acceptodds
Withdrawn from ICLR 2027

Flow-CIR: Generative Feature Rectification for Noise-Robust Composed Image Retrieval

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Composed image retrieval aims to retrieve a target image according to a reference image and a textual modification. Existing methods usually assume that training triplets are correctly matched, while real-world annotations may contain noisy triplet correspondence caused by subjective modification texts, ambiguous visual changes, and imperfect target selection. Such noise is especially challenging because a suspicious triplet is not always completely wrong; it may still preserve partial semantic cues that are useful for learning fine-grained visual transformations. Most noise-robust CIR methods mainly identify, suppress, or down-weight unreliable triplets, which reduces harmful supervision but may also discard informative signals from hard samples. In this paper, we propose Flow-CIR, a feature-level generative rectification framework for noise-robust composed image retrieval. Instead of simply forgetting suspicious triplets, Flow-CIR learns a conditional semantic vector field from reliable triplets and uses it to generate pseudo-target features for suspicious samples. To avoid over-correction, the rectified target feature is softly mixed with the observed target representation and is applied only after a warm-up stage. Since rectification is performed in the compact feature space, Flow-CIR introduces no additional inference cost. Experiments on FashionIQ and CIRR under different noise ratios demonstrate the robustness and effectiveness of the proposed framework.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Composed Image Retrieval, Noise-Robust Learning, Generative Feature Rectification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_cPTXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates supervised composed image retrieval in the presence of noisy triplet correspondence. The authors propose Flow-CIR, which categorizes training triplets into "reliable" and "suspicious" groups based on query-target similarity. It utilizes reliable triplets to learn a conditional vector field that generates pseudo-target features for suspicious triplets, subsequently performing soft fusion between these generated features and the observed target features. This method applies feature rectification solely during the training phase, while adhering to the standard retrieval pipeline during inference. The authors conducted experiments on the FashionIQ and CIRR datasets using varying levels of synthetic noise and report results regarding component ablation, parameter sensitivity, and qualitative analysis.

Critical strengths
  1. To address the issue in noisy triplet supervision where "suspicious samples may still contain some useful semantic information," Flow-CIR proposes an active feature correction approach rather than simply discarding or down-weighting such samples; the motivation behind this is quite clear.
  2. The method performs feature rectification during the training phase and introduces no additional modules or computational steps during inference. The authors conducted experiments across various noise levels and provided component ablation studies as well as parameter sensitivity analyses.
Critical weaknesses
  1. The fidelity score derived from Equation (6) in the paper evaluates to zero when the normalized similarity is either 0 or 1. Consequently, a sample with the highest query-target similarity could receive the lowest fidelity score and be flagged by the model as a "suspicious sample." This contradicts the paper's explanation that "low similarity implies lower reliability." The authors need to clarify whether the actual implementation aligns with the formula used to calculate the fidelity score.
  2. The paper specifies the synthetic noise ratio σ\sigma but does not explain how the hyperparameter ρ\rho is selected or whether it is related to σ\sigma. Since ρ\rho determines the number of samples undergoing rectification in each batch, this information is crucial for reproducing the experiments and interpreting the results.
Questions for authors
  1. Was Equation (6) implemented exactly as presented in the paper? If not, please specify the formula actually used and the corresponding code.
  2. Please specify the exact value and selection method for ρ\rho, as well as its relationship with the noise ratio σ\sigma.
Reviewer_YSdrOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This is a paper that trains a model for composed image retrieval. The training set consists of triplets (ref-image, modifier text, target image). The claim is that such datasets are noisy. The author identify noisy examples and refine them with a feature-level flow-matching that maps composed representations into easier/correct targets. The flow-matching model is trained seperately on data identified as non-noisy to map the compoosed example to the target. Performance is evaluated on CIRR and FashionIQ - two known datasets for this task.

Critical strengths
  • The motivation is clear: the paper identifies noisy training samples and improves them by refining the target features. It is true that collecting training triplets for composed retrieval is quite challenging, therefore, expecting noise is a realistic assumption.
  • There is no inference time overhead. The proposed method only acts in the training.
  • The performance gains are more visible at the large noise rate, exactly where they are expected.
Critical weaknesses

W1 - some baselines that I 'd expect to see:

  • in the linear combination of (16), mix the target feature with the composed feature
  • it is unclear what the noise-aware selection is in Table 3 because it is not described but we should see the baseline that instead of refining the noisy examples, it skips them from the training
  • no flow matching, just a one-step MLP regressor trained on reliable pairs

W2 - the suspicious ratio rho seems a key parameter but its value is never mentioned. Is this a constant among all noise values in Tables 1 and 2? If it follows the noise this indicates leaking, which is unrealistic.

W3 - the fidelity score in (6) is not discussed at all. This is a non monotonic function with a peak around \bar{s}_i = 0.7. This is counterintuitive as one would expect the largest values to be chosen as noisy and not realiable. Right now, the largest values will be added to the realiable set.

W4 - minor, but there are several repetitions in the text, the same thing said again and again, often in nearby places, eg see top of sec 3.3 in page 4.

W5 - details about the architecture of the composition module C_\phi are missing. |

W6 - strong baseline and training repice: the base recipe in Table 3, noise = 0.5, is already achieving higher performance than some of the approaches of Tab1-2. Does this have to do with borrowing the BLIP-2 encoders? What are the encoders of the other approaches? At the moment, this make it sound like an apple vs oranges comparison.

