acceptodds
Under review as a conference paper at ICLR 2027

ControlTrace: Recovering Control Fields for Hidden-Content Recognition

Abstract

Spatially conditioned diffusion models can embed words and contours in natural-looking images, but vision-language models (VLMs) may fail to recognize the hidden content. Transformation-based recovery depends on parameter and view selection. To evaluate hidden-content recovery and recognition, we construct FreqBlind, a 6,000-image benchmark spanning contours, real words and non-words across three conditioning strengths. The evaluated transformation-based methods show limited recognition of contour patterns and weakly conditioned hidden content. To address this limitation, we propose ControlTrace to recover the grayscale control field used during generation. An 8.4M-parameter U-Net predicts this field from the carrier image, and a VLM then identifies its content. With Qwen2.5-VL-7B-Instruct, ControlTrace achieves 60.2% open-ended contour recognition accuracy across the three conditioning strengths, exceeding the best of the three evaluated prior methods by 26.9 percentage points. On an A100 GPU, the complete pipeline adds only 7.4 ms (5.3%) to direct VLM inference. Recovered fields have lower pixel errors and higher structural similarity than the evaluated transformation views. Across four evaluated VLMs, ControlTrace retains its overall contour recognition advantage. Recognition remains stable under the tested JPEG compression, Gaussian noise and downsampling. These results support control-field recovery for hidden-content recognition in the evaluated setting.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Hidden-content recognition, Vision-language models, Conditional diffusion models, Control-field recovery, Image reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1QhhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines cases where VLMs miss a word or object outline embedded in a ControlNet-generated scene. SemVink and SMSP try to expose the hidden pattern through downsampling and low-pass views, respectively. ControlTrace trains an 8.4M-parameter U-Net using carrier images as inputs and their original grayscale control fields as reconstruction targets. For a new image, the network predicts the field without needing the original generation inputs. Its prediction becomes the visual input to a VLM, which identifies the hidden content. The comparison uses FreqBlind, where 6,000 images are divided equally between contour and text targets. Both groups cover three conditioning strengths, with real words and pseudowords making up the text group. Using Qwen2.5-VL-7B, ControlTrace recognizes 60.2% of the contours, compared with 33.3% for SMSP. Their contour scores are furthest apart at the weakest setting.

Critical strengths

The paper includes extensive experiments. The real-word, anagram, and pseudoword tests examine the reader’s reliance on familiar words. Natural and synthetic images with no embedded target show how often the reader reports one anyway. Hidden-content recognition has a practical role in moderation, since the classifiers and VLMs tested on Hateful Illusion often missed hidden hateful messages.

ControlTrace takes just 7.4 ms more per image (5.3%) than direct reading in the reported tests. The tested AVR implementation needs seven VLM calls for each image; ControlTrace needs only one to read the recovered field.

Critical weaknesses

The authors train a U-Net to recover control fields from the corresponding generated images, using standard reconstruction losses. Without the edge head, weak-contour accuracy is just 0.27 percentage points lower. The deeper versions they tested do no better. NAFNet reaches 56.51% overall contour accuracy when trained to recover the same control fields, leaving a 3.70-point gap to ControlTrace. Although the supervision-target ablation supports choosing control fields, the improvement beyond a basic supervised reconstruction model needs to be better established.

With SD1.5/QR Code, SMSP gets 22.56% contour accuracy; ControlTrace manages only 4.78%. This is a concern for moderation if an attacker can choose or fine-tune the control encoder to exploit that sensitivity. I would have liked to see the authors try training on several generators or adapting the model to a new encoder.

Questions for authors

Why is ControlTrace absent from the HC-Bench evaluation? Can the authors report its results under the same evaluation protocol as the baselines?

With the same training data and comparable tuning budgets, how much does the full method improve over a U-Net trained to recover control fields using only L1 and cosine losses, and over a comparably tuned NAFNet? Do the main conclusions still hold when the gradient loss and auxiliary edge head are removed together?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1Qhh: loss and backbone ablations, HC-Bench, and encoder training

Thank you for the close reading of our ablations. Your questions led us to test what the contour gain depends on and to evaluate ControlTrace on HC-Bench; all new results are now in the revised PDF, with changes marked in blue.

