Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How AI image and video generators work

    Many AI image generators are diffusion models: they build a picture through a series of denoising steps, often in a compressed "latent" space, guided by a text prompt.[1][2] Newer designs swap in transformers that work on patches of the compressed image, and video models such as Meta's Movie Gen use large transformers over video tokens.[3][4] Provenance tools such as C2PA Content Credentials and Google's SynthID watermark, and the EU AI Act's marking rules, aim to make the output identifiable as AI-made.[5][6][7]

    Editor reviewedUpdated Explainer16 claims · 100% primaryArtificial intelligence, Frontier AI, Tech policy
    Reading level: 3 passages differ; the rest is the same for everyone.

    This explainer covers how image and video generators are built, where their training data comes from, and the tools and rules for identifying their output. It does not compare or review individual products. Choose a reading level in each section.

    How diffusion models make images

    Many image generators are diffusion models, which build a picture through a long series of small clean-up, or “denoising”, steps.[1] A 2020 paper showed the approach could produce high-quality pictures.[8] To save computing power, latent diffusion systems do this work on a compressed version of the image, and a text prompt steers the process.[2]

    Diffusion models decompose image formation into a sequential application of denoising autoencoders; Ho, Jain and Abbeel’s 2020 denoising diffusion probabilistic models paper demonstrated high-quality synthesis with this family.[8][1] Training directly in pixel space often cost hundreds of GPU days, and sampling is expensive because it needs many sequential evaluations.[1] Latent diffusion moved the process into the latent space of a pretrained autoencoder and added cross-attention layers so that text or other inputs can condition generation.[2]

    From images to video

    Researchers swapped the image-cleaning network for a transformer that works on small patches of the compressed image, and found that bigger transformers made better images.[3] Meta’s Movie Gen applies a transformer to video. Its largest version has 30 billion parameters and makes 16-second 1080p clips with matching sound.[4]

    Diffusion Transformers (DiTs) replace the U-Net backbone of a latent diffusion model with a transformer over latent patches, and image quality (FID) improved consistently with more compute per forward pass.[3] Meta’s Movie Gen paper describes a 30B-parameter transformer trained with a maximum context of 73,000 video tokens, equal to 16 seconds at 16 frames per second, generating 1080p video with synchronised audio.[4]

    Training data

    Image generators learn from very large collections of pictures paired with captions: LAION-5B, an openly released dataset used to train such models, contains 5.85 billion CLIP-filtered image-text pairs, 2.32 billion of them in English.[9] After the Stanford Internet Observatory reported in December 2023 that LAION-5B contained links to suspected child sexual abuse material, LAION withdrew the dataset. In August 2024 it released Re-LAION-5B with 2,236 such links removed.[10]

    Who owns the rights to training images is a separate question, followed in the AI copyright lawsuits tracker. Whether a model’s weights and training details are public is covered in the open-weight vs closed models explainer.

    Provenance and watermarks

    The C2PA standard, called Content Credentials, attaches a signed record of who made a file and how it was edited.[5][11] That record can be stripped off, and adding it is optional.[12] Google’s SynthID instead hides an invisible watermark inside AI-generated images, video, audio and text.[6]

    A C2PA Content Credential (manifest) is a cryptographically bound structure of assertions about an asset’s origin and modifications.[11] The C2PA explainer acknowledges that provenance metadata can be removed; “durable” credentials pair a hard binding (cryptographic hashing) with a soft binding such as a watermark or fingerprint so the manifest can be recovered.[12] Google says SynthID image and video watermarks are designed to survive cropping, filters, frame-rate changes and lossy compression.[6] Its SynthID Detector portal, launched in 2025, checks content from Google and partners including OpenAI, NVIDIA and Kakao.[13]

    Labelling rules

    The EU AI Act sets two duties in Article 50. Providers of systems that generate synthetic audio, images, video or text must mark the output in a machine-readable format so it is detectable as AI-generated.[7] Deployers that publish deepfake images, audio or video must disclose that the content is artificial, with lighter rules for evidently artistic, satirical or fictional work.[14] The Act’s recitals list watermarks, metadata, cryptographic provenance methods, logging and fingerprints as possible techniques.[15] These transparency rules took effect in August 2026.[16]

    Next in the crash course

    More in AI

    Questions readers ask

    What is a diffusion model?

    A diffusion model generates an image through a sequence of denoising steps. A 2020 paper by Ho, Jain and Abbeel showed that this approach could produce high-quality images.[1][8]

    How do AI video generators differ from image generators?

