{"id":"94696196-9714-4c14-bc93-e9278fa5fc48","arxiv_id":"2412.17671","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"B-Free trains an AI-image detector on self-conditioned Stable Diffusion 2.1 reconstructions of COCO photos plus inpainting-based augmentation, improving generalization to unseen generators and calibration.","lead":"This paper presents a new way to build training data for AI-image detectors: fake images are made from real photos with the same content, so the detector learns generator artifacts instead of content or format biases. The resulting detector generalizes to unseen generators with higher accuracy and better-calibrated scores than current methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training fakes come from SD2.1 self-conditioned reconstruction only; the paper does not directly verify artifact overlap with unseen generators, so the 'bias-free' claim rests on untested representativeness.","rationale":"I agree with the reader that the weakest point is the representativeness of SD2.1 self-conditioned artifacts. This is the load-bearing condition for the central claim of a bias-free, general training paradigm. The paper provides extensive empirical support (27 generators, including GANs and autoregressive models), and honestly states the limitation in Sec. 6, so a conditional verdict is appropriate. My proposed retraining experiment would directly test the representativeness assumption by varying the source generator while holding the rest of the pipeline fixed. Given the strength of the empirical results and the acknowledged limitation, I do not see a reason to change the reader's conditional verdict; adding this robustness check and reporting error bars would increase confidence.","tokens_in":24182,"tokens_out":15756,"duration_ms":148686,"concrete_test":"Retrain the full B-Free pipeline (DINOv2+reg, end-to-end, same content-augmentation protocol) using self-conditioned fakes from a different diffusion inpainting backbone, e.g., SDXL-inpainting, instead of SD2.1. Evaluate on the same evaluation suite (Synthbuster, the seven new generators, WildRF, GenImage-unbiased, FakeInversion, SynthWildX, and the autoregressive models), and compare the average balanced accuracy to the SD2.1-trained model. If the average bAcc shifts by more than 3 points, the paradigm is sensitive to the choice of the source generator, undermining the claim of general 'bias-free' artifacts. If the average is within 3 points, the recipe is robust across diffusion families and the representativeness assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of B-Free is that training on semantically aligned SD2.1 self-conditioned fakes yields a general real-vs-fake boundary. This requires that the artifacts in these reconstructions be representative of artifacts in unseen generators. The paper's evidence for representativeness is indirect: high end-to-end accuracy on 27 generators. No analysis shows that the feature distributions or spectral signatures of unseen fakes (e.g., StyleGAN, FLUX) overlap with SD2.1 training fakes. The high accuracy on GANs (e.g., StyleGAN 88.5 bAcc in Table 10) is particularly surprising since GANs do not use a diffusion process or the SD VAE; if the detector exploits a coarse cue (e.g., spectral envelope), the paradigm may not be capturing 'subtle artifacts' as claimed. The authors acknowledge in Sec. 6 that a completely different synthesis process would likely cause failure. Consequently, the 'bias-free' generality is conditional on an untested overlap between SD2.1 self-conditioned artifacts and the artifact manifolds of all current and future generators.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes B-Free, a training paradigm for AI-generated image detection. The authors construct a dataset of 51,517 real images from COCO and 309,102 synthetic images produced by self-conditioned inpainting with Stable Diffusion 2.1, plus content-augmented variants obtained by inpainting objects with same- or different-category replacements and background restoration. A Vision Transformer based on DINOv2+registers is fine-tuned end-to-end on this dataset. The paper reports balanced accuracy above 90% on average across 27 unseen generators, including recent models such as FLUX and SD3.5, along with improved calibration (lower NLL/ECE) compared with state-of-the-art detectors. Ablations vary the alignment strategy, augmentation type, architecture, and training dataset.","tokens_in":24289,"tokens_out":6380,"duration_ms":61919,"significance":"If the results hold, the paper makes a useful empirical contribution: it demonstrates that semantic alignment of real and fake training images, combined with content-based augmentation, yields detectors that generalize substantially better than standard ProGAN- or LDM-trained detectors. The evaluation is broad (27 generators, multiple independent datasets, ablations controlling architecture and training set), and the authors provide code and data, which supports reproducibility. The main weakness is that the interpretative claim of