In the summer of 2022, latent-diffusion image generators were released for general use and the technology to produce photographic-grade images from text prompts alone reached anyone who wanted it. Within months, the production, distribution, and regulation of sexual imagery had been remade.

Overview

AI-generated erotica (Japanese: AI 生成エロ, AI seisei ero; English: AI-generated pornography) is the umbrella term for sexual representations produced by deep-learning image- and video-generation models. The article covers post-2022 diffusion-model image generators (Stable Diffusion, NovelAI Diffusion, Midjourney and others) and their adult-content use: the technical history, additional-training techniques for likeness reproduction, the collisions with copyright, image rights, and child-protection law, and the international regulatory landscape.

The production-side description: (1) a diffusion model pre-trained on a large image-and-text dataset functions as the base; (2) the user provides a natural-language prompt or a reference image; (3) the model iteratively denoises a random tensor into a finished image; (4) the resulting image, if so directed, contains sexual subject matter.

The technical lineage runs through three stages. The first is the generative-adversarial network (GAN) line from 2014. The second is the diffusion-model line (denoising-diffusion probabilistic models, DDPM) of around 2020. The third is the latent-diffusion-model line (LDM) of 2022, in which the working representation is a low-dimensional latent space that reduces compute cost enough for consumer GPUs to run inference. Robin Rombach et al.’s 2022 CVPR paper High-Resolution Image Synthesis with Latent Diffusion Models is the standard reference.

In adult-content use, base models are combined with additional-training and conditioning techniques — LoRA (Low-Rank Adaptation), DreamBooth, Textual Inversion, ControlNet, and others — to reproduce a particular style, character, composition, or pose intentionally. These techniques were published as scientific work; their carry-over into sexual-content generation has produced new questions of copyright, image rights, and child protection.

Etymology

AI seisei (AI 生成) is the Japanese translation of AI-generated, used for outputs of machine-learning models. AI poruno renders the English AI-generated pornography / AI porn / synthetic pornography. The Japanese-language usage stabilised through late 2022 and 2023, in step with the general release of Stable Diffusion.

The AI in this context is not classical symbolic-AI: it is a deep-neural-network statistical-generative model. For strict accuracy, image-generation AI, generative AI, or diffusion model are sometimes used.

Technical history

Pre-history (2014–2021): GAN and StyleGAN

Ian Goodfellow et al.’s 2014 generative-adversarial-network paper introduced the generator-vs-discriminator adversarial training pattern that produces new images close to a training distribution. NVIDIA’s StyleGAN and StyleGAN2 lines from 2018 onward reached photographic quality in face generation, and their adult-side applications included real-person face replacement (deepfake) and fictional-face generation services such as This Person Does Not Exist.

This generation of technology required substantial same-domain training data per model, did not follow arbitrary text prompts, and gave poor pose / composition control. Sexual-image generation in this generation was correspondingly limited.

2022: Stable Diffusion 1.x and the boom

In August 2022, a coalition led by CompVis, Stability AI, and Runway released Stable Diffusion 1.4 (1.5 in October). The model was trained on LAION-5B (5.85 billion image-text pairs), distributed under an open licence (CreativeML Open RAIL-M), and ran on a consumer GPU with around 8 GB of VRAM — the first high-quality text-to-image model to combine the three properties.

From the release date forward, derivative work for adult content proliferated on 4chan, Reddit, Twitter, and dedicated communities. Civitai and similar model-share platforms collected tens of thousands of derivative models and LoRA. DLsite and FANZA began receiving AI-generated CG sets.

In October 2022, Anlatan released the illustration-specialised NovelAI Diffusion as a paid service. The model was trained on Danbooru-style illustration datasets and surpassed contemporary alternatives in two-dimensional bishōjo-illustration quality by a substantial margin. The weights leaked in October of the same year, and derivative models built on the leaked weights (the NAI line) became one of the standard substrates for generative-adult-content production.

2023–2024: SDXL, Pony Diffusion, the LoRA ecosystem

Stability AI’s SDXL (Stable Diffusion XL), released in July 2023, brought 1024×1024 generation, improved complex-prompt adherence, and substantially better hand rendering. Through 2024, SDXL-derived NSFW-specialised models — Pony Diffusion XL, AnimagineXL, AutismMix and others — pushed two-dimensional adult image generation further.

