{"id":"5a39e128-72d1-4049-9986-32497f21a651","arxiv_id":"2507.00373","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A text-prompted ROI image coder with a user-controlled quality knob and latent mask attention achieves strong RD and machine-vision results, though headline numbers use ground-truth masks.","lead":"This paper presents a learned image compression system where a user types a text description, such as 'house' or 'person', to tell the encoder which regions to keep sharp, and can dial a quality knob for the rest of the image. It matters because ROI-based compression is currently tied to fixed masks, while real users and downstream machine tasks need flexible control over what is preserved.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA results in Sec. IV-A are obtained with GT masks substituted for the text-derived masks; without a quantitative evaluation of TMA mask quality and its effect on RD/mAP, the central customization claim is unsupported.","rationale":"The reader's weakest_assumption precisely identifies the load-bearing gap: the quantitative SOTA evidence relies on GT masks, while the text-derived masks that actual users would get are only qualitatively validated. My read of the manuscript confirms this. Section IV-A explicitly replaces the text-derived similarity with GT segmentation results for all methods, so the headline RD and mAP comparisons do not exercise the TMA module. The only quantitative mask-quality evidence is IoU for two examples in Fig. 3, which is far too thin to support the 'customizable' claim. The training procedure also uses category annotations (GT masks) in stages 2–3, meaning the compression network has only seen clean masks during training; Lseg-based text masks at inference will differ. These issues are addressable with additional experiments, so CONDITIONAL is the right verdict rather than REJECT; the paper's LMA component and ablations provide partial support for the compression-side contribution. I agree with the reader's assessment and do not propose changing the verdict.","tokens_in":17136,"tokens_out":3545,"duration_ms":37862,"concrete_test":"Re-run the Sec. IV-A COCO validation experiments with TMA-generated masks (using each COCO category name as the text prompt) instead of GT masks, and report mean IoU of TMA masks against GT masks plus the ROI-PSNR, PSNR, and mAP curves over BPP; if these drop materially below the GT-mask curves, the central customization claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline contribution is text-customizable ROI compression, and the abstract/Conclusion claim SOTA performance for the proposed paradigm. However, all quantitative RD and mAP results in Fig. 5 are produced by replacing the similarity p in Eq. (2) with COCO GT segmentation masks (Sec. IV-A, 'we replace the p in Eq. (2) with the GT segmentation results'). This means the measured gains come from a near-perfect mask, not from the actual TMA output. The TMA module is only demonstrated qualitatively on a few Kodak images (Figs. 1, 6) and with IoU on two single examples (Fig. 3); no average mask accuracy is reported, and no end-to-end RD or mAP curve is given for text-derived masks. Moreover, the compression network is trained in stages 2–3 with COCO category annotations (i.e., GT masks), while at inference the masks come from the Lseg-based SG sub-module, introducing a train/test distribution shift. If text masks have inaccurate boundaries or miss objects, the ROI-PSNR, PSNR, and mAP gains in Fig. 5 will not transfer to the actual customizable system, collapsing the main claim regardless of whether the GT-mask experiments are internally correct.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a customizable ROI-based deep image compression paradigm with three main components: a Text-controlled Mask Acquisition (TMA) module that derives ROI masks from a user text prompt via an Lseg/CLIP-based similarity map and an adjustable binarization step; a Customizable Value Assign (CVA) mechanism that encodes a user-chosen quality trade-off factor σ into the non-ROI mask values and the weighted distortion loss; and a Latent Mask Attention (LMA) module that fuses a latent spatial prior from the mask with a latent RDO prior (importance map) to guide the latent representation. The method is trained on COCO with category annotations and tested on COCO validation and Kodak. The quantitative rate-distortion (RD) and machine-vision results in Fig. 5 are obtained by replacing the similarity map in Eq. (2) with ground-truth COCO segmentation masks, while