a 'bias-free' paradigm is stronger than the evidence directly supports: the representativeness of SD2.1 self-conditioned artifacts for unseen generators is inferred only from end-task accuracy, not from a direct analysis of the learned decision cues.","major_comments":[{"comment":"The central claim that B-Free is 'bias-free' and detects 'subtle artifacts introduced by AI generation' rests on the premise that self-conditioned SD2.1 reconstructions contain artifacts representative of all unseen generators. The paper provides only indirect evidence for this premise: high end-to-end accuracy on 27 generators. Figure 5 shows power spectra only for SD2.1 real/reconstructed/self-conditioned images; it does not compare with spectra or feature distributions of unseen generators such as StyleGAN, FLUX, or SD3.5. The strong balanced accuracy on GANs (e.g., StyleGAN 88.5% in Table 10) is not explained by the diffusion-specific rationale, and Sec. 6 concedes that a completely different synthesis process would likely cause failure. The authors should either (i) provide a direct artifact-overlap analysis (e.g., spectral signatures, feature-space distances, or per-generator error analysis) to support the representativeness assumption, or (ii) revise the abstract and Sec. 4 to state the scope more precisely (e.g., 'diffusion-like generators') instead of claiming a bias-free paradigm. As written, the abstract's causal statement 'allowing any differences to stem solely from the subtle artifacts introduced by AI generation' is not validated.","section":"Sec. 4.1 and Sec. 6"}],"minor_comments":[{"comment":"The construction of 'GenImage unbiased' is underspecified; please state the exact JPEG quality factor (or distribution of qualities) used for the fake images, how it was matched to the real class, and release the modified subset for reproducibility.","section":"Sec. 3.1"},{"comment":"All reported numbers are point estimates from single runs with no confidence intervals or repeated seeds; the 'bold with a margin of 1%' criterion is not interpretable without variance estimates, particularly for the small WildRF subsets and for ablation variants in Table 2 that differ by only 1-2%.","section":"Tables 2, 5, 7-10"},{"comment":"There are typos ('discrminate' in Sec. 1, 'adpots' and 'deatils' in Sec. B of the supplement, 'is it very likely' in Sec. 6); these should be corrected.","section":"Sec. 1 and Supplementary Material"},{"comment":"Please add axis labels and specify whether the power spectra are azimuthally averaged; without this detail the qualitative claim about low-frequency artifacts is hard to reproduce.","section":"Figure 5"},{"comment":"With M=15 bins and small test sets, ECE estimates can be biased; please report the average bin counts or use a debiased ECE estimator.","section":"Eq. (1)"},{"comment":"For a fair comparison, indicate the inference resolution used for each baseline and whether the authors' multi-crop averaging is also applied to the baselines.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds on prior work by the same group (e.g., FakeInversion [7], DMID [10], CLIP-based detection [11]), but the proposed dataset construction and augmentation strategy are distinct enough to constitute a novel contribution. The 'bias-free' terminology is likely to attract scrutiny; the editor may wish to require either additional artifact-overlap experiments or more modest wording in the abstract. The paper is within scope for a computer-vision forensics venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: B-Free is a genuinely useful empirical contribution to AI-generated image detection. The core idea—use self-conditioned SD2.1 reconstructions of real images, then augment with inpainted content edits, and fine-tune a DINOv2+reg ViT end-to-end—works better than text-paired or autoencoder-only alignment. On 27 generators, including GANs, recent diffusion models (FLUX, SD3.5), and autoregressive models, they get high balanced accuracy and, unusually, decent calibration at threshold 0.5 without per-generator threshold tuning. That is the thing worth knowing: the recipe is simple and the payoff is large.\n\nWhat's new: the combination of empty-mask self-conditioned inpainting plus category-consistent and category-inconsistent object inpainting as augmentation. The ablation in Table 2 cleanly isolates the contribution of each piece; the power-spectrum comparison (Fig. 5) gives a plausible mechanistic rationale for why self-conditioning beats autoencoder reconstruction. They also compare against a strong concurrent baseline [38] and show the diffusion pass matters. Reproducibility looks good: public code, data, and detailed implementation notes. Credit where due—this is careful, systematic work.\n\nSoft spots, in proportion: the 'bias-free' label overclaims. You train on SD2.1 artifacts only; the paper doesn't show those artifacts are representative of StyleGAN or FLUX in any direct feature-space sense. The high accuracy on GANs is strong indirect evidence, and the authors do admit in Sec. 6 that fundamentally different future synthesis processes would likely break it, so the core claim is honestly scoped. But the title and abstract push further than the evidence.