":"Section 4.3"},{"comment":"The training descriptions say '10,000 steps of gradient ascent' and '5,000 steps of gradient ascent'; these should be 'gradient descent.'","section":"Section B.1"}],"recommendation":"major_revision","confidential_remarks":"The circular text-accuracy metric is the main load-bearing issue. I would not reject outright because the architecture is plausible and the perceptual-quality comparisons are reasonable; however, the central claim requires re-evaluation against ground-truth text. If the authors can provide such validation, the paper could be acceptable after revision. If not, the claims should be substantially downgraded to 'OCR-consistent rendering.'"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about PICD. First, the core idea is genuinely new: it splits screen-content compression into lossless OCR text plus diffusion rendering of the rest, and it works. Second, the headline 'highest text accuracy' is weaker than it looks, because the scoring OCR is the same engine that defines the ground truth.\n\nThe contribution is real. Prior screen-content codecs either rendered text directly on the decoded image (Mitrica, Tang) or used OCR losses (Lai), but none combined lossless text with a generative diffusion decoder. The three-level conditioning (LoRA fine-tuning on screenshots, hybrid ControlNet/StableSR adaptor, instance guidance) is a plausible recipe, and the ablations in Table 1 support each piece. The natural-image variant also holds up, with the best FID on Kodak/CLIC among the perceptual codecs. The paper is honest: Section B.6 admits failure when OCR fails, and Table 3 shows direct text rendering actually gets higher text accuracy (0.463 vs 0.445) at the cost of much worse FID/CLIP. So the claim should be 'best perceptual-quality-plus-text-accuracy,' not 'best text accuracy' unqualified.\n\nThe stress-test is right: because Z is defined as h(X) with Tesseract, a misread becomes ground truth, and the decoder is trained to reproduce that misread. The Jaccard metric then measures self-consistency with Tesseract, not true text fidelity. The authors are aware, but the paper's framing underplays it. A simple fix would be to score with a second OCR engine, or report human transcription accuracy on a subset. That said, the same OCR-based metric is standard in screen-content coding (Tang, Lai, etc.), so this is not a fatal flaw; it's a limitation that should be surfaced in the paper.\n\nThe 'optimality' arguments in Section 3.1 are more like consistency checks. H(Z|Y)=0 is true by construction once Z is lossless, and p(Xhat)=p(X) assumes the diffusion posterior is exact. Neither is a proof with error bounds, but they don't need to be; the empirical results are the argument. No code is released, so independent reproduction of the BD numbers isn't possible.\n\nBottom line: this is a solid systems paper worth a serious referee. I'd send it to review with a request for code and a non-circular text-accuracy check. If I were working on screen-content or perceptual compression, I'd cite it.","headline":"Genuinely new combination of lossless OCR text coding and diffusion rendering, with credible experiments; the 'best text accuracy' claim is inflated by a circular OCR-based metric.","tokens_in":15492,"tokens_out":2600,"would_cite":true,"duration_ms":26398,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Codec wins text accuracy and visual quality on screens and photos","keywords":["perceptual image compression","screen content coding","diffusion model","diffusion rendering","text accuracy","optical character recognition","lossless text coding"],"falsifier":"Take a set of screenshots whose intended text strings are known independently, deliberately including glyphs a standard OCR engine tends to misread (stylized fonts, small sizes, low contrast), run PICD, and measure character-level agreement between the decoded image and the intended strings. If the decoded text matches the OCR output at encoding time but not the intended text, the 'lossless text' claim has failed at the extraction stage; the paper's own failure-case section predicts such cases exist.","tokens_in":14459,"feed_emoji":"🖼️","tokens_out":9933,"duration_ms":92793,"temperature":0.7,"pith_summary":"PICD is a proposed image codec that splits a screen image into a losslessly compressed text stream and a compressed image, then uses a pre-trained diffusion model to render both back into one picture. The claim is that this split lets one codec reach the top of both axes at once—text accuracy and perceptual quality—where existing generative codecs produce sharp but wrong characters on screenshots, and existing screen codecs preserve text but look blurry at low bitrate. The paper reports that on screen benchmarks PICD has the highest text accuracy and lowest FID among the perceptual codecs compared, and that on natural-image benchmarks it reaches the lowest FID among all compared codecs. If the claim holds, screen content no longer forces a choice between readable text and visually convincing reconstruction, and the same codec can switch to natural images by dropping the text branch.","feed_headline":"Codec wins text accuracy and visual quality on screens and photos","feed_subtitle":"Encoding text separately and letting a diffusion model redraw it beats prior codecs on legibility and visual quality at once.","key_machinery":"The mechanism is 'diffusion rendering,' implemented as a three-level conditioning stack on top of a pre-trained Stable Diffusion model. Domain level: the base model is fine-tuned (LoRA, rank 256) on screenshots with text-content prompts so it can generate screen layouts instead of natural-image scenes. Adaptor level: a custom adaptor feeds the decoded image and the glyph image into the UNet, using ControlNet's feature encoder for glyphs, both that encoder and Stable Diffusion's VAE encoder for the decoded image, pixel shuffle for a lossless transform, and SPADE layers for injection, because vanilla ControlNet is too weak for low-level control and StableSR alone hurts text accuracy. Instance level: at each DDPM sampling step, the paper takes the posterior mean $E[X_0 \\mid X_t]$, runs a differentiable OCR (PARSeq) on crops of it and re-encodes it through the codec, then back-propagates two guidance terms so the intermediate estimate agrees with the true text and with the compressed image. Text $Z$ itself is extracted by Tesseract, compressed losslessly with cmix, and turned into a glyph image for conditioning.","core_discovery":"The central discovery is that text fidelity and perceptual quality do not have to compete inside one lossy codec if the text is taken out of the lossy path entirely. PICD encodes the OCR output $Z$ losslessly, compresses the image given $Z$, and then treats decoding as a conditional generation problem: sample $\\hat{X} \\sim p(X \\mid \\bar{X}, Z)$ from a diffusion model, where $\\bar{X}$ is the compressed image and $Z$ is the glyph-rendered text. The paper argues this is near-optimal for text because $H(Z \\mid Y)=0$ makes $H(Y \\mid Z)+H(Z)=H(Y)$, so preserving text costs no extra rate, and optimal for perception because the rendered marginal distribution matches the source. Empirically, Table 2 reports BD-TEXT of 0.107 on SCI1K and BD-FID of -20.68 among perceptual codecs on screen images, and BD-FID of -74.55 on Kodak, the lowest among all compared codecs. The natural-image version removes the glyph and OCR guidance and uses BLIP captions as prompts, making the same framework a perceptual codec for ordinary photos.","pith_inferences":["A natural extension the authors do not pursue: the same side-channel idea could handle other structured screen elements—diagrams, tables, formulas, code blocks—turning the codec into a general semantic codec whenever a reliable extractor and renderer exist.","The paper uses Tesseract for extraction and PARSeq for guidance; since the two OCR engines differ, a testable upgrade is to use one stronger neural OCR for both roles and measure whether text accuracy rises.","The rate-optimality argument implies a falsifiable prediction the paper does not test: once the text stream is fixed, additional image bitrate should buy visual quality but almost no further text accuracy."],"forward_implications":["If PICD's comparisons are right, screen-content coding can use a perceptual diffusion-based codec without giving up text legibility, removing the main reason screen content needed separate, blurrier codecs.","Because the rate argument makes text preservation cost no extra bitstream rate, further gains for screen content should come from better rendering and better image compression rather than from spending additional bits on text regions.","The natural-image simplification means the same trained framework, with glyph and OCR guidance removed and BLIP captions as prompts, already acts as a competitive perceptual codec for ordinary photos.","Among the text-preserving tools compared in Table 3, direct text rendering achieves slightly higher text accuracy than diffusion rendering but much worse FID and CLIP similarity, so the paper's approach is the one that