text-driven masks are demonstrated only qualitatively on a few Kodak images and with IoU on two examples in Fig. 3. The paper claims state-of-the-art ROI-PSNR, PSNR, and detection/segmentation mAP even when the same GT mask is supplied to all methods, and argues that the σ knob lets users manage the ROI/non-ROI quality trade-off.","tokens_in":17441,"tokens_out":6310,"duration_ms":69011,"significance":"If the full pipeline works as claimed, the paper addresses a genuinely useful gap: ROI-based deep image compression in which end users can define the ROI by a text prompt and control the ROI/non-ROI quality balance via a single parameter. The component ablations (MR, IG, fusion, η, σ) are informative and give some independent support for the architectural choices. The promise of public code and the zero-shot demonstrations on out-of-vocabulary text prompts are also strengths. However, because every quantitative RD and mAP curve substitutes GT masks for the proposed text-derived masks, the central claim that text-customizable ROI compression achieves the reported gains is not yet supported. The significance is therefore conditional on closing that evaluation gap with end-to-end text-mask experiments.","major_comments":[{"comment":"The quantitative backbone of the paper is built on oracle masks: the text states 'we replace the p in Eq. (2) with the GT segmentation results while conducting our method.' Consequently, the PSNR, ROI-PSNR, and mAP curves in Fig. 5 measure the compression network with near-perfect masks, not the proposed text-controlled mask acquisition pipeline. Since the main contribution is text-customizable ROI selection, the results do not yet support the claim that users can achieve these gains by typing a prompt. Please add an end-to-end evaluation on COCO validation in which the TMA module (SG+AB) produces masks from text prompts (e.g., the COCO category names) and report mean mask IoU, PSNR/ROI-PSNR vs BPP, and detection/segmentation mAP vs BPP, overlaid on the GT-mask curves. This is the minimum evidence needed to connect the system-level claim to the experiments.","section":"§IV-A, Eq. (2), Fig. 5"},{"comment":"Mask-quality evidence for TMA is limited to two images with IoU reported for a few η values, and no average over the dataset is given. Moreover, the choice η=0.85 is made by sweeping η on the COCO validation set, which is the same set on which the final RD and mAP numbers are reported; this is a form of tuning on the test set. Please report average mask IoU and boundary metrics over a hold-out split (or cross-validation), compare against the Lseg baseline without the AB sub-module, and show how end-to-end RD/mAP varies with η over the hold-out split. Without this, the reader cannot judge whether the qualitative examples in Figs. 1, 3, and 6 are representative.","section":"§III-B, Fig. 3, and §IV-D"},{"comment":"The training protocol states that stages 2 and 3 use COCO images 'together with the category annotations,' but it does not specify whether the mask m fed to the compression loss is the GT annotation mask or the output of the TMA module. If GT masks are used in training, there is a train/test distribution shift because inference uses Lseg-derived masks, and the decision in §IV-A to substitute GT masks at evaluation time is then a consequence of this mismatch. Please clarify the training mask source, and if GT masks are used, either fine-tune with on-the-fly TMA masks or quantify robustness by perturbing GT masks (e.g., dilation/erosion, random omissions) and measuring the resulting change in RD and mAP. This will bound the impact of TMA mask errors on the headline results.","section":"§III-D"},{"comment":"The σ-based trade-off is enforced by construction: Eq. (2) assigns non-ROI mask values σ and Eq. (4) multiplies the distortion at each pixel by U(m), so increasing σ directly reduces the loss weight on non-ROI pixels. The trends in Fig. 11 are therefore a direct reflection of the loss definition rather than an emergent property of the learned system. To substantiate the claim of 'effective management' of the quality trade-off, report rate-distortion curves for several σ values at matched total BPP (e.g., by adjusting λ per σ) and report the measured bit allocation to ROI and non-ROI. This would demonstrate that σ is a meaningful user control beyond its explicit role in the objective.","section":"§III-B Eq. (4) and §IV-D Fig. 11"}],"minor_comments":[{"comment":"The RD and mAP curves do not include error bars or confidence intervals, and no tabulated mean values are