- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This work proposes a systematic study of Implicit Co-occurring Objects (ICO) in text-to-image (T2I) models, demonstrating that ICOs are observed in the generations of recent, diverse T2I architectures. The authors investigate several hypotheses regarding ICOs' sources, eventually indicating a strong impact of co-occurrences in the training data, early denoising steps, and standard sampling hyperparameters. The authors propose several control approaches that reduce leakage with small degradation in generation fidelity. Additionally, the authors show that ICOs cannot be freely suppressed without degrading fidelity when training data already contains object co-occurrences.
Critical strengths
I think the paper's main motivation is strong. It is worth investigating the phenomenon of implicit co-occurring objects (ICOs), which can be harmful as the biased outputs of text-to-image models are often used as training data for downstream tasks (e.g., object detection). The authors, using 7 diverse T2I pipelines on 80 COCO categories, show that the analyzed models generate ICOs.
Critical weaknesses
W1. Lack of novelty
I consider some of the main claims from this work to be incremental relative to prior studies, raising concerns about novelty.
Some of the proposed hypotheses are not new and have already been investigated in similar problems involving Diffusion Models. Precisely:
- Hypothesis 1 - training data patterns. Liu et al. [1] study concepts that are generated in images despite not being specified in the prompt and attributed them to the training data.
- Hypothesis 4 - sampling dynamics. Roos et al. [2] present a correlation between the model's bias level and the sampling hyperparameters: the guidance scale π€ and the number of timesteps.
Also, the authors claim that their work is the first systematic study, but there are studies that weaken this claim. Wu et al. [3] investigate "unguided objects" in Stable Diffusion for biases - the objects that are generated but not mentioned in the prompt, along with their co-occurrence.
The method related to steering text embeddings has already been used in a similar setup. Koh et al. [4] investigate entangled content and introduce a delta vector to modify text embeddings.
[1] Liu, Zhili, et al. "Implicit concept removal of diffusion models." ECCV 2024.
[2] Roos, Nathan, et al. "How I Met Your Bias: Investigating Bias Amplification in Diffusion Models." WACV 2026.
[3] Wu, Yankun, et al. "Stable diffusion exposed: Gender bias from prompt to image." AIES 2024.
[4] Koh, Eunseo, et al. "Translation of Text Embedding via Delta Vector to Suppress Strongly Entangled Content in Text-to-Image Diffusion Models." ICCV 2025.
W2. Presentation issues
It is problematic to understand the paper during reading, i.a.:
- There are many details missing in the description. E.g., what are and in the co-occurrence rate metric?
- In Stage 1 (Section 5), the authors design controls inspired by the observations from Section 4. To me, it's not clear which observations suggest which methods.
- Missing descriptions of methods used in Stage 3.
- Some hyperparameter values are provided without reasoning: Spearman , detector thresholds.
W3. Lack of motivation for some choices
- The choice of methods for Stage 3 seems to be ad hoc. No reasoning is provided for why they are good candidates.
- Choice of metrics is unusual. The authors provide their own for the object co-occurrence. Why do you not compare methods in overall generation quality (FID)? Also, why are standard concept unlearning metrics (Unlearning Accuracy, Retain Accuracy) not used for concept-erasure comparisons?
- What's the goal of continuous control in this exact setup? I cannot find information about the motivation.
Questions for authors
Q1. Please explain how your experiments and claims relate to my novelty concerns from W1.
Q2. Could you fill in the presentational gaps that I mentioned in W2?
Q3. Please provide motivation for the choices I listed in W3.
5 comments
Response to W2: presentation
We will improve our draft and clarify the confusing points.
(1) Notation in the co-occurrence rate.
- π is the set of MS-COCO 2017 train+val images with at least one instance annotation. That is 122,218 human-annotated photos (117,266 train + 4,952 val).
- For an image I β π, we write c β I when category c is annotated in I. This is presence per image.
- So Ο(c, o) = |{I β π : c β I β§ o β I}| / |{I β π : c β I}| is the fraction of COCO photos containing c that also contain o.
We will state this explicitly in Section 2.
(2) Which observations motivate which controls. Each Stage 1 control targets one finding from Section 4. We will add this mapping to the Stage 1 table.
| Stage 1 control | Observation it targets (Section 4) |
|---|---|
| Isolation phrasing | H1: models reproduce training-data co-occurrence, and fine-tuning on isolated objects on a plain surface causally reduces leakage. Isolation phrasing requests that same low-co-occurrence part of the data (an isolated object on a plain background) at inference time. |
| Template diversity | H1 (inference-time proxy): if leakage is tied to one overrepresented caption pattern ("a photo of"), paraphrasing it should change the leak rate. |
| Embedding-space steering | H2: text embeddings carry a weak co-occurrence signal, so we remove the prompt embedding's component along the leak-object direction. |
| Negative prompting | H2 + H4: guidance modulates leakage (H4), and its effectiveness tracks the text encoder's co-occurrence signal (H2). |
| Early CFG suppression | H3 + H4: ICOs are planned early (20β30% of the steps for most models), so negative guidance is applied only before each model's median first-detection step. |
| Detect and correct | H3: ICOs become detectable early, so at each model's median first-detection step we decode the preview image, detect the leak object, and apply negative guidance for the remaining steps only to images where it is detected. |
| Guidance and step reduction | H4: higher guidance and more steps increase leakage. |
| Scheduler substitution | H4: the sampling trajectory modulates leakage. |
(3) Description of Stage 3 methods. We will add an appendix subsection describing each method and its hyperparameters. In brief:
- Loss-guided sampling: loss-based guidance in the spirit of universal guidance, with CLIP as the loss.
- During the first 8 of 20 DDIM steps, the predicted clean image is decoded.
- Its CLIP ViT-L/14 similarity to "a photo of a dining table" is backpropagated to the latent.
- The latent is moved against that gradient (scale 40) before the regular denoising step.
- Cross-attention suppression: the leak object never appears in the prompt, so there is no leak token to suppress.
- A custom attention processor boosts the prompted category's token(s) by 6Γ.
- It scales every other ("unprompted") token by 0.3Γ. The weights are then renormalised.
- This is applied in all cross-attention layers at every step, so the image is pulled toward the prompted object alone.
- Post-hoc inpainting: on images where OWLv2 detects a dining table, the detected boxes (dilated by 24 px, feathered) are inpainted with the model family's inpainting checkpoint (SD-1.5-inpainting, SDXL-inpainting-0.1, FLUX.1-Fill-dev). The inpainting prompt is the original prompt.
- Post-hoc SDEdit: the finished image is re-noised to strength 0.75, past the point where ICOs are committed (Section 4). It is then denoised again with the original prompt, using the base model's own image-to-image pipeline. No negative prompt is used.
- Best-of-N: N = 4 candidates are denoised jointly up to step 6 of 20.
- OWLv2 checks their predicted clean images for a dining table, and only the first clean candidate is completed. If none is clean, candidate 0 is completed.
- The selector uses the same detector as our evaluation, but it sees only the early preview, never the final image. A clean preview does not guarantee a clean final image, so the selection does not trivially optimise the evaluation metric.
(4) Hyperparameter choices. Sorry for the confusion.
- Spearman threshold. Spearman Ο β₯ 0.3 is not a tuned hyperparameter. It is a conventional statistical boundary for a moderate correlation (0.1 small, 0.3 moderate, 0.5 large). We use it only as a descriptive label and report the exact Ο and p-values alongside it.
- Detector threshold. We use OWLv2 with a fixed confidence threshold of 0.15 throughout.
- Robustness. Our conclusions do not depend on this choice:
- Replacing the fixed cutoff with an area-under-curve leak rate over thresholds 0.15β0.95 gives the identical ranking of all seven models by Excess Leakage Score.
- On a stratified sample of 300 SD1.5 images (150 OWLv2-positive, 150 negative), OWLv2 agrees with BLIP in VQA mode ("Is there a dining table?") on 88.3% of images.
Response to W3: motivation
(1) Choice of Stage 3 methods. Stage 3 deliberately goes beyond our four hypotheses. It covers the main families of inference-time control:
- gradient guidance (loss-guided sampling);
- attention manipulation (cross-attention suppression);
- post-hoc editing (inpainting, SDEdit);
- search (best-of-N).
Where a method needed a timing choice (guided steps, checkpoint step, SDEdit strength), we reused the choices from Section 4 rather than tuning it. We will add this rationale to Section 6.
(2) Choice of metrics.
- Counterparts of erasure metrics. Erasure metrics assume the concept is named in the prompt, which never holds for ICOs, so we report their counterparts:
- 1 β leakage for Unlearning Accuracy;
- fidelity (the detection rate of the prompted object on the 79 non-target categories) for Retain Accuracy;
- in addition, preservation of the target object when it is itself prompted.
- Quality metrics. We added FD-DINOv2 (FID's FrΓ©chet distance on DINOv2 features, which agrees better with human judgment), ImageReward, CLIPScore, LPIPS diversity and Vendi. Please see our response to Reviewer giV8 W2 for a detailed table of comparison on those metrics.
- Results. Our controls (steering, isolation, and the two combined) leave CLIPScore within 0.3 of the baseline and do not reduce diversity. The cost is moderate in ImageReward, and there is a measurable distribution shift (FD-DINOv2 to the baseline images).
(3) Goal of continuous control.
- Why a dial. ICOs are not always unwanted. Context aids realism, and a synthetic-data pipeline may want to set a co-occurrence rate rather than remove it.
- Mechanism. The dial is a single signed strength. On the six leak-prone prompts in SD1.5, it moves the dining-table rate monotonically from 50% to 8% at fidelity (baseline 37%). The same rate needs a different strength for each prompted category, so the dial is calibrated per category.
- Effect. Run on all COCO categories, the dial has a large range almost only for the leak-prone categories and close to none for most others. The dial modulates the ICOs a trained model already encodes in finer intervals.
Thank you for your valuable feedback. Below we address the weaknesses.
Response to W1: novelty
We thank the reviewer for these pointers. The works are relevant, and we agree that several individual components of our study, such as probing the training data, varying sampling parameters or editing text embeddings, have precedents. We will discuss each work explicitly and clarify our novelty with respect to it in the revision.
We would like to emphasis that the contribution we claim is the systematic study as a whole: ICOs measured at scale (seven architecturally distinct models from 2022 to 2026, 80 categories, 56,000 images, four independent probes) and compared against real-world co-occurrence statistics; four hypotheses on their origin tested side by side on the same data, including a causal training intervention; controls derived from these findings and compared with nine concept-erasure methods; and a theoretical account of why suppression trades off against faithfulness to the data. To our knowledge, no prior work brings these together for this problem, and none offers such a theory.
Below we note how each work differs from ours.
Liu et al. (2024) and H1. They fine-tune away unwanted artifacts such as watermarks. On the contrary, ICOs are not artifacts, they are natural scene objects that track real-world co-occurrence. And our control is training-free.
Roos et al. (2026) and H4. Our H4 agrees with their finding that guidance and timesteps modulate bias, and extends it to object co-occurrence. We will add a discussion in the revision.
Wu et al. (2024) and "first systematic study". They study how unguided objects differ across gendered prompts, while we study ICOs themselves, at a large scale, against real-world co-occurrence rates.
Koh et al. (2025) and embedding steering. SSDV adds a delta vector to one word's token, while we project the target direction out of the whole prompt embedding. We have conducted an analysis: at matched leakage, SSDV costs more quality (ImageReward β0.69 vs. β0.17). We will include a detailed discussion in the revision.