Diffusion Models Can Create New Images Without Infringing on Training Sources

A study of language models, especially those designed to create images and video, might suggest that copyright infringements by diffusion-based models might be impossible to prove. 

The authors say “locating a part of the training data that can be held responsible for a generated sample, can become impossible if a model is trained on a sufficiently large corpus of data.”

In other words, a large training set is all that is needed to render copyright traceable to any single source impossible. 

The study methodology relies on omitting an attributed unit from the training set (one image or all images by the same creator), the induced counterfactual sample cannot be directly attributed to that same unit. 

source: Nature Communications

In the above example, a generated image created from a smaller data set is more attributable to a single training image.

A large data set, on the other hand, creates a new image that shows no trace of a single training image.  

Model / family	Main generation	Diffusion?	Basic architecture / approach
GPT-4 / GPT-4o / GPT-5 family	Text, multimodal	No*	Primarily autoregressive Transformer for language
Claude 3/4 family	Text, code, multimodal	No*	Autoregressive Transformer
Gemini 1–4 family	Text, multimodal	No*	Primarily autoregressive Transformer
Llama family	Text, code	No	Autoregressive Transformer
Mistral / Mixtral family	Text, code	No	Autoregressive Transformer, including MoE
DeepSeek family	Text, code, reasoning	No	Autoregressive Transformer/MoE
Gemini Diffusion	Text/code	Yes	Text diffusion Transformer; generates blocks of tokens and iteratively refines them (Google DeepMind)
DiffusionGemma	Text/code	Yes	Non-sequential diffusion Transformer based on Gemma research (Google DeepMind)
Stable Diffusion 1.x/2.x/XL	Images	Yes	Latent diffusion, historically U-Net-based
Stable Diffusion 3/3.5	Images	Yes	Diffusion Transformer / MM-DiT
FLUX.1 / FLUX family	Images	Yes	Diffusion/flow-based Transformer
DALL·E 2	Images	Yes	Diffusion
DALL·E 3	Images	Yes	Diffusion-based image generation
Imagen 2/3/4	Images	Yes	Diffusion-based image generation; Google explicitly describes Imagen as a diffusion model (Google DeepMind)
Midjourney	Images	Yes*	Widely understood to use diffusion/related denoising techniques, though Midjourney does not disclose its architecture in detail
Adobe Firefly	Images/video	Yes*	Diffusion-based generation among its models; architecture varies by model
Sora	Video	Yes	Diffusion Transformer (DiT); starts from noisy video patches and denoises them (OpenAI)
Sora 2	Video	Yes*	Diffusion-based video generation; details of current implementation are less fully disclosed
Veo 2/3/3.1	Video	Yes*	Diffusion/Transformer-based video generation; Google describes Veo as drawing on its diffusion-model research (Google DeepMind)
Runway Gen-2/3	Video	Yes*	Diffusion-based video generation
Kling	Video	Yes*	Diffusion/Transformer-based video generation
Luma Dream Machine	Video	Yes*	Diffusion/Transformer-based video generation
MusicGen	Music/audio	No	Autoregressive Transformer over discrete audio tokens
AudioCraft / related models	Audio/music	Mixed	Different models use different architectures
Genie 2	Interactive worlds/video	Yes	Autoregressive latent diffusion—an interesting hybrid (Google DeepMind)

“Our findings provide a compelling case that attribution, the task of locating a unit of data within the training set that can be held responsible for a generated sample, is practically impossible on contemporary generative diffusion models,” say authors Zheng Dai and David Gifford.