provided; please state the number of seeds and add standard deviations or a table of representative operating points.","section":"Fig. 5 and Fig. 11"},{"comment":"The notation Q(p; 1, σ) in Eq. (2) is not defined; the piecewise definition Q(phw) in Eq. (3) depends on phw and η, but the parameters 1 and σ are not explained there. Please rewrite the definition to clarify the role of each parameter.","section":"Eq. (2)-(3)"},{"comment":"The abstract and Section IV-A use the phrase 'uniform mask' to describe the experimental setup; since the masks are COCO GT segmentation masks, please clarify whether 'uniform' means identical across methods or a constant-valued mask.","section":"Abstract and §IV-A"},{"comment":"Minor text issues include 'repeting' in Section III-B, 'tread-off' in Section III-B, 'iamge' in the caption of Fig. 7, and inconsistent spacing in 'CV A' in the abstract; please also make the range of phw and the role of η explicit in Eq. (3).","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core architecture is interesting, but the central claim of text-customizable ROI compression is not quantitatively validated because all headline results use GT masks. This is a substantial but fixable gap; I would be willing to review a revised version that adds end-to-end text-mask evaluation and addresses the training/evaluation mask mismatch."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a genuinely useful integration: text-prompted ROI masks (via Lseg/CLIP), a user-set quality knob for the non-ROI, and a learned latent fusion of a mask prior and an importance map. If the full pipeline holds together, it makes ROI compression meaningfully more practical. Second, the headline numbers are not actually produced with the proposed text masks. The authors replace the similarity map in Eq. (2) with COCO ground-truth segmentation masks for the RD and mAP comparisons in Fig. 5. So the SOTA performance they advertise supports the compression architecture given a perfect mask, not the customizable system the paper is about.\n\nWhat is genuinely new: among the cited work, this is the first to let a user specify the ROI with free text and to combine that with a changeable non-ROI mask and a latent-space fusion of mask prior and RDO prior. The component ablations (MR, IG, fusion, eta, sigma) are reasonable and mostly support the design choices. The special-case tests with the whole image as ROI or non-ROI are a nice sanity check, and the Kodak zero-shot examples with terms like \"house,\" \"fence,\" and \"grass\" show the TMA can generalize to categories outside COCO, at least qualitatively.\n\nThe soft spots are real but fixable. The central gap is the missing quantitative evaluation of the TMA module. There are two IoU examples and a few crops, but no average mask accuracy on COCO and no end-to-end RD or mAP curves using text-derived masks. Training uses COCO category annotations (GT masks) in stages 2–3 while inference uses Lseg similarities, so there is a train/test distribution shift that could eat into the claimed gains. The sigma trade-off is also partly built into the weighted distortion loss, so the sigma response curves are partly by construction; the module ablations give some independent support, so I consider this a moderate concern rather than fatal. The paper has no error bars or tabulated numbers, making effect sizes hard to judge. Eta is tuned on the validation set, and the code is only a promised URL, not a public repo.\n\nWho is this for? Researchers working on learned image compression with ROI or machine-vision objectives will find the architecture and ablations useful, but they should not take the headline numbers at face value. The paper deserves a serious referee, but it needs major revision: quantify TMA mask quality (e.g., average IoU against GT on COCO), report end-to-end RD and mAP with text masks, include standard deviations, and make code and weights public. Send it to peer review with those requirements; the core idea is worth engaging and the paper is honest enough about its procedure that the gaps are addressable. It is not a reject on the merits.","headline":"Useful integration of text-prompted ROI masks with latent mask attention, but the headline SOTA curves are run with ground-truth masks, so the customizable claim is not yet supported.","tokens_in":17928,"tokens_out":3464,"would_cite":false,"duration_ms":38532,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Text prompts, not fixed masks, can drive ROI image compression.","keywords":["ROI-based image compression","learned