TerraVis Annotations: licensing This dataset combines material that is licensed differently. Each part is listed below with the fields and rows it covers. The dataset is published by the TerraVis authors: Shuai Fu, Jing Gu, Jian Zhou, Zicheng Duan, Gengze Zhou and Qi Wu. 1. Human annotations and metadata Fields: image_id, benchmark, prompt_id, model, ratings, has_violation, rating_mean, has_violation_majority. Under the Amazon Mechanical Turk (AMT) Participation Agreement (last updated 25 March 2020, section 3(c)), rights in work product from AMT tasks vest in the Requester. No AMT Worker IDs, assignment or HIT IDs, or other worker information are included. Copyright (c) 2026 the TerraVis authors. To the extent copyright or database rights exist in these fields, they are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/legalcode 2. Generated images Field: image. The model field gives the model that generated each image. Every image in this dataset is AI-generated. The images are not photographs of real people or events. The licence for an image depends on the model that generated it: - gpt-image-1.5, gemini-3-pro-image-preview, qwen-image-2512 and stable-diffusion-3.5-large (3,600 images): CC BY 4.0, see section 2a. - flux.2-dev (900 images): the terms in section 2b. These are not a Creative Commons licence. The model providers' terms bind the dataset authors, including when they distribute the images. This licence passes on no conditions from those terms except the FLUX.2 [dev] restrictions in section 2b. 2a. Images licensed under CC BY 4.0 Rows: model = gpt-image-1.5, gemini-3-pro-image-preview, qwen-image-2512 or stable-diffusion-3.5-large (3,600 images). To the extent the dataset authors hold any rights in these images, they are licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/legalcode None of these four providers claims ownership of the images or requires a non-commercial licence for them. - gpt-image-1.5: created with OpenAI's gpt-image-1.5 model through the OpenAI Images API. OpenAI does not claim ownership of API outputs and sets no licence conditions for people who receive them. - gemini-3-pro-image-preview (Nano Banana Pro): generated through the Gemini API, including the Gemini Batch API, under the Google APIs Terms of Service (last modified 9 November 2021; https://developers.google.com/terms), the Gemini API Additional Terms of Service (versions effective 18 December 2025 and 23 March 2026; https://ai.google.dev/gemini-api/terms, earlier version at https://ai.google.dev/gemini-api/terms-archive/terms_12_18_25) and the Generative AI Prohibited Use Policy (last modified 17 December 2024; https://policies.google.com/terms/generative-ai/use-policy). Google does not claim ownership of generated content and may generate the same or similar content for others. Its terms leave the dataset authors responsible for how anyone they share the images with uses them. 437 of the Nano Banana Pro images are the original JPEG bytes returned by the API and carry Google's C2PA content credentials. The other 463 were decoded and re-saved as PNG by the generation code, which removed any C2PA metadata they carried. - qwen-image-2512: generated locally with the open-weight model Qwen/Qwen-Image-2512 from the Qwen team, Alibaba Cloud (https://huggingface.co/Qwen/Qwen-Image-2512). The model is released under the Apache License 2.0, which covers the model weights and code and places no conditions on images generated with it. - stable-diffusion-3.5-large: generated locally with stabilityai/stable-diffusion-3.5-large under the Stability AI Community License Agreement, last updated 5 July 2024 (https://huggingface.co/stabilityai/stable-diffusion-3.5-large/blob/main/LICENSE.md). Under its section IV(c)(iii), the dataset authors own these outputs as between themselves and Stability AI, to the extent permitted by applicable law. That licence does not treat model output as a Derivative Work, so its distribution and attribution terms do not apply to these images. Its section IV(b) makes the dataset authors' use of the outputs subject to the Stability AI Acceptable Use Policy (https://stability.ai/use-policy). 2b. flux.2-dev images Rows: model = flux.2-dev (900 images). These images are NOT licensed under CC BY 4.0, CC BY-NC 4.0 or any other Creative Commons licence. The dataset authors generated them locally with FLUX.2 [dev] (https://huggingface.co/black-forest-labs/FLUX.2-dev) under the FLUX Non-Commercial License: v2.0 of 25 November 2025 (https://bfl.ai/legal/non-commercial-license-terms), and v2.1, which the model repository has shown since 17 February 2026 (https://huggingface.co/black-forest-labs/FLUX.2-dev/blob/main/LICENSE.md). Both versions have the same terms for outputs. Black Forest Labs claims no ownership of outputs (section 2(d)). Under section 4(a), the dataset authors may not permit, assist or cause any third party to use, copy or distribute data produced by the model for the purposes listed below. The terms below pass that restriction on. To the extent the dataset authors hold any rights in these images, you may use, copy, modify and redistribute them, provided that you do not use, copy, modify or redistribute them, in whole or in part, or permit, assist or cause anyone else to do so: (a) for any commercial or production purpose. This includes any revenue-generating activity, use in direct interactions with end