Internet image generator archive
From melted MNIST digits to Nano Banana 2: a curated timeline of models, artifact-heavy outputs, and the most direct places to still try them.
Run links prioritize promptable browser demos. Some historic models are closed, sleeping, unofficial, or replaced by current products; each entry says what kind of link it is.
Digits, bedrooms, fake faces
The tiny digit-grid era: noisy strokes, melted numbers, visible mode collapse.
Goodfellow et al. made generator-versus-discriminator training the start of the modern image synthesis story.
Vibe: Blurry digits that look like a copier learning to dream.
Demo context: Community Space; small model, should wake quickly.
The dog-slug, eye-feather, recursive texture look that made AI weirdness visible.
Google turned classifier internals into public visual culture before text-to-image went mainstream.
Vibe: Over-recognized textures, paisley eyes, and animal parts everywhere.
Demo context: Community Streamlit demo; may need a short wake-up.
Almost a bedroom, if you squint.
Radford, Metz, and Chintala gave GANs a stable convolutional recipe and a recognizable generated-image texture.
Vibe: Symmetric smears, room blobs, glossy training-set averages.
Demo context: Community Space; current demo is not the original paper UI.
Sharper ImageNet GANs still produced cursed plush-object hybrids.
DeepMind showed that scaling GANs could produce much sharper and more diverse ImageNet samples.
Vibe: Hyperreal animals and objects with suspicious local anatomy.
Demo context: Community Space; class-based prompt/play controls vary.
Convincing until you looked at ears, jewelry, hair, and backgrounds.
NVIDIA made latent-space controls, face interpolation, and synthetic portraits central to AI imagery.
Vibe: Almost-photo faces with melting earrings and impossible backgrounds.
Demo context: One-click face refresh; not a prompt interface.
Sketches, