Dehate Shared Task: Image Dehatification

• The paper introduces the Dehate Shared Task, a challenge that uses segmentation to identify and mitigate hateful regions in synthetically generated images.
• It employs a multimodal dataset combining Hatenorm text prompts, Stable Diffusion images, and DAAM-derived hate attention maps to create dehated images.
• Evaluation is based on Intersection over Union (IoU) scores, highlighting the complexity of precise localization and the potential for advanced segmentation improvements.

The Dehate Shared Task is a segmentation-based content moderation challenge hosted at the Defactify 3 workshop, the Third Workshop on Multimodal Fact Checking and Hate Speech Detection, co-located with AAAI 2024. It addresses image dehatification: given an image generated from a hateful prompt, a system must predict a mask, or an equivalent blurred component, corresponding to the hateful components in the image. The task is defined together with the DeHate dataset, a multimodal resource of $2411$ synthetic instances produced from Hatenorm text, stable-diffusion-2-base image generation, and DAAM-derived hate attention maps, and with DeHater, a vision-language baseline for multimodal dehatification (Dalal et al., 26 Sep 2025).

1. Scope, motivation, and formal task definition

The task is framed within the broader problem of hateful multimodal content, especially images paired with text such as tweets, prompts, and captions on social media. The motivating claim is that large-scale AI systems increasingly operate on multimodal data and are often trained on web-scale corpora that contain hateful, abusive, or NSFW visual content. This produces both ethical and trustworthiness challenges: models can learn to reproduce or amplify harmful stereotypes, and platforms struggle with moderating images, which are harder to filter than text. In response, the Dehate Shared Task emphasizes hate mitigation, not merely hate detection, and explicitly moves beyond binary hate/no-hate classification toward localized, region-level mitigation (Dalal et al., 26 Sep 2025).

The task is described in the paper as an image dehatification problem conceptualized as unsupervised image masking. Operationally, it is a segmentation, localization, and mitigation task on images generated from hateful text. The released resource provides the original generated image and the associated blurred output or mask; the shared-task formulation focuses on predicting the blurred component or segmentation mask indicating hateful regions. The paper describes a single main track, namely segmentation-based dehatification of generated images. No separate tracks for detection versus removal, or unimodal versus multimodal settings, are described. The focus is on localization and masking, not classification (Dalal et al., 26 Sep 2025).

Evaluation uses Intersection over Union (IoU) on the test set. If $M_{\text{pred}}$ is the binary mask predicted by a system and $M_{\text{gt}}$ is the ground-truth blurred or hateful component mask, then

$$\mathrm{IoU}(M_{\text{pred}}, M_{\text{gt}}) = \frac{|M_{\text{pred}} \cap M_{\text{gt}}|}{|M_{\text{pred}} \cup M_{\text{gt}}|}.$$

The task ranks participants by the IoU computed between the predicted blurred component and the ground-truth blurred component in the test dataset. This implies ranking by average IoU across test images (Dalal et al., 26 Sep 2025).

The paper does not give a formal mathematical definition of hate as a semantic category, but it does provide an operational definition of hateful regions. Hateful text comes from the Hatenorm dataset, prompts are built by amalgamating tokens indicative of hateful terms, and the DAAM pipeline computes cross-attention maps between text tokens and image pixels. Pixels whose DAAM attention values exceed a threshold for hateful spans are treated as hateful regions. In conceptual form,

$$M(x,y) = \begin{cases} 1 & \text{if } A(x,y) \geq T \ 0 & \text{otherwise.} \end{cases}$$

Pixels with $M(x,y)=1$ are then blurred or erased and become the ground-truth hateful component used in training and evaluation (Dalal et al., 26 Sep 2025).

2. Dataset composition and source data

The textual source for the task is the Hatenorm dataset, described as a parallel corpus of hate texts and their normalized counterparts. It is designed to “make the texts less hateful and more benign” while preserving semantics, and it captures “varying degrees of hatefulness” while identifying key hateful phrases. Within Dehate, Hatenorm supplies hateful spans used to construct image generation prompts and provides a linguistically grounded basis for associating textual hate with generated images (Dalal et al., 26 Sep 2025).

From this source, the authors create the DeHate dataset, which comprises two distinct components: Images Generated from Prompts and Images with Blurred Hateful Components. Image generation uses the stable-diffusion-2-base latent diffusion model. Prompts are constructed by amalgamating tokens indicative of hateful terms from tweets in Hatenorm, and when tweets are too long for the prompt size limit, the authors select relevant segments focusing on hateful content. The blurred versions are produced by a DAAM pipeline that generates hate attention maps linking pixels to prompt components, then applies a global heatmap threshold to produce a binary mask of pixels associated with hateful content (Dalal et al., 26 Sep 2025).