image compression","text-controlled mask acquisition","latent mask attention","rate-distortion optimization","quality trade-off factor","zero-shot mask acquisition","machine vision evaluation"],"falsifier":"Run the pipeline on a held-out validation set twice, once with ground-truth masks and once with masks produced only from the text prompt, and compare the ROI-PSNR and detection mAP curves; if the text-mask curves fall below the ground-truth-mask curves by more than the margin the paper reports over its anchors, the practical text-customization claim fails even if the GT-mask experiments are correct.","tokens_in":16957,"feed_emoji":"🖼️","tokens_out":7151,"duration_ms":74314,"temperature":0.7,"pith_summary":"The paper tries to make ROI-based image compression customizable: a user types the thing they care about (\"person\", \"fence\", \"grass\") and that text becomes the mask that steers where bits go. It also gives the user a single number, the quality trade-off factor $\\sigma \\in [0,1]$, that sets how much quality the non-ROI gets relative to the ROI. The authors claim that fusing a latent mask prior with a latent rate-distortion prior improves both reconstruction quality and downstream object detection and instance segmentation at the same bitrate, compared with existing ROI codecs. The practical payoff is that ROI selection no longer needs a per-image segmentation step, and new categories can be handled without retraining.","feed_headline":"Text prompts choose the region a codec preserves","feed_subtitle":"It lets users pick regions by text and tune ROI versus background quality, beating fixed-mask codecs.","key_machinery":"The load-bearing machinery is the LMA module's two-prior fusion in latent space. The Mask Representation sub-module turns the single-channel mask into a latent spatial prior $l$ with the same shape as the latent $y$, so mask guidance no longer crosses from image space to latent space; the Importance-map Generation sub-module produces a latent RDO prior $i$ that flags which latent details most affect reconstruction. Their element-wise sum forms attention map $s$, which is multiplied onto $y$; the CVA mechanism controls the mask's non-ROI value through $\\sigma$, and a weighted distortion loss using the same mask reinforces the trade-off during training.","core_discovery":"The central discovery is that the ROI mask can be replaced by a text-derived similarity map and still drive a deep codec effectively, provided the mask is converted into latent-space guidance rather than used as an image-space outline. The proposed pipeline builds a mask $m$ from prompt-to-pixel similarities, then learns a latent spatial prior $l$ from $m$ and a latent RDO prior $i$ from the image's intermediate features; adding $l$ and $i$ and multiplying the sum onto the latent representation tells the entropy model where bits matter. With ground-truth masks, this arrangement reports higher ROI-PSNR and average PSNR than the compared ROI codecs, and with uniform masks it still beats the anchors, indicating the prior fusion itself carries much of the benefit.","pith_inferences":["If text-mask quality is the bottleneck, pairing the same codec with a stronger open-vocabulary segmentation model should push text-driven results closer to the ground-truth-mask curves.","The uniform-mask result suggests the latent RDO prior may be the dominant contributor; a clean test would ablate the IG sub-module while keeping the mask channel constant and measuring the gain.","The same latent-space prior fusion is a candidate for video ROI coding, where object prompts could define preservation targets per frame.","The reported mAP above the source image is likely a de-emphasis effect: aggressive non-ROI compression removes distracting context; a background-blur or background-masking baseline could test whether compression is needed at all."],"forward_implications":["A user can switch ROI from \"person\" to \"vehicle\" for the same image without retraining or re-segmenting, changing only the text prompt.","Setting $\\sigma$ small spends nearly all bits on the ROI; setting it near 1 approaches whole-image quality, giving one knob for the ROI/non-ROI trade-off.","Because the uniform-mask case still beats the anchors, the latent priors can benefit even users who never supply a meaningful mask.","On datasets without ground-truth masks, text prompts still produce usable masks, extending ROI compression to new categories.","Better boundary preservation in the ROI and suppression of non-ROI interference translate into higher mAP for