\n\nSpecific technical concerns: no error bars anywhere; given many of the gains are >10 points, that's minor, but for calibration metrics and the smaller datasets it would help. They use a modified GenImage (fake images re-compressed) and don't report the original benchmark, so the comparison on that dataset is not exactly the standard. Also, the contribution bullet 'superior to 90%' holds on average but not on all datasets—FakeInversion is 86.2 bAcc, GenImage 89.3. That should be softened or reported per-dataset.\n\nWho this is for: anyone building or evaluating AI-image detectors, and researchers studying dataset bias in forensics. It deserves a serious referee. The missing error bars and benchmark modifications are fixable; the methodology itself is sound.\n\nRecommendation: send it to review, with a request for variance estimates, original GenImage numbers, and a more careful title. The empirical finding will stand.","headline":"Solid, useful training-data recipe for AI-image detection; the empirical case is strong, but the 'bias-free' framing overpromises and missing error bars need attention.","tokens_in":24905,"tokens_out":2951,"would_cite":true,"duration_ms":25370,"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":"This paper claims that training on self-conditioned Stable Diffusion fakes, which regenerate real photos so that real and fake differ only in synthesis artifacts, yields a detector that generalizes to unseen generators and stays…","keywords":["AI-generated image detection","deepfake detection","training data bias","semantic alignment","self-conditioning","Stable Diffusion inpainting","content augmentation","detector calibration"],"falsifier":"A concrete test: evaluate B-Free on outputs from a pure pixel-space autoregressive generator that does not use a latent autoencoder or iterative denoising, and where a detector trained on that generator's own fakes is highly accurate. If B-Free's balanced accuracy falls to chance (around 50%) while the control detector performs well, the claim that SD2.1 self-conditioned artifacts span the space of generative artifacts is falsified. A complementary observation is to measure the averaged difference spectrum between real and fake images for a new generator: if that spectrum is indistinguishable from real-versus-real differences at low frequencies, the detector should lose its main cue.","tokens_in":23922,"feed_emoji":"🖼️","tokens_out":8575,"duration_ms":70488,"temperature":0.7,"pith_summary":"AI-generated image detectors often fail in the real world because they learn spurious cues from their training data, such as compression, resolution, or subject matter, rather than the subtle artifacts left by synthesis. This paper argues the fix is not a better algorithm but a better training set, and proposes B-Free, a paradigm in which fake images are created by regenerating real photographs through Stable Diffusion's conditioning procedure. Because each fake is a self-conditioned reconstruction of a specific real image, the two differ only in generation artifacts, so the detector cannot shortcut on content or format. Trained end-to-end on these aligned fakes plus inpainting-based content augmentation, a ViT-based detector reaches about 96.4% balanced accuracy averaged over 27 generators, including unseen models like FLUX, Stable Diffusion 3.5, and autoregressive architectures, and remains well-calibrated at a fixed threshold.","feed_headline":"Image detector hits 96% accuracy by training on regenerated real photos","feed_subtitle":"New training data pairs each fake with its real source photo, giving detectors a bias-free learning signal.","key_machinery":"The load-bearing device is the self-conditioned reconstruction: an SD2.1 inpainting model applied with an empty mask to a real photograph, which forces the diffusion process to reproduce the same scene instead of generating a new one from a text prompt. This yields a fake image that is pixel-level aligned with its real counterpart, so the only differences are the low-level artifacts introduced by synthesis. Crucially, unlike plain autoencoder reconstructions, running the full diffusion pass deposits forensic clues at the lowest frequencies, as shown by averaged power-spectrum differences, giving the detector a broader artifact band to exploit. The second component is content augmentation through inpainting, replacing an object with a new one of the same or a different category, with or without restoring the original background, which forces the detector to rely on localized editing artifacts rather than global scene statistics. Together, these mechanisms let a detector trained on only one generative model, SD2.1, separate real from fake for 27 generators.","core_discovery":"B-Free builds a dataset of 51,517 real images from MS-COCO and 309,102 fakes, produced by passing each real image through the Stable Diffusion 2.1 