Q2 / W1. Basic U-Net, removing both terms, and a comparably tuned NAFNet. The main conclusions hold without the gradient loss and the edge head. All rows except SMSP use the same four-level RGB U-Net, data, learning rate and 30 epochs; accuracy (%) is the mean of three seeds, and intervals pair the 50 contour targets.

Training lossWeak contours (1,000)All contours (3,000)Weak: Δ vs. submitted, pp [95% CI]
L1 + cosine (basic U-Net)45.3760.44−0.03 [−1.10, 1.03]
L1 + cosine + gradient45.1360.73−0.27 [−1.33, 0.80]
L1 + cosine + edge head44.0359.23−1.37 [−2.53, −0.27]
L1 + cosine + gradient + edge head (submitted ControlTrace)45.4060.21–
SMSP (no training)12.7033.33–

Appendix C.2 (Table R4) reports this ablation, and Section 5.3 now notes that removing both terms leaves accuracy unchanged. Five- and six-level U-Nets did not improve weak contours.

For the backbone comparison (Appendix C.2, Table R5), both networks used L1 + cosine, the same data and 30 epochs; at seed 42 each tried three learning rates, and validation MAE selected 0.0001 (U-Net) and 0.0002 (NAFNet) for all seeds.

Contour accuracy (%)U-NetNAFNetU-Net − NAFNet, pp [95% CI]
Weak (1,000)45.8039.07+6.73 [3.90, 9.73]
All (3,000)60.3255.91+4.41 [2.56, 6.36]

Both backbones far exceed SMSP, U-Net leads under equal tuning, and the overall gap is similar to the paper's (4.41 vs 3.70 pp). Together with Table 12, where the submitted U-Net reaches only 13.27% on weak contours when trained to output the s = 2.0 carrier instead of the field, this places the gain in the recovery target rather than in the losses or the backbone.

Q1. HC-Bench. Table 8 was meant only to check our baseline implementations. We have now evaluated the three frozen checkpoints on the same 84 images (56 objects, 28 text), reader, prompts and scorer. Accuracy (%):

MethodObjects, HC-Bench promptText, HC-Bench promptObjects, FreqBlind promptText, FreqBlind prompt
Direct3.573.571.797.14
SemVink (40 px)28.5735.7130.3628.57
AVR (vote)3.573.578.9321.43
SMSP19.6410.7119.6442.86
ControlTrace (Ours)4.1715.483.5720.24

ControlTrace fails on HC-Bench objects: they come from photographs (continuous tones, light or dark foregrounds), while we trained only on white-on-black icons. In a test designed after seeing these failures, we fine-tuned ControlTrace without HC-Bench images (Table R2). Object accuracy (%, HC-Bench prompt):

ControlTrace trainingObjects (56)
Submitted model (frozen)4.17
Fine-tuned further, original recipe4.76
Fine-tuned on carriers made with HC-Bench-like settings4.76
Same, but with independent grayscale photo controls (both polarities)23.21

The photo controls add +18.45 pp [8.93, 29.17] over HC-Bench-like settings alone. Table 8 and Table R2 (Appendix B.2) now include these rows.

W2. Training on several encoders. The SD1.5/QR Code failure is in recovery, not reading (foreground MAE 0.121 with QR Code Monster vs 0.957 with QR Code, reader fixed). We mixed QR Code Monster and QR Code training pairs (9,600 each) and trained 10 epochs with three seeds, generating both encoders' training carriers at the strength s given in each row. Tests use the paper's SD1.5 sets (900 carriers each; weak = the 300 at s = 1.0). Accuracy (%; Appendix E.3, Table R9):

Mixed-encoder trainingQR Code: weakQR Code: allQR Code Monster: weakQR Pattern (unseen): all
Fine-tune submitted model, s = 1.04.448.89not run37.04
From scratch, s = 1.05.3316.6734.6735.48
From scratch, s = 1.53.6730.5923.1132.37
From scratch, s = 2.03.1131.7413.6730.48
SMSP (no training)5.0022.5612.6725.89

Adding the second encoder helps at medium and strong conditioning (from scratch at s = 2.0: QR Code, all strengths, 31.74% vs SMSP 22.56%) but not on the weakest images, where no method exceeds about 5%. Training at stronger s also lowers weak QR Code Monster accuracy (34.67% → 13.67%) and QR Pattern accuracy (35.48% → 30.48%). A known encoder can thus be covered at some cost; robustness to an encoder chosen to defeat the recovery model remains open, as Section 6 now states.