LoRA is the additional-training technique that adds a low-rank matrix delta (typically tens of megabytes) to a base model rather than rewriting its hundreds of millions of parameters, allowing the reproduction of a particular character, illustrator, or real person from a few dozen reference images and a few hours of training on a consumer GPU. LoRA files for specific anime and game characters, specific illustrators’ styles, and specific real persons now circulate in very large numbers.

ControlNet (2023) adds pose, line-drawing, depth, and similar conditioning inputs, enabling the explicit specification of composition, pose, and body placement in generated images.

2024 onward: video generation

In 2024, Stable Video Diffusion, OpenAI Sora, Runway Gen-3, Kling and other text-to-video models were released in succession. Their adult-content carry-over has produced (1) style-transfer derivatives of existing video material; (2) wholly synthetic AI video; (3) still-to-short-video derivatives. Video remains substantially more compute-intensive than still images, and quality is less consistent, but the rate of advance is rapid.

Principal models

Model	Origin	Year	Notes
Stable Diffusion 1.5	Runway / Stability AI	2022	The general-purpose entry point; 512×512, CreativeML Open RAIL-M.
NovelAI Diffusion	Anlatan	2022	Illustration specialised; weight leak drove derivative ecosystem.
SDXL 1.0	Stability AI	2023	1024×1024; improved prompt adherence.
Pony Diffusion XL	PurpleSmart.ai	2023	NSFW-tag trained; de facto standard for two-dimensional adult generation.
Midjourney	Midjourney, Inc.	2022	Closed commercial; NSFW prohibited by policy.
Stable Video Diffusion	Stability AI	2023	Still-to-short-video generation.
Sora	OpenAI	2024	High-quality text-to-video; commercial use under review.

The closed commercial models (Midjourney, OpenAI’s DALL-E, Adobe Firefly) prohibit sexual content by policy and implement prompt-stage and output-stage filtering. The open Stable Diffusion line distributes the base model neutrally, with sexual-image generation delivered through derivative models and user-side filter disabling.

Ethical and legal questions

Copyright in the training data

The image-text pairs in LAION-5B and similar datasets were collected from the open web without explicit rights clearance from the underlying rightsholders. Beginning in 2023, suits in the United States (Getty Images vs Stability AI and others) have raised the copyright question of the training stage.

Japanese copyright-law article 30-no-4 generally permits the use of copyrighted works for information-analysis purposes, but the enjoyment-purpose overlap clause and the not unduly harming the rightsholder’s interests limitation set the practical edges of the permission. The Agency for Cultural Affairs Council of Cultural Affairs Copyright Subdivision published Considerations on AI and Copyright in March 2024, structuring the discussion of training and generation stages.

In the adult-content context covered here, training data likely include unlicensed adult manga, doujinshi, and commercial art collections; the dependency-and-similarity analysis of output works against original works is being argued case by case.

Real-person likeness and image rights

LoRA- and DreamBooth-trained reproduction of real-person likenesses is liable on image-rights, publicity-rights, and defamation grounds, both at generation and at publication. Generating sexual imagery of real entertainers, streamers, or non-celebrity individuals without their consent, and posting it to social-media or streaming platforms, falls within the scope of revenge-porn law, defamation, and insult-offence provisions.

The United States enacted the Take It Down Act at the federal level in 2025, criminalising the posting and circulation of non-consensual sexual images including AI-generated material and imposing a 48-hour take-down obligation on platforms. The EU Digital Services Act and AI Act operate in the same direction.

AI-generated sexual imagery of real persons in combination with deepfake techniques is internationally recognised as a particularly serious human-rights violation. The editorial position of this article is that no AI-generated sexual imagery of any real person should be produced or distributed without that person’s consent; the practice is unlawful and ethically impermissible.