inpainting model with an empty mask, so the diffusion process regenerates the same scene rather than generating a new one from text. The resulting self-conditioned images carry the forensic artifacts of diffusion synthesis while preserving pixel-level semantic content, and averaged power spectra show these artifacts extend to lower frequencies than those from a plain autoencoder reconstruction. Additional content augmentation, which replaces an object with one of the same or a different category and optionally restores the original background, further improves generalization and calibration. Fine-tuned end-to-end on this data, a DINOv2 Vision Transformer with registers, evaluated at 504 by 504 pixels without resizing, achieves an average balanced accuracy of 96.4% and an average AUC of 99.3% across 27 generators, outperforming the second-best state-of-the-art method by 20.7 percentage points in balanced accuracy, and maintains about 92% accuracy on web-scraped viral images months after upload.","pith_inferences":["Editorial inference: The paradigm can be turned into a continuously updating self-labeling pipeline. Any new diffusion-based generator can regenerate an existing corpus of real images to produce training data for the next detector without manual annotation; the paper trains only on SD2.1 fakes, so a natural extension is to pool self-conditioned reconstructions from several diffusion models and mea","Editorial inference: The power-spectrum rationale yields a testable predictor. For a new generator, one can compute the average difference between its output spectrum and that of real images, then compare it to the corresponding spectrum for SD2.1 self-conditioned fakes. Generators whose difference spectrum lies far from the training distribution, particularly at low frequencies, should be the har","Editorial inference: The paper's success on autoregressive models is attributed to the similarity between their tokenizers and Stable Diffusion's latent embedders. A discriminating stress test is a pure pixel-space autoregressive generator that shares no latent projection or iterative denoising module with Stable Diffusion; if the detector's balanced accuracy collapses on such a model, the general","Editorial inference: The content-augmentation recipes (same-category vs. different-category inpainting, background restoration) are coarse but suggest a broader design space. Varying the fraction of locally edited images, or mixing self-conditioned fakes from multiple diffusion steps, could offer a principled way to tune the artifact distribution that the detector must learn."],"forward_implications":["A detector trained on semantically aligned, self-conditioned fakes generalizes to generators never seen in training, including GANs, FLUX, Stable Diffusion 3.5, and autoregressive models, sustaining above 90% balanced accuracy on most of them.","Content-based augmentation through inpainting substantially improves robustness to JPEG compression, resizing, and blurring, and keeps accuracy high on images that have been reposted online over time, with accuracy around 92% even after many days from the initial post.","The B-Free training data improves existing architectures: CLIP- and RINE-based detectors gain large accuracy increases when trained on the B-Free dataset instead of ProGAN or Latent Diffusion data, suggesting that data design can matter more than architectural novelty.","Because the method improves calibration metrics (lower NLL and ECE), predictions can be thresholded at 0.5 without per-generator calibration, making the detector directly usable in realistic settings where calibration images are unavailable.","The approach reduces the gap between AUC and balanced accuracy, meaning high ranking performance translates into actually usable fixed-threshold decisions rather than merely good ordering."],"supporting_citations":[{"why":"Supplies the Stable Diffusion 2.1 inpainting model whose empty-mask self-conditioning is the core mechanism that creates semantically aligned fake images.","marker":"[39]"},{"why":"FakeInversion introduced the semantic-alignment principle for training detectors and provides the reverse-image-search dataset used for evaluation.","marker":"[7]"},{"why":"The ProGAN-based dataset and its augmentation scheme are the standard training data that this work replaces and compares against, establishing the generalization baseline.","marker":"[46]"},{"why":"Supplies the Latent Diffusion dataset used as an alternative training source in the ablation study that isolates the contribution of the training data.","marker":"[10]"},{"why":"The Vision Transformer with registers is the architecture used for the detector, and its design is credited for preserving local patch information needed to expose low-level artifacts.","marker":"[13]"},{"why":"DINOv2 provides the self-supervised pretrained backbone that is fine-tuned end-to-end on the B-Free data.","marker":"[34]"},{"why":"Synthbuster is a raw-format evaluation dataset whose real and fake images share the same compression history, used