W7 - I know that both CIRR and FashionIQ have been established datasets and there have been tens of papers evaluating on them, but they are a distraction. There are a few reasons for that and I will explain. I have personally gone through tons of examples in those and in other CIR datasets.

  • For a large number of queries, the text solely does all the work, there is no composition required. You can clearly understand this qualitatively (a photo of a soft drink and "show three bottles of soft drink", the positives are the photos with 3 bottles of soft drinks)
  • There is a large amount of noise in the ground truth, in terms of false negatives. The labels include one or few positives, while all the rest are not labeled.
  • For most of the queries what you need from the image is to recognise the main object catogory, once you do this, then you don t need anything else from the image.
  • The above point has to do with the class definition, which is at the semantic-category level. The two photos, query and positive need to show an object of the same semantic-category, which is the object id does not matter or spatial layout or anything else about specific appearance of the image. These datasets do not reflect a real-world setting (toy datasets to help progress would be ok, but all the above issues made them just bad datasets), while there have been 2 datasets out there that focus on the correct direction. This is, we need to retrieve objects of the same identity (the same particular object) under different conditions. [a] Psomas et al, neurips 2025, i-CIR: Instance-Level Composed Image Retrieval [b] Rosasco CVPR 2024 ConCon-Chi: Concept-Context Chimera Benchmark for Personalized Vision-Language Tasks. Both [a] and [b] can really drive progess in CIR while CIRR and FashionIQ have been providing a very noisy measurement.
Questions for authors

Experimens are requires in W1, clarificiations in W2,3,5,6, while W6 may require some more direct comparisons too. W7: the fact that a lot of papers have followed that path, to me, does not mean that we need to continue accepting papers that do it wrongly, when there are alternative datasets or possibilities to create new ones.

Reviewer_u7qsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies noisy triplet correspondence in composed image retrieval (CIR). Noting that suspicious triplets may still contain useful semantics, the authors propose Flow-CIR, a training-time feature-level rectification framework. During training, Flow-CIR first separates reliable and suspicious samples by ranking their within-batch fidelity, then learns a conditional vector field from reliable triplets and uses it to generate pseudo-target features for soft rectification of suspicious samples. The authors compare Flow-CIR with several noise-robust baselines on FashionIQ and CIRR under multiple synthetic noise ratios, and validate each component through ablation studies and visual analyses.

Critical strengths
  1. The authors' research perspective is reasonably well motivated. Rather than treating suspicious triplets as noise to be discarded, they view them as "rectifiable supervision," shifting the problem from whether a sample is trustworthy to how to extract useful signal from an untrustworthy one.
  2. The method is practically appealing. The proposed modules operate only during training and are removed at inference, adding no retrieval latency. Moreover, since they act on features, they could in principle be combined with different composition backbones.
Critical weaknesses
  1. Eq. (6) doesn't hold up: the curve peaks around 0.707 and returns to zero at 1, so a perfectly clean triplet scores exactly the same as a completely wrong one. That is not a harmless quirk — the suspicious set is the bottom-ρB\rho B by this score, so the best-matched triplet in each batch lands in the same lowest bucket as the worst ones and can be sent to the rectifier.
  2. The rectification direction is not guaranteed. Flow is trained with v⋆=Ft−Fcv^\star=F_t-F_c, i.e., it is asked to transport FcF_c to the very noisy target that the method has just decided to correct. If the field is learned successfully on reliable triplets, its integral is simply FtF_t and the correction is identically zero. More importantly, both the reliability decision and the correction are driven by the model's own similarity, which is weak evidence for the claimed effect.
  3. The method is essentially a compositional application of linear interpolation. With a straight path and velocity Ft−FcF_t-F_c, the one-step solution is just linear interpolation, and K=3K=3 only approximates it numerically; yet the paper provides no analysis to support its claim of being "different from fixed interpolation." The ablation shows that the flow loss contributes only 0.1 to 0.2 points, less than sample selection and the memory queue, which raises doubts about its effectiveness.
  4. The experimental scope and depth are seriously insufficient, and the central claim remains unvalidated. The main evidence comes only from random-replacement synthetic noise on two benchmarks, FashionIQ and CIRR; there is no test with different backbones and no report on the accuracy of the suspicious-set decision. In addition, the authors do not provide a rich experimental design around their own core mechanism.
Questions for authors
  1. Eq. (6) bothers us. The score drops back to zero as sˉ\bar s gets close to 1, yet the text says the calibration is supposed to suppress scores below the background boundary. Those two do not line up. So how are the reliable and suspicious sets actually picked during training? It is possible we are reading the equation wrong, and we would like to be corrected if so.
  2. Since the ground-truth noise indices are known in the synthetic setting, why is there no report on the precision and recall of the suspicious set against the true noisy set? This result would directly test the premise of the method, and we ask the authors to provide it.
  3. If the field is learned well on reliable triplets, integrating it should land on FtF_t, which means the correction is zero. What we cannot see is what makes the extrapolation on suspicious samples move toward something semantically right, instead of just collapsing toward the batch mean. Some explanation of that would help.
  4. The ablation is missing a "memory queue only, no rectification" arm, so we cannot tell whether the gains come from the rectification itself or simply from a larger negative pool. Adding that control would settle it. It would also help to see whether the framework still works when the backbone is swapped.

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

Related papers

Discussion (0)

Sign in to comment.