Child protection and CSAM

The most serious question raised by AI-generated erotica is child sexual abuse material (CSAM). In December 2023, David Thiel of the Stanford Internet Observatory reported that the LAION-5B dataset used to train Stable Diffusion had contained at least 1,008 real-CSAM images, with substantial impact on the research community and regulators. LAION temporarily withdrew the dataset and published a CSAM-removed reprocessing as Re-LAION-5B in 2024.

Japan’s Child Pornography Act (1999, amended 2014) is currently interpreted as covering only depictions of real children; wholly drawn fictional minor depictions are not within the criminal statute. The interpretive boundaries with AI-generated material — what happens when training data contained real CSAM; what happens when LoRA is trained on photographs of real children; what happens when an existing child’s face is composited onto a different body — have not yet settled.

The United Kingdom, the United States, Canada, Australia and others criminalise drawings, illustrations, and AI-generated images depicting children sexually, regardless of whether the depicted child is real or fictional, and treat AI-generated material on the same footing as other depictions. Japanese legislative discussion of an equivalent extension is in progress as of 2025.

The editorial position of this article is that any AI-generated image depicting a child sexually is treated as falling within the regulatory and ethical bar, regardless of whether the depicted subject is held to be real or fictional, and is not introduced positively in this work.

Copyright in the output

Whether an AI-generated image qualifies as a work of authorship under copyright law is being argued differently in different jurisdictions. The United States Copyright Office has held since 2023 that pure AI-generated material without identifiable human creative contribution is not protectable. Japan’s Cultural Affairs Agency similarly takes the position in its 2024 statement that AI-output authorship is decided case by case according to the degree of human creative input.

In practice, AI-generated CG sets distributed commercially carry real difficulties: enforcement against third-party copying is hard; platform-side responsibility allocation is unclear; and the line with derivative-work (fan-art) framing is fuzzy.

Jurisdictions and platforms

Regulation

The EU’s AI Act (2024) imposes risk-tier regulation on AI systems and a transparency obligation (disclosure of AI generation) for generative-AI outputs. The act phases into force from 2026.

The United States has no comprehensive federal AI statute, but in addition to the 2025 Take It Down Act, state-level statutes — Tennessee’s 2024 ELVIS Act (against unauthorised AI voice and likeness generation), California’s AB 2602 (consent requirements for AI replication of actors) — lead the field.

China’s Provisional Administrative Measures for Generative Artificial Intelligence Services (2023) imposes registration, content-control, and AI-output-transparency obligations on generative-AI service providers. Sexual-content generation is broadly prohibited under that statute and the related Provisions on the Ecological Governance of Network Information Content.

Platforms

DLsite introduced an AI-generated tag for the explicit labelling of AI-generated works in 2023, with filter and exclusion-search support. Fantia, pixiv FANBOX, and Civitai followed similar policies. Pixiv has progressively strengthened the search separation and tag labelling of AI-generated work since 2022.

The closed commercial models (Midjourney, OpenAI DALL-E, Adobe Firefly) prohibit NSFW output in their terms of service and enforce filtering. In the open Stable Diffusion line, the base model is distributed neutrally and adult generation is delivered through derivative models and user-side disabling of filters; the practical landscape is as already described.

Cultural and social impact

The spread of AI-generated erotica is producing a multi-sided shift: (1) competition with hand-drawn artists in the doujin CG-set market; (2) the re-definition of original-work and character rights inside fan-art production; (3) the diffusion of AI-assistive techniques into the eromanga production pipeline; (4) a re-balancing of freedom-of-expression against child-protection and human-rights protection.

The hand-drawing community has voiced (1) objection to the use of their work in training data without consent; (2) concern at the downward pressure on prices that mixed-supply markets exert; (3) concern at the duplication of artists’ personal styles through LoRA. Other artists have adopted AI tooling as an assistive component, and a categorical pronouncement either way is difficult at the current stage.

In expression-regulation discourse, AI-generated erotica raises the foundational question of whether the creator is a human or a machine; raises the question of the regulator’s response when the real / fictional boundary is technically blurred; and raises the question of the effectiveness of jurisdictional regulation against models distributed across borders.

Related terms

• Deepfake
• Revenge-porn (legal)
• Eromanga
• DLsite
• FANZA