to rule out format bias in the experiments.","marker":"[2]"},{"why":"CoDE and its D3 dataset provide a text-driven, large-scale training-data comparison showing that B-Free's alignment yields better calibration and accuracy.","marker":"[3]"},{"why":"The concurrent dataset-alignment work is the direct alternative that uses only autoencoder reconstruction without the diffusion pass, and the paper's power-spectrum comparison shows why self-conditioning performs better.","marker":"[38]"},{"why":"Provides the SynthWildX and CLIP-based evaluation protocol and dataset used to measure in-the-wild performance on social-network images.","marker":"[11]"}],"fun_headline_variants":["Bias-free training: 96.4% accuracy on 27 AI image generators","Regenerate real photos, train detectors, beat SOTA by 20.7 points","Bias-free training regenerates real photos to spot any AI fake","Regenerate real photos to train detectors that beat 27 generators","Bias-free training: pair each fake with its real source to catch any fake"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the artifacts present in Stable Diffusion 2.1 self-conditioned reconstructions, produced by empty-mask inpainting, are representative enough of the artifacts of unseen generators, including GANs, FLUX, SD3.5, and autoregressive models, for a detector trained only on those fakes to learn a general real-versus-fake boundary.","fun_headline_variants_meta":{"raw":{"variants":["Bias-free training: 96.4% accuracy on 27 AI image generators","Regenerate real photos, train detectors, beat SOTA by 20.7 points","Bias-free training regenerates real photos to spot any AI fake","Regenerate real photos to train detectors that beat 27 generators","Bias-free training: pair each fake with its real source to catch any fake"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001328,"raw_usage":{"total_tokens":5419,"prompt_tokens":974,"completion_tokens":4445,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":4344}},"tokens_in":590,"tokens_out":4445,"duration_ms":29653,"temperature":1.0,"reasoning_tokens":4344,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:17:20.840685+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: evaluate B-Free on outputs from a pure pixel-space autoregressive generator that does not use a latent autoencoder or iterative denoising, and where a detector trained on that generator's own fakes is highly accurate. If B-Free's balanced accuracy falls to chance (around 50%) while the control detector performs well, the claim that SD2.1 self-conditioned artifacts span the space of generative artifacts is falsified. A complementary observation is to measure the averaged difference spectrum between real and fake images for a new generator: if that spectrum is indistinguishable from real-versus-real differences at low frequencies, the detector should lose its main cue.","supporting_citations":[{"cited_title":"Stable Diffu- sion","cited_arxiv_id":null,"evidence_quote":"Supplies the Stable Diffusion 2.1 inpainting model whose empty-mask self-conditioning is the core mechanism that creates semantically aligned fake images."},{"cited_title":"FakeInversion: Learning to Detect Images from Un- seen Text-to-Image Models by Inverting Stable Diffusion","cited_arxiv_id":null,"evidence_quote":"FakeInversion introduced the semantic-alignment principle for training detectors and provides the reverse-image-search dataset used for evaluation."},{"cited_title":"On the de- tection of synthetic images generated by diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the Latent Diffusion dataset used as an alternative training source in the ablation study that isolates the contribution of the training data."},{"cited_title":"Vision Transformers Need Registers","cited_arxiv_id":null,"evidence_quote":"The Vision Transformer with registers is the architecture used for the detector, and its design is credited for preserving local patch information needed to expose low-level artifacts."},{"cited_title":"DINOv2: Learning Robust Visual Features without Supervi- sion","cited_arxiv_id":null,"evidence_quote":"DINOv2 provides the self-supervised pretrained backbone that is fine-tuned end-to-end on the B-Free data."},{"cited_title":"Synthbuster: Towards detection of diffu- sion model generated images","cited_arxiv_id":null,"evidence_quote":"Synthbuster is a raw-format evaluation dataset whose real and fake images share the same compression history, used to rule out format bias in the experiments."},{"cited_title":"Contrasting Deepfakes Diffusion via Contrastive Learning and Global-Local Similarities","cited_arxiv_id":null,"evidence_quote":"CoDE and its D3 dataset provide a text-driven, large-scale training-data comparison showing that B-Free's alignment yields better calibration and accuracy."},{"cited_title":"Raising the Bar of AI-generated Image Detection with CLIP","cited_arxiv_id":null,"evidence_quote":"Provides the SynthWildX and CLIP-based evaluation protocol and dataset used to measure in-the-wild performance on social-network images."}],"review_version":1}