keeps both metrics high simultaneously."],"supporting_citations":[{"why":"MLIC is the base learned image codec that PICD fine-tunes to accept the glyph/text condition.","marker":"[20]"},{"why":"ControlNet supplies the conditioning architecture adapted to inject the glyph image and decoded image into the diffusion UNet.","marker":"[55]"},{"why":"StableSR provides the SPADE-based adaptor design that the hybrid adaptor draws on for low-level control.","marker":"[45]"},{"why":"PerCo is the diffusion-based perceptual codec baseline whose opensource implementation is compared in the experiments.","marker":"[11]"},{"why":"The OCR-loss baseline and the source of the PARSeq neural OCR used for instance-level guidance.","marker":"[25]"},{"why":"TSA-SCC is the prior screen-content work that also encodes text losslessly, providing the direct-text-rendering baseline.","marker":"[43]"},{"why":"Tesseract is the OCR engine that extracts the text stream Z at encoding time.","marker":"[42]"},{"why":"cmix losslessly compresses the concatenated text stream.","marker":"[21]"},{"why":"Blau-Michaeli defines the perception-distortion optimality condition p(hat X)=p(X) that the paper uses to argue perceptual optimality.","marker":"[7]"},{"why":"BLIP generates captions used as text prompts in the natural-image simplification of PICD.","marker":"[28]"}],"fun_headline_variants":["Codec keeps text lossless, diffusion redraws screen images","Text encoded separately, diffusion renders final image","PICD: Screen and photo codec with diffusion rendering","Lossless text encoding enables diffusion-based perceptual codec"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire text branch depends on the OCR engine reading the source image correctly at encoding time: a single misread character is stored losslessly, rendered crisply, and counted as accurate, so the reported text accuracy is bounded by OCR accuracy, not by the codec, and the paper's supplementary material concedes that text rendering fails if the OCR algorithm fails.","fun_headline_variants_meta":{"raw":{"variants":["Codec keeps text lossless, diffusion redraws screen images","Text encoded separately, diffusion renders final image","PICD: Screen and photo codec with diffusion rendering","Lossless text encoding enables diffusion-based perceptual codec"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000479,"raw_usage":{"total_tokens":2392,"prompt_tokens":988,"completion_tokens":1404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":1347}},"tokens_in":604,"tokens_out":1404,"duration_ms":11369,"temperature":1.0,"reasoning_tokens":1347,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:53:43.317230+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of screenshots whose intended text strings are known independently, deliberately including glyphs a standard OCR engine tends to misread (stylized fonts, small sizes, low contrast), run PICD, and measure character-level agreement between the decoded image and the intended strings. If the decoded text matches the OCR output at encoding time but not the intended text, the 'lossless text' claim has failed at the extraction stage; the paper's own failure-case section predicts such cases exist.","supporting_citations":[{"cited_title":"Mlic: Multi-reference entropy model for learned image compression","cited_arxiv_id":null,"evidence_quote":"MLIC is the base learned image codec that PICD fine-tunes to accept the glyph/text condition."},{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"ControlNet supplies the conditioning architecture adapted to inject the glyph image and decoded image into the diffusion UNet."},{"cited_title":"Learned Image Compression with Text Quality Enhancement","cited_arxiv_id":"2402.08643","evidence_quote":"The OCR-loss baseline and the source of the PARSeq neural OCR used for instance-level guidance."},{"cited_title":"Li, Xiao Wen Wu, Ruizhi Chen, Haochen Li, Guo Lu, and Limin Cheng","cited_arxiv_id":null,"evidence_quote":"TSA-SCC is the prior screen-content work that also encodes text losslessly, providing the direct-text-rendering baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Tesseract is the OCR engine that extracts the text stream Z at encoding time."},{"cited_title":"A machine learning per- spective on predictive coding with paq8","cited_arxiv_id":null,"evidence_quote":"cmix losslessly compresses the concatenated text stream."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BLIP generates captions used as text prompts in the natural-image simplification of PICD."}],"review_version":1}