Thanks for running these extra experiments. It was good to see that the simpler U-Net does about as well on FreqBlind. That makes me more convinced that recovering the control field is a useful way to pick up hidden contours. I’m glad you also ran the HC-Bench comparison, since it shows how much harder things can get on a different dataset.

That’s the one point I’d bring out a bit more in the main text. The frozen model gets 4.17% on HC-Bench objects, and that goes up to 23.21% after fine-tuning with separate photo-based controls. Putting those numbers side by side would help readers see where the original model struggles and how much fine-tuning helped here. Overall, I’m more confident in the main finding after reading the response, and I still lean toward acceptance.

AuthorsOpenReview ↗

Response to Reviewer 1Qhh (follow-up): thank you

Thank you for reading our response and the new experiments so carefully, and for your continued support.

Your comments have enriched the paper and made it clearer. They led us to compare against a basic U-Net and a comparably tuned NAFNet, evaluate ControlTrace on HC-Bench and fine-tune it with photo controls, and train on a second encoder; all of these results are now in the revised PDF (Appendices B.2, C.2 and E.3). Together, they show where the contour gain comes from and where the frozen model needs further training.

Reviewer_N9ZhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an interesting method to uncover hidden structures in images: first recovering the hidden structure from a messy image, and then recognising the recovered structure using a vision language model. This is especially interesting, as previous methods try transform the image to make the hidden strcutrre easier to detect. So what the paper does it first train a model to learn the "hidden" structure of the image, the "control field", and in a second step removing much of the distracting texture using ControlTrace, making it much easier for the VLM to read.

Critical strengths
  • The provided method shows strong gains in the most difficult hidden-content cases where hidden shapes only faintly are visible (in the paper called low conditioning strength), reaching 45.4% contour accuracy compared to 12.7% of SMSP and 2.8% of SemVink.
  • Finally a paper that uses anonymous.4open.science for code! props!
Critical weaknesses

Major weaknesses

  • as you state yourself poor transfer across generators, weak advantages on text and false-positive risk as well as vulnerabilitty to targeted adverbial attacks limit broader use / real-world applications... I could imagine training other CNNs on similar tasks to be similarly effective.

Minor weaknesses:

  • license of code / models / dataset (benchmark FreqBlind) is missing completely (or I couldn't find a LICENSE file in the GitHub repository at least) - maybe it is hidden somewhere in the GitHub repository and I cannot find, but this definitely belongs to - at least - the appendix of the paper (stating under which license you release code, dataset and models)
  • FreqBlind contains only 50 contour and 50 word targets, prompt prefixes differ across methods, and AVR any-view coverage is not an available single-answer selector; broader targets, prompt-sensitivity controls, and strictly matched answer protocols are needed...
  • you use generators of only one family (SD) - a comparison with flux / moss-image / z-image / qqwen-image etc. would greatly increase your transferability claims
  • human assessment / annotation of the lesser obvious examples would be interesting (just a small sample size; would help
  • can you explain the discrepancy between s = 1.0 36.30 of hidden contours of conrolTrace accuracy vs the reported 45.4% accuracy in Contour on s = 1.0 in Table 1? I don't quite get it.
  • the codebase is fine, you have a requirements.txt - but for more "standardised" approach you could create a project (like uv init) and state the architecture you use, so metrics can be verified 1:1.
Questions for authors
  1. Were the final U-Net depth, loss weights, training allocation, baseline views, and transformation parameters fixed before seeing FreqBlind recognition results, and can you report an evaluation with that protocol?
  2. Can you report precision, recall, and false-positive behavior under a tuned none threshold at representative prevalence, especially given that AVR has much lower COCO false positives than ControlTrace in Table 4?
  3. why did you not use bigger, less-overtrained datasets in contrast to coco? e.g., coyo-labeled-300m, data comp-12B or alike?
5 comments
AuthorsOpenReview ↗

Response to Reviewer N9Zh (1/2): other networks, held-out targets, precision, and COCO

Thank you for the careful review and for noting the anonymous code release. Your questions led us to test how far our results generalize; all new results are now in the revised PDF, with changes marked in blue.