object detection and instance segmentation at matched bitrate."],"supporting_citations":[{"why":"supplies the transformer-based deep codec used as Phase 2 baseline and compression backbone","marker":"[34]"},{"why":"contributes the importance-map generation (IG) sub-module that produces the latent RDO prior","marker":"[25]"},{"why":"provides the language-driven segmentation encoders adopted for the Similarity Generation sub-module","marker":"[55]"},{"why":"supplies the text-image alignment whose zero-shot capability lets text masks work on unseen categories","marker":"[24]"},{"why":"is the state-of-the-art ROI-based deep codec anchor compared in RD and mAP experiments","marker":"[9]"},{"why":"is the mask-concatenation ROI codec anchor that motivates latent-space mask guidance","marker":"[8]"},{"why":"provides the COCO training categories, ground-truth masks for fair comparison, and the validation set for RD/mAP","marker":"[26]"},{"why":"defines the object detection evaluation via Faster R-CNN mAP on reconstructed images","marker":"[59]"},{"why":"defines the instance segmentation evaluation via Mask R-CNN mAP on reconstructed images","marker":"[60]"}],"fun_headline_variants":["Text picks ROI, codec tunes the rest","Type your region, codec preserves it","Custom ROI compression via text prompts","Latent mask guidance lets text define ROI","Text-driven ROI for deep compression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative gains against the anchors come from substituting ground-truth segmentation masks for the text-produced similarities, so the whole customization story rests on text masks being accurate enough to reproduce those gains in practice.","fun_headline_variants_meta":{"raw":{"variants":["Text picks ROI, codec tunes the rest","Type your region, codec preserves it","Custom ROI compression via text prompts","Latent mask guidance lets text define ROI","Text-driven ROI for deep compression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1268,"prompt_tokens":983,"completion_tokens":285,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":222}},"tokens_in":599,"tokens_out":285,"duration_ms":3807,"temperature":1.0,"reasoning_tokens":222,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:18:11.651074+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline on a held-out validation set twice, once with ground-truth masks and once with masks produced only from the text prompt, and compare the ROI-PSNR and detection mAP curves; if the text-mask curves fall below the ground-truth-mask curves by more than the margin the paper reports over its anchors, the practical text-customization claim fails even if the GT-mask experiments are correct.","supporting_citations":[{"cited_title":"Transformer-based image compression,","cited_arxiv_id":null,"evidence_quote":"supplies the transformer-based deep codec used as Phase 2 baseline and compression backbone"},{"cited_title":"Asymmetric learned image compression with multi-scale residual block, importance scaling, and post-quantization filtering,","cited_arxiv_id":null,"evidence_quote":"contributes the importance-map generation (IG) sub-module that produces the latent RDO prior"},{"cited_title":"Language-driven semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"provides the language-driven segmentation encoders adopted for the Similarity Generation sub-module"},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"supplies the text-image alignment whose zero-shot capability lets text masks work on unseen categories"},{"cited_title":"Transformer-based variable-rate image compression with region-of- interest control,","cited_arxiv_id":null,"evidence_quote":"is the state-of-the-art ROI-based deep codec anchor compared in RD and mAP experiments"},{"cited_title":"Roi-based deep image compression with swin transformers,","cited_arxiv_id":null,"evidence_quote":"is the mask-concatenation ROI codec anchor that motivates latent-space mask guidance"},{"cited_title":"Microsoft coco: Common objects in context,","cited_arxiv_id":null,"evidence_quote":"provides the COCO training categories, ground-truth masks for fair comparison, and the validation set for RD/mAP"},{"cited_title":"Faster r-cnn: Towards real-time object detection with region proposal networks,","cited_arxiv_id":null,"evidence_quote":"defines the object detection evaluation via Faster R-CNN mAP on reconstructed images"},{"cited_title":"Mask r-cnn,","cited_arxiv_id":null,"evidence_quote":"defines the instance segmentation evaluation via Mask R-CNN mAP on reconstructed images"}],"review_version":1}