blog & research · model guides
AI image & video
model detection
⤹ who made this thing?when amige. flags something as “looks like Midjourney v6” or “looks like GPT-4o,” these are the pages that explain what that means. version histories, visible tells in the output, and how amige. reads the marks each model leaves. a trained model routes each scan to the detectors strongest for it, a panel of independent detectors built by different teams weighs in, and the reads fuse into a calibrated best guess. the whole machine lives at the machine.
- Midjourney →imagethe category leader for stylized image generation. moody lighting, painterly textures, the famous 'Midjourney face.' V7 and V8 are meaningfully harder to detect than V5 was.
- ChatGPT (GPT-4o, GPT-5) →textthe highest-traffic AI product in history. flat burstiness, heavy bullet structure, and the documented post-launch vocabulary spike are the practical tells.
- Sora →videoOpenAI's text-to-video. Sora 2 added audio and better physics. short clips, morphing background objects, physics anomalies are the residual tells.
- Claude →textAnthropic's frontier text family. heavy em-dash use, characteristic hedges, structured bullets. Opus 4.8 is the current flagship.
- Gemini →textGoogle DeepMind's frontier model. embedded everywhere via Search AI Mode, Workspace, Pixel. SynthID by default.
- DALL-E →imageOpenAI's pre-GPT-Image image gen, deprecated May 2026. still in the wild on social media. DALL-E 2's color-stripe watermark is a giveaway when present.
- Stable Diffusion →imagethe open-weights family + Civitai ecosystem. SDXL through 3.5. the hardest detection case because of the fine-tune long tail.
- Flux →imageBlack Forest Labs' flow-matching model. the strongest open-weights challenger to Midjourney. high prompt adherence, hard to spot by eye.
- Nano Banana →imageGoogle's image gen inside Gemini. SynthID by default. when the watermark survives, the maker call gets a lot easier.
- DreamID →imageByteDance + Tsinghua face-swap model. a DreamID flag means deepfake rather than generic AI-generated content.
- Imagen →imageGoogle DeepMind's diffusion image family, distinct from Nano Banana. clean high-key 'Google product photography' aesthetic. SynthID baked in.
- Veo →videoGoogle's text-to-video. Veo 3 added synchronized native audio. SynthID on both video frames and the audio track, which makes verification unusually durable.
- Runway →videothe indie / creator-favorite video gen. Gen-1 through Gen-4.5. most embedded in professional film and TV pipelines. no invisible watermark on consumer tier.
- Luma (Dream Machine) →videothe free-tier consumer video gen. 25M+ users reported during 2025. smooth motion, soft textures. the 'I tried it for fun' tool.
- Pika →videothe TikTok-native social video gen. Pikaffects (Cake-ify, Crumble, Explode) are the signature. stylized over photoreal, on purpose.
- Ideogram →imageowned the in-image text rendering niche from 2023 until Nano Banana caught up. design-y compositions, crisp typography. marketing / SMB workflows.
- Adobe Firefly →imageAdobe's enterprise-grade 'commercially safe' image (and video) gen. trained on licensed Adobe Stock content. C2PA signed on every output. ~24B assets generated.
- GPT-5 →textis this GPT-5? OpenAI’s default model writes flatter and more human than GPT-4o, which is exactly why it’s harder to flag. the tells that still show, and why every verdict stays a probability.
- Grok Imagine →videoxAI’s image-and-video gen inside the Grok app. short clips with native audio, waxy skin on stills, image-to-video artifacts on clips. best known for the Spicy-mode deepfake scandal, and it ships no robust watermark, so detection leans on the pixels and the context.
DALL-E, Stable Diffusion, Claude, Gemini, Flux, Nano Banana, DreamID, pages coming in priority order by search demand. if a model you care about is missing, scan something and share the permalink. data routes scope.