W1. Would another CNN be similarly effective? Largely yes, as the gain comes from recovering the control field, not from a particular network. Weak-contour accuracy (%, mean of three seeds; Appendix C.2, Tables 12 and R5):

Recovery networkTraining targetWeak contours (1,000)
U-Net, matched learning-rate searchControl field45.80
NAFNet, matched learning-rate searchControl field39.07
Submitted U-Net (Table 12)Control field45.40
Submitted U-Net (Table 12)s = 2.0 carrier, same prompt and seed13.27
SMSP (no training)–12.70

The other limits are real; Section 6 now scopes the claims to them:

  • ControlTrace is reliable only for the control types and generator family it was trained on (HC-Bench objects: 4.17%, 23.21% after photo-control fine-tuning, Appendix B.2; generators: W4).
  • It trails SMSP on FreqBlind words (61.62% vs 63.37%).
  • PGD at 8/255 lowers its weak-contour accuracy from 43.33% to 9.00%.
  • False positives: Q2 and Q3.

Q1 / W3. Fixed before seeing FreqBlind results? No, except the AVR views. Checkpoints were selected by validation loss, but FreqBlind accuracy informed U-Net depth, the gradient-loss weight, the training allocation and some transform parameters. With these, prompts and scoring aliases frozen, we tested 50 new contour targets and 50 new words (25 real, 25 pseudo-words), each in 20 new scenes at three strengths (3,000 carriers per domain, same generator). Accuracy (%):

Held-out setControlTraceSMSPControlTrace − SMSP, pp [95% CI]
Contours, weak (s = 1.0)12.572.00+10.57 [5.47, 16.50]
Contours, all strengths32.4814.47+18.01 [11.67, 24.57]
Words, weak (s = 1.0)22.6715.50+7.17 [4.20, 10.20]
Words, all strengths57.0859.77−2.69 [−6.97, 1.22]

The contour gap is smaller than on FreqBlind (+32.70 pp weak, +26.88 pp overall); Section 6 now cites both, and Appendix C.5 (Table R6) gives details. On words, ControlTrace again trails SMSP overall, as on FreqBlind.

Q2. Precision, recall and AVR. We split the existing Table 4 answers retrospectively into a calibration half and an evaluation half. The gate reports content only when the reader names something and the recovered field's gradient energy exceeds the 99th percentile of 5,000 calibration COCO negatives. On the evaluation half (500 weak positives, 5,000 other COCO negatives), precision follows from recall, FPR and prevalence (%; Appendix G.1, Table R12):

MethodRecallCOCO FPRPrecision at 0.1 / 1 / 10% prevalence
AVR (vote)0.201.420.01 / 0.14 / 1.54
ControlTrace66.205.481.19 / 10.88 / 57.31
ControlTrace + gate52.200.2815.73 / 65.32 / 95.39

AVR's low FPR comes from almost never reporting content (1 of 500 positives), so its precision is lowest.

Q3. Why COCO? COCO supplies the negative photos and the scene captions. We replaced each with COYO-Labeled-300M (original captions) and DataComp xlarge CommonPool.

False-positive rate (%) on 1,000 ordinary photos per source, with the none-capable prompt (no field embedded; photos not checked by hand; ControlTrace seed 42; Appendix G.1, Table R11):

MethodCOCOCOYODataComp
SemVink9.45.46.3
AVR (vote)1.31.41.8
SMSP10.09.08.3
ControlTrace (Ours)6.23.83.7
ControlTrace + gate (Q2)0.40.00.6

After retraining on each caption source (three models per source, 9,600 carriers each, same controls and seeds; COCO retrained the same way), FreqBlind contour accuracy (%; Appendix C.6, Table R7):

FreqBlind contoursCOCOCOYODataComp
Weak (1,000)45.4042.4343.20
All (3,000)60.4659.4659.62
Weak: Δ vs. COCO, pp [95% CI]–−2.97 [−4.27, −1.73]−2.20 [−3.77, −0.67]
All: Δ vs. COCO, pp [95% CI]–−1.00 [−1.82, −0.14]−0.83 [−1.86, 0.29]

The submitted model on 900 new carriers per source (60 held-out captions each), accuracy (%; also Table R7):

MethodCOCOCOYODataComp
SemVink22.0025.4423.78
AVR (vote)29.6730.2231.67
SMSP36.7841.4440.89
ControlTrace (Ours)65.7063.9666.41
AuthorsOpenReview ↗

Response to Reviewer N9Zh (2/2): prompts, other generators, human assessment, Table 1 vs 4, and release

W3. Prompts and AVR. With one shared prompt for all methods, weak-contour accuracy (%, 1,000 carriers; Appendix A.2, Table R1):

PromptDirectSemVinkAVR (vote)SMSPControlTrace (Ours)
Per-method prompts (Table 1)1.12.82.912.745.4
One shared prompt1.23.12.814.445.5

AVR answers by plurality vote; any-view coverage needs the true label, so it is only an oracle diagnostic.