The dataset contains, for each instance, the original generated image and the corresponding blurred image, where hateful regions have been anonymized. An implicit mask is derivable from the difference or from the DAAM attention map. The resource is explicitly released as part of the shared task.

Split	Instances
Train	1687
Test	724
Total	2411

The construction procedure uses two image-space modifications. In an erasure stage, masked pixels are set to black $[0,0,0]$. In a blur stage, those pixels are replaced by the average color of a local box around each pixel, yielding what the paper calls a “nuanced blurring effect.” This paired-image design is central to the task because it provides both the original hateful image and its dehated counterpart (Dalal et al., 26 Sep 2025).

3. Stable Diffusion, DAAM, and the construction of hateful-region masks

The Dehate Shared Task depends on a specific synthetic-data pipeline built from Stable Diffusion and the Digital Attention Analysis Module (DAAM). Stable Diffusion 2-base provides synthetic images conditioned on hateful prompts, which allows control over content without scraping real harmful imagery. The paper also refers conceptually to “watermarked, stability-enhanced, stable diffusion techniques,” although the exact watermarking details are not elaborated (Dalal et al., 26 Sep 2025).

DAAM, based on Tang et al. (2022), is used to interpret Stable Diffusion through cross-attention maps. For each prompt token, DAAM aggregates attention maps from the denoising module into a two-dimensional heatmap highlighting pixel-token correlations. In the Dehate pipeline, this makes it possible to isolate the regions corresponding to hateful spans in the prompt. The resulting hate attention maps are then thresholded to derive binary masks and to blur or remove the associated regions in the image (Dalal et al., 26 Sep 2025).

The pipeline proceeds in three stages. First, hateful text from Hatenorm, or tokens indicative of hate, conditions Stable Diffusion 2-base to generate an image $I$. Second, DAAM computes cross-attention maps between prompt tokens and image latents during denoising, and the attention maps for hateful spans are combined into a heatmap $A(x,y)$. Third, thresholding yields a mask $M(x,y)$, and the masked pixels are altered. The paper describes a two-step image modification: initially setting high-attention pixels to black, then, for heatmap value $M_{\text{pred}}$0, computing a local average color and assigning it to produce a blurred region (Dalal et al., 26 Sep 2025).

This pipeline has a dual role. It is used to build the dataset, and it also defines what participating systems are expected to approximate. A plausible implication is that Dehate formalizes mitigation as a supervised approximation to a generative-attention-based annotation procedure rather than as direct manual segmentation. The paper’s examples underscore this connection by showing raw Stable Diffusion images, dehated versions, and model outputs (Dalal et al., 26 Sep 2025).

4. DeHater baseline model

DeHater is the baseline system supplied for the shared task. It is a vision-LLM designed to produce masks for hateful image regions from images together with textual prompts or hate spans. Its encoder is a frozen CLIP model, which acts as a joint language-image encoder. CLIP provides spatial image feature maps and text features, including token embeddings and a pooled CLS-like representation (Dalal et al., 26 Sep 2025).

The decoder is transformer-based and is inspired by U-Net. U-Net-style skip connections connect intermediate CLIP feature maps to corresponding decoder stages, and CLIP activations, including the CLS representation, are injected into internal activations at each block. The stated purpose is to enhance contextual alignment between text and image. Conditioning is performed with Feature-wise Linear Modulation (FiLM), where a textual conditional vector $M_{\text{pred}}$1, derived from hate spans, modulates decoder features (Dalal et al., 26 Sep 2025):

$M_{\text{pred}}$2

To construct the conditioning vector, multiple hate span embeddings from CLIP’s text encoder are combined through a learnable projection network:

$M_{\text{pred}}$3

This architecture condenses several token-level hate representations into a single vector that conditions decoding (Dalal et al., 26 Sep 2025).

The output is a binarized image or segmentation mask. The decoder produces a per-pixel score map $M_{\text{pred}}$4, and after sigmoid and thresholding it yields a binary mask $M_{\text{pred}}$5 marking hateful regions. The paper states that “the output of our architecture is the production of a binarized image, which represents the final masked output.” Given an image and the associated prompt or hate spans, DeHater encodes image and text via CLIP, conditions the decoder on hateful spans, and produces a mask that can be applied to generate a dehated image through blurring or erasure (Dalal et al., 26 Sep 2025).