users or in ways that affect end users, and training, fine-tuning or distilling models for commercial use; (b) for military purposes; (c) for surveillance, including any research or development relating to surveillance; (d) for biometric processing; (e) in any way that infringes, misappropriates or otherwise violates, or is likely to infringe, misappropriate or otherwise violate, any third party's legal rights, including rights of publicity or "digital replica" rights; (f) in any unlawful, fraudulent, defamatory or abusive activity; (g) to generate unlawful content, including child sexual abuse material or non-consensual intimate images; or (h) in any way that violates applicable law, including privacy, security and biometric-information laws, the EU General Data Protection Regulation, the California Consumer Privacy Act and the EU Artificial Intelligence Act. In short: non-commercial, non-production use only, such as academic research and evaluation, and none of the uses in (b) to (h). If you redistribute any of these images, you must do so under this section 2b (or terms at least as restrictive), keep the statement that the images are AI-generated, and require your recipients to do the same. Rights to use FLUX.2 [dev] itself come only from Black Forest Labs under its own licence. This dataset grants none. 2c. Requests The dataset authors also ask the following of everyone who uses the images. These are requests, not conditions of this licence, except where section 2b already makes them conditions for the flux.2-dev images. - Please do not use the images to train image-generation models. OpenAI, Black Forest Labs and Stability AI restrict the dataset authors' own use of these outputs to develop competing or foundational generative models, and Google restricts using its services to develop competing models. Those restrictions bind only the dataset authors. Section 2b(a) separately forbids training models for commercial use with the flux.2-dev images. - Please do not present the images as made by a human or as real photographs, and do not use them in ways that the OpenAI Usage Policies, the Google Generative AI Prohibited Use Policy, the Stability AI Acceptable Use Policy or the Black Forest Labs Usage Policy (https://bfl.ai/legal/usage-policy) prohibit. - Please keep the C2PA content credentials in the Nano Banana Pro JPEGs, and do not try to remove invisible watermarks such as SynthID. 3. Prompts Field: prompt. Prompt text stays under its original licence, whatever licence applies to the rest of the dataset. No restriction or request in section 2 or in the dataset card applies to it. - COCO-T2I prompts are MS-COCO 2014 validation-set captions (annotations (c) COCO Consortium, CC BY 4.0, https://cocodataset.org/#termsofuse), as selected for the COCO-T2I subset of SeeTRUE (Yarom et al., "What You See is What You Read? Improving Text-Image Alignment Evaluation", NeurIPS 2023; https://huggingface.co/datasets/yonatanbitton/SeeTRUE; CC BY 4.0). CC BY 4.0 licence text: https://creativecommons.org/licenses/by/4.0/legalcode - GenAI-Bench prompts are from GenAI-Bench (Li et al., "GenAI-Bench: Evaluating and Improving Compositional Text-to-Visual Generation", 2024; https://huggingface.co/datasets/BaiqiL/GenAI-Bench), licensed under the Apache License 2.0. The full licence text is in licenses/Apache-2.0.txt. Modifications. The prompt field holds the prompt text given to each model. A possible "Generate an image: " prefix on 151 Gemini Batch API requests (see the dataset card) is not part of the field. The field differs from the upstream text in these cases (the file holding the rows is given in brackets): - COCO-T2I 006 [data/coco_t2i-00000-of-00002.parquet], all models: "a women that is walking down a sidwalk" became "a women that is walking down a sidewalk." - COCO-T2I 029 [data/coco_t2i-00000-of-00002.parquet], all models: "A dog and cat lying together on an orange couch." (two spaces after "lying") became "A dog and cat lying together on an orange couch." - COCO-T2I 053 [data/coco_t2i-00000-of-00002.parquet], all models: "A kitchen with a stainless steel stove.and white cabinets." became "A kitchen with a stainless steel stove and white cabinets." - COCO-T2I 149 [data/coco_t2i-00001-of-00002.parquet], gpt-image-1.5 and flux.2-dev only: "A refrigerator sitting in front of a vandalized wall." became "A refrigerator standing in front of a damaged wall." The other three models received the original text. - GenAI-Bench 0817 [data/genai_bench-00006-of-00012.parquet], every model except gemini-3-pro-image-preview: "A boy in a vivid orange hoodie and black basketball shorts kisses a girl in a sky-blue tutu and glittery flats." became "An adult man in a vivid orange hoodie and black basketball shorts kisses an adult woman in a sky-blue tutu and glittery flats." gemini-3-pro-image-preview received the original text. 4. No warranty; no endorsement The dataset is provided "as is", without warranties of any kind. Upstream material is provided "as is" by its authors, as set out in CC BY 4.0 section 5, Apache License 2.0 sections 7 and 8, and the COCO terms of use. OpenAI, Google, Black Forest Labs, Alibaba Cloud (Qwen team), Stability AI, Amazon (Amazon Mechanical Turk), the COCO Consortium, the SeeTRUE authors and the GenAI-Bench authors are named only to say where material came from. None of them is affiliated with or endorses this dataset.