W4. Other generator families. On Z-Image-Turbo and Qwen-Image-2512 (900 carriers each; the publishers' grayscale control adapters; no retraining), accuracy (%) by conditioning scale (Appendix E.4, Table R10):

Generator, scaleDirectSemVinkAVR (vote)SMSPControlTrace (Ours)
Z-Image, 0.450.331.670.331.672.22
Z-Image, 0.9053.3342.0055.3356.0028.78
Z-Image, 1.3560.0060.3362.6769.0023.67
Z-Image, all37.8934.6739.4442.2218.22
Qwen-Image, 0.452.000.332.332.671.33
Qwen-Image, 0.9033.6711.0031.0038.006.11
Qwen-Image, 1.3562.0047.6763.0064.6722.89
Qwen-Image, all32.5619.6732.1135.1110.11

Neither adapter hides much content: at the two stronger scales, direct reading comes within 9 points of the best method, and no method works at the weakest. Frozen ControlTrace − SMSP: −24.00 pp [−29.93, −18.00] on Z-Image and −25.00 pp [−30.41, −19.48] on Qwen-Image. We chose the Z-Image test after viewing pilot images (test list locked beforehand); Qwen-Image used the same targets, scenes and noise, with all settings fixed before generation and no pilot. A FLUX.2-klein-4B pilot was at chance (4, 5 and 5 of 10 contour pairs; preset bar 8), so we ran no recognition test. MOSS-Image takes no image or control input, so it cannot embed a control field.

W5. Human assessment. Ten readers new to FreqBlind named the hidden object in 20 random weak contours (s = 1.0): 32% at full size, 48% with a 96-px thumbnail. Images the VLM named (ControlTrace: ≥2 of 3 seeds; Appendix B.4, Table R3):

Images (human naming rate)DirectSemVinkAVR (vote)SMSPControlTrace (Ours)
9 (≥50%)01127
11 (<50%)00003
20 (total)011210

W6. Table 1 vs Table 4. Both count the same 1,000 weak positives; Table 4 also allows the answer "none" (Appendix G.1):

Weak positives (1,000)Correct
Table 1: name required, three-seed mean45.40%
Seed 42: name required451 (45.10%)
Seed 42: "none" allowed, as in Table 4363 (36.30%)

W2 / W7. Licenses and reproducibility. The repository root now has LICENSE files (Appendix A.1):

MaterialLicense
Our code and the three checkpointsMIT
Our FreqBlind manifests, annotations and evaluation recordsCC BY 4.0
Third-party artwork, COCO captions and photos, fonts, model weightsOriginal terms

It is now a uv project (pyproject.toml, uv.lock; conda/pip lockfiles too) with a reference training configuration and a CPU script recomputing 125 result summaries from saved records. The per-image records of the new model experiments are now released as well, with a CPU script that recomputes their tables. Appendix C gives the architecture (four-stage U-Net, 48/96/192/384 channels).

The author response is thorough and technically helpful especially on protocol transparency, prompt controls, precision/recall, licensing and reproducibility. It also adds useful held-out, caption-source, generator-family, and small human-evaluation experiments.my initial reservations are proven to be well-founded regarding limited transferability: ControlTrace is strongest on SD-style control fields, weaker on text, less competitive on Z-Image or Qwen T2I models, and remains vulnerable to adversarial perturbations. These results sharpen the scope of the contribution, but they do not fully address the generality concerns. Therefore, I keep my score unchanged.

Reviewer_15xPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ControlTrace separates hidden-content recognition into grayscale control-field recovery and VLM reading. An 8.37M-parameter U-Net learns from paired synthetic carriers and their generating control fields, using reconstruction, gradient, and auxiliary edge losses. FreqBlind contains 50 contour and 50 word targets, each rendered with 20 scene prompts drawn from a pool of 200 at three conditioning strengths. The study compares recognition, field fidelity, inference cost, reader transfer, encoder transfer, and perturbation sensitivity.