The paper does not explicitly write down the loss function in LaTeX, and it does not present detailed ablation studies such as with-versus-without FiLM, with-versus-without DAAM-derived masks, or CLIP-versus-other encoders. This omission is significant because the benchmark is defined clearly, but the optimization details of the baseline are only partially specified (Dalal et al., 26 Sep 2025).

5. Shared-task procedure and benchmark results

The evaluation protocol follows a conventional shared-task release pattern. Initially, labeled training data—images together with blurred versions or masks—were provided. Later, an unlabeled test set was released. Participants trained on the labeled data and submitted predictions for the test images. The paper reports $M_{\text{pred}}$6 registrations and 5 official teams submitting results on the test set, in addition to the baseline (Dalal et al., 26 Sep 2025).

The winning system, UniteToModerate, combines NExT-Chat and UniFusion. NExT-Chat provides initial masks via a pix2emb method, and UniFusion performs hierarchical fusion of visual and reference features to refine masks. Other submitted teams were PaulJane, Markans, Sanskarfc, and rachitmodi (Dalal et al., 26 Sep 2025).

The leaderboard reported in the paper is as follows:

Rank	Team	IoU score
1	UniteToModerate	0.55
2	PaulJane	0.51
3	Baseline (ours)	0.49
4	Markans	0.48
5	Sanskarfc	0.47
6	rachitmodi	0.44

The baseline DeHater obtains an IoU of 0.49. Two teams surpass it, while three teams fall below it. The best score, 0.55 IoU, is explicitly interpreted in the paper as evidence that the task is challenging and that there remains considerable room for improvement. The paper also points qualitatively to the difficulty of precise hateful-region localization in complex generated scenes and to the need for more advanced segmentation and multimodal fusion methods, as illustrated by the NExT-Chat plus UniFusion system (Dalal et al., 26 Sep 2025).

6. Ethical motivations, limitations, and relation to adjacent shared tasks

Ethical motivation is central to the Dehate Shared Task. The stated aim is to support ethical AI and content moderation by identifying and removing hateful elements in images rather than simply flagging them, while allowing non-hateful portions of images to remain visible. The use of synthetic data generated via Stable Diffusion and Hatenorm is presented as a way to avoid distributing real-world harmful images while still capturing the semantics of hateful content for research (Dalal et al., 26 Sep 2025).

The paper also notes several limitations. Real-world datasets are described as “complex, messy, and unstructured,” and they may carry biases. The authors briefly mention using a Debiased LLM and Dataless Model Merging to enhance dataset diversity and reduce inherent biases in the textual prompts. At the same time, the benchmark remains English-only and is built from a specific hate normalization corpus. IoU scores around 0.5–0.55 indicate that precise hateful-region segmentation is difficult. The paper further notes the risk of misuse: any system that understands hateful content could theoretically be repurposed to generate or target such content. Future directions mentioned include using an LLM to justify the output of the mitigation pipeline and extending the approach to other languages and modalities (Dalal et al., 26 Sep 2025).

Within the broader shared-task landscape, Dehate is distinctive because it operationalizes hate mitigation as region-level masking in images rather than as text-only classification. Nearby tasks in the same research ecosystem remain largely text-centered. The VLSP 2019 Hate Speech Detection for Social Good shared task, for example, is a Vietnamese Facebook multi-class classification problem with labels HATE, OFFENSIVE, and CLEAN (Vu et al., 2020). HASOC 2021 includes both binary HOF/NOT and fine-grained HATE/OFFN/PRFN/NONE formulations, and joint multilingual transformer training has been shown to improve English, Hindi, and Marathi performance in that setting (Mohtaj et al., 2022, Bhatia et al., 2021). Shared tasks for Dravidian languages extend offensive-language detection to low-resource, code-mixed, and transliterated settings, including targeted-versus-untargeted label schemes (Chakravarthi et al., 2021). CHiPSAL 2025 further couples hate speech identification with target detection in Devanagari languages (Purbey et al., 2024), while Faux-Hate links hate speech to fake narratives and predicts target and severity in code-mixed Hindi-English text (Bhaskar et al., 18 Dec 2025).

Against that background, Dehate’s contribution is the shift from classifying whether hateful content is present to identifying where it is in an image and suppressing it through masking or blurring. This suggests a different moderation paradigm: one oriented toward localized visual intervention rather than solely toward categorical detection (Dalal et al., 26 Sep 2025).