OneIsFake
← Back to OneIsFake

Model generations & vendors

Last updated 26 August 2026

Every round in OneIsFake is drawn from a generator with a birthday. A 2023 image gives away far more than a 2026 one — so knowing the generations is half the skill. This is the map of what's out there.

How to read this page

Generators are grouped by vendor, then ordered oldest to newest. Each model carries the medium it produces and a difficulty band: how hard a typical output from it is to call today, not how advanced it was on release.

This is a reference to the publicly released families that matter for spotting synthetic media. It is not a claim that every model listed appears in every round, and it is not a complete inventory of our content pipeline.

Obvious2022 – 2023

Melted hands, gibberish signage, waxy skin. Most people can call these in under a second once they know what to look for.

Subtle2024 – 2025

Anatomy and text mostly fixed. The tells move to lighting logic, background physics and an over-composed, too-clean look.

Hard2025 – today

Frontier image and video models. Single-frame inspection often fails; context, provenance and cross-checking do the work.

Vendors and models

OpenAI

Vendor site

Moved from a dedicated diffusion image model (DALL·E) to image generation inside the multimodal chat models, plus Sora for video.

  • DALL·E 2ObviousImage2022
  • DALL·E 3ObviousImageOctober 2023

    Big jump in prompt following; still weak at embedded text and hands.

  • GPT-4o image generation / GPT-imageHardImageMarch 2025

    Native multimodal generation; marketed on accurate in-image text rendering.

  • SoraSubtleVideoDecember 2024
  • Sora 2HardVideoSeptember 2025

    Adds synchronised audio and markedly better physical plausibility.

  • GPT-4oSubtleTextMay 2024
  • GPT-4.1HardTextApril 2025
  • GPT-5 classHardTextAugust 2025 onwards

Google DeepMind

Vendor site

The widest spread of media types of any vendor — text, image, video and audio — with SynthID watermarking applied across outputs.

  • Imagen 2ObviousImage2023
  • Imagen 3SubtleImage2024
  • Imagen 4HardImage2025
  • Gemini 2.5 Flash Image (“Nano Banana”)HardImageAugust 2025

    Conversational image editing; very strong at keeping a subject consistent across edits.

  • Veo 2SubtleVideo2024
  • Veo 3HardVideo2025

    Native audio generation alongside video.

  • Gemini 1.5 / 2.0SubtleText2024
  • Gemini 2.5HardText2025
  • Gemini 3 classHardTextNovember 2025 onwards

Anthropic

Vendor site

Text only — Anthropic ships no public image or video generator, so Claude shows up in our fake-news and fake-review rounds rather than the picture rounds.

  • Claude 3 familySubtleTextMarch 2024
  • Claude 3.5 / 3.7SubtleText2024 – 2025
  • Claude 4 / 4.5 classHardText2025 onwards

    Latest public flagship at time of writing is the Opus 4.5 generation.

DeepSeek

Vendor site

Open-weight Chinese lab whose reasoning models made long, confident, well-structured text cheap to produce at scale.

  • DeepSeek-V3SubtleTextDecember 2024
  • DeepSeek-R1SubtleTextJanuary 2025

    Reasoning model; early versions leaked visible chain-of-thought phrasing into answers.

  • DeepSeek-V3.2 and laterHardText2025 – 2026

Moonshot AI (Kimi)

Vendor site

Open-weight long-context models. Strong at bulk content, which is why they turn up in synthetic article and comment sets.

  • Kimi K1.5SubtleTextJanuary 2025
  • Kimi K2HardTextJuly 2025

    Open-weight release; a common backend for high-volume generated text.

  • Later K2.x / K3 releasesHardText2026

    Version numbering moves fast and is not always announced in English — treat specific labels loosely.

Black Forest Labs (FLUX)

Vendor site

Founded by former Stable Diffusion researchers. FLUX is the open-weight image model you are most likely to meet in the wild today.

  • FLUX.1 [schnell] / [dev] / [pro]SubtleImageAugust 2024

    schnell is the fast, few-step variant and tends to look flatter than pro.

  • FLUX.2HardImageNovember 2025

    Multi-reference consistency and much better in-image text.

Midjourney

Vendor site

The most stylistically opinionated generator. Its house look is itself a tell — and the reason so many people think all AI images look like that.

  • v5ObviousImage2023
  • v6 / v6.1SubtleImage2024
  • v7 (and niji variants)HardImageApril 2025

    Explicitly improved hand and body coherence over v6.

Stability AI

Vendor site

The original open-weight lineage. Old checkpoints are still everywhere, which is why 2022-era artifacts have not disappeared from the internet.

  • Stable Diffusion 1.5ObviousImage2022

    Still widely run locally and in hobby tooling.

  • SDXLObviousImage2023
  • Stable Diffusion 3SubtleImage2024
  • Stable Diffusion 3.5 (Large / Turbo / Medium)SubtleImageOctober 2024

Caveats worth stating plainly

  • Release dates are approximate: vendors stage previews, limited access and general availability weeks or months apart.
  • Version numbering moves faster than documentation, particularly for open-weight labs. Treat specific point releases as indicative.
  • Open weights mean an old model never really retires. Stable Diffusion 1.5 output is still being produced today, which is why the “obvious” era still shows up in circulation.

Next: what each one gets wrong

Knowing the lineup is step one. Step two is knowing the artifacts each generation leaves behind — see how to spot AI.