Critical strengths
  1. The recovery target produces a substantial, well-supported contour-recognition gain. Aggregate accuracy is 60.21% versus 33.33% for SMSP, with 45.4% versus 12.7% under weak conditioning. The matched supervision-target comparison in Table 12 improves weak-contour accuracy from 13.27% for stronger-conditioning scene targets to 45.40% for control-field targets. NAFNet also retains a large gain, supporting the recovery target beyond the particular U-Net architecture.

  2. The empirical investigation is unusually explicit about its boundaries. It includes three training seeds, target-cluster uncertainty, four VLM readers, paired anagram controls, negative sets, and several transfer settings. AVR single-answer voting is distinguished from oracle any-view coverage. The accessible release includes inspectable answer records; independently rescoring all 6,000 released seed-42 responses with the stated rule reproduces its contour and word results. This is a consistency check of saved outputs, not a rerun of model inference.

Critical weaknesses
  1. FreqBlind is not independent of configuration selection. Appendix C.4 states that its recognition results informed the learned configuration, and Appendix D states that downstream results also informed some transformation settings. Thus the issue affects more than one side of the comparison, and its effect on the ranking is uncertain. Seed variability and item-bootstrap intervals remain conditional on the selected configurations. The large reported gap is informative, but its precise magnitude lacks a clean final estimate with configurations fixed in advance.

  2. The external validation does not yet resolve how far the main advantage extends. Table 8 uses HC-Bench to check the baseline implementations but does not report ControlTrace on that external set. The generator study demonstrates partial transfer, including success with SDXL/QR Code Monster and failure with SD1.5/QR Code (4.78% versus 22.56% for SMSP). The authors acknowledge this dependence, and the main claims are appropriately scoped. Nevertheless, evaluation of the frozen recovery model on an existing external benchmark would help distinguish a useful encoder-specific recovery result from a more broadly applicable recognition method. Comparisons also concern the stated baseline implementations, including fixed-view AVR QA, rather than every variant of the original systems.

Overall assessment. I lean toward weak acceptance. The substantial contour gains, informative target-supervision comparison, and careful reporting make a useful empirical contribution in the evaluated setting. Configuration selection and incomplete external validation limit confidence in the exact effect size and breadth. I would not make a separately fine-tuned VLM or universal generator robustness a prerequisite for acceptance.

Questions for authors
  1. Can you provide a final evaluation on targets and scenes that did not inform configuration selection, with all methods' configurations fixed in advance? Please keep configuration-selection uncertainty separate from the seed and item uncertainty already reported.

  2. Why is ControlTrace absent from the HC-Bench comparison in Table 8? Could its frozen model be evaluated on the same 84-image subset, reader, prompts, and scoring protocol, or could you explain why that comparison would be inappropriate?

  3. In the SD1.5/QR Code failure, do field-recovery metrics and examples indicate a failure of recovery or a reader interface problem? This would make the method's applicability boundary more informative than simply reporting the recognition decline.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 15xP (1/2): held-out targets

Thank you for rescoring the released answers yourself. Your questions led us to test where the contour advantage holds and where it breaks down; all new results are now in the revised PDF, with changes marked in blue.

Q1 / W1. Held-out evaluation with configurations fixed in advance. Checkpoints were selected by validation loss, but FreqBlind accuracy informed the recovery configuration and some transform settings. We froze the checkpoints, all baseline settings, prompts and scoring aliases, then used the same generator to render 50 new contour targets and 50 new word targets (25 real words, 25 pseudo-words absent from the wordfreq lexicon), each in 20 new scenes at three strengths (3,000 carriers per domain). Contour accuracy (%):

StrengthDirectSemVinkAVR (vote)SMSPControlTrace (Ours)Ours − SMSP, pp [95% CI]
1.0 (weak)0.301.500.502.0012.57+10.57 [5.47, 16.50]
1.52.608.905.7015.1036.37+21.27 [13.53, 28.97]
2.06.2019.3016.5026.3048.50+22.20 [13.90, 30.53]
All3.039.907.5714.4732.48+18.01 [11.67, 24.57]

The ranking holds, but the gap is smaller than on FreqBlind (+26.88 pp overall, +32.70 pp weak). The new targets are harder: the reader named 31/50 of their clean control fields (FreqBlind: 39/50), and on nameable targets the gap is still +25.75 pp [18.14, 33.85]. Since targets and scenes changed together, we cannot separate target difficulty from selection effects.