    Some use large transformers that work on compressed video tokens. Meta's Movie Gen, for example, is a 30-billion-parameter transformer that generates 16-second, 1080p clips with synchronised audio.[3][4]

    What data are image generators trained on?

    Large collections of images paired with text. LAION-5B, an openly released dataset, contains 5.85 billion image-text pairs; in 2024 LAION re-released it with 2,236 links to suspected child abuse material removed.[9][10]

    Can you tell whether an image was made by AI?

    Sometimes. C2PA Content Credentials record an image's origin and edits, but the metadata can be removed and adding it is optional. Google's SynthID embeds invisible watermarks that its SynthID Detector can check.[11][12][6][13]

    Does the EU require AI images to be labelled?

    Yes. Article 50 of the AI Act requires providers to mark synthetic output in a machine-readable way and requires deployers to disclose deepfakes. The transparency rules took effect in August 2026.[7][14][16]

    Sources

    · 16 claims from 6 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      Diffusion models build an image through a sequence of denoising steps, and working directly on pixels made training powerful models cost hundreds of GPU days. confirmedas of 2021-12-20

    2. [2]

      Latent diffusion runs the diffusion process in the compressed latent space of a pretrained autoencoder and uses cross-attention layers to condition generation on inputs such as text. confirmedas of 2021-12-20

    3. [3]

      Diffusion Transformers (DiTs) replace the U-Net backbone of latent diffusion models with a transformer that operates on latent patches, and their image quality improved as they were scaled up. confirmedas of 2022-12-19

    4. [4]

      Meta's Movie Gen paper (October 2024) describes a 30-billion-parameter transformer that generates 1080p video with synchronised audio, with a context of 73,000 video tokens, equal to 16 seconds at 16 frames per second. confirmedas of 2024-10-17

    5. [5]

      The Coalition for Content Provenance and Authenticity (C2PA) publishes an open technical standard, called Content Credentials, for establishing the origin and edits of digital content. confirmedas of 2026-11-07

    6. [6]

      Google DeepMind's SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video, and Google says image and video watermarks are designed to survive cropping, filters and lossy compression. confirmedas of 2026-11-07

      • SynthID · Google DeepMind · Primary source · How SynthID works (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · AI-generated image and video (retrieved 2026-11-07)
    7. [7]

      Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text to mark the outputs in a machine-readable format so they are detectable as artificially generated or manipulated. confirmedas of 2024-07-12

    8. [8]

      A June 2020 paper by Ho, Jain and Abbeel showed high-quality image synthesis using denoising diffusion probabilistic models. confirmedas of 2020-06-19

    9. [9]

      LAION-5B, an openly released dataset used to train image generators, contains 5.85 billion CLIP-filtered image-text pairs, 2.32 billion of them in English. confirmedas of 2022-10-16

    10. [10]

      After a December 2023 Stanford Internet Observatory report found links to suspected child sexual abuse material in LAION-5B, LAION withdrew the dataset and in August 2024 released Re-LAION-5B with 2,236 such links removed. confirmedas of 2024-08-30

    11. [11]

      A C2PA Content Credential is a cryptographically bound record of an asset's provenance containing assertions such as when and where it was created and how it was modified. confirmedas of 2026-11-07

    12. [12]

      The C2PA explainer says provenance metadata can be removed, that "durable" credentials add watermarking or fingerprinting to recover it, and that adding provenance is optional. confirmedas of 2026-11-07

    13. [13]

      Google launched the SynthID Detector verification portal in 2025 and says it is expanding access to everyone, to check content made with AI from Google or partners including OpenAI, NVIDIA and Kakao. confirmedas of 2026-11-07

      • SynthID · Google DeepMind · Primary source · SynthID Detector (retrieved 2026-11-07)
      • SynthID · Google DeepMind · Primary source · SynthID Detector (retrieved 2026-11-07)
    14. [14]

      Article 50(4) requires deployers of AI systems that generate or manipulate deep-fake image, audio or video content to disclose that it is artificially generated or manipulated, with lighter rules for evidently artistic, satirical or fictional works. confirmedas of 2024-07-12

    15. [15]

      The AI Act's recitals list watermarks, metadata, cryptographic provenance methods, logging and fingerprints as possible techniques for marking AI-generated content. confirmedas of 2024-07-12

    16. [16]

      The AI Act's transparency rules took effect in August 2026, when the Act became generally applicable on 2 August 2026. confirmedas of 2026-08-02

    Revision history (1)
    1. Page created.

    Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Mar 7, 2027.

    Cite this page

    "How AI image and video generators work." ContentLora, updated Nov 7, 2026. https://contentlora.com/explain/how-ai-image-and-video-generators-work

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.