Word accuracy (%):

StrengthDirectSemVinkAVR (vote)SMSPControlTrace (Ours)Ours − SMSP, pp [95% CI]
1.0 (weak)0.1012.300.3015.5022.67+7.17 [4.20, 10.20]
1.52.7067.1019.9073.8070.53−3.27 [−9.43, 2.67]
2.018.1082.2054.7090.0078.03−11.97 [−18.47, −6.27]
All6.9753.8724.9759.7757.08−2.69 [−6.97, 1.22]

As on FreqBlind (−1.74 pp), ControlTrace trails SMSP on words overall and leads only at strength 1.0.

Accuracy (%) of each frozen checkpoint:

Held-out setSeed 42Seed 43Seed 44
Contours32.7732.9331.73
Words56.0357.9057.30

Intervals resample the 50 targets (20,000 times) under fixed configurations, so they exclude selection uncertainty; the held-out tests are our selection-free estimates. Appendix C.5 (Table R6) reports both tests, and Section 6 now cites the held-out estimates alongside FreqBlind.

AuthorsOpenReview ↗

Response to Reviewer 15xP (2/2): HC-Bench and SD1.5/QR Code

Q2 / W2. HC-Bench with the frozen model. Table 8 was meant to check our baselines. We have now evaluated the three frozen checkpoints on the same 84 images (56 objects, 28 text), reader and scorer, under both prompts. Accuracy (%):

MethodObjects, HC-Bench promptText, HC-Bench promptObjects, FreqBlind promptText, FreqBlind prompt
Direct3.573.571.797.14
SemVink (40 px)28.5735.7130.3628.57
AVR (vote)3.573.578.9321.43
SMSP19.6410.7119.6442.86
ControlTrace (Ours)4.1715.483.5720.24

ControlTrace fails on HC-Bench objects; Table 8 (Appendix B.2) now includes these rows. The cause is the control type, not the encoder: all 60 HC-Bench images with generation metadata used SD1.5/QR Code Monster, our training encoder, but HC-Bench objects come from photographs (continuous tones, light or dark foregrounds), while we trained only on white-on-black icons. For 43–46 of the 56 objects (depending on the seed), no recovered pixel reached 128/255. In a test designed after seeing these failures, we fine-tuned ControlTrace without HC-Bench images (Table R2). Object accuracy (%, HC-Bench prompt):

ControlTrace trainingObjects (56)
Submitted model (frozen)4.17
Fine-tuned further, original recipe4.76
Fine-tuned on carriers made with HC-Bench-like settings4.76
Same, but with independent grayscale photo controls (both polarities)23.21

The photo controls add +18.45 pp [8.93, 29.17] over HC-Bench-like settings alone, placing ControlTrace above SMSP (19.64%) but below SemVink (28.57%). Section 6 now states that recovery is reliable only for the control types seen in training; all comparisons concern our stated baseline implementations, including fixed-view AVR QA (Appendix B.2).

Q3. SD1.5/QR Code: recovery or reading? Recovery (Appendix E.3, Table R8). With the seed-42 checkpoint and reader fixed, we compared 900 carriers per encoder, matched by target, scene and strength:

SD1.5 encoderForeground MAE (0–1)Accuracy, all 900 (%)Accuracy, 702 with nameable targets (%)
QR Code Monster (training encoder)0.12163.7874.64
QR Code0.9574.785.98

The 702 carriers belong to the 39 targets whose clean fields the reader names, so the reader is not the bottleneck. For a weak airplane in a kitchen, it answered "airplane" for the clean field and the QR Code Monster recovery, and "dog" for the QR Code recovery.

AuthorsOpenReview ↗

Follow-up for Reviewer 15xP

Thank you again for your careful review, including rescoring our released answers yourself. We understand that this is a busy period.

Following your questions, our two comments above report a held-out evaluation with all configurations fixed in advance, ControlTrace on HC-Bench, and the SD1.5/QR Code analysis; the revised PDF includes them (Appendices B.2, C.5 and E.3). As the discussion period closes on November 18 (AoE), we would be grateful to know whether these address your questions, and we would be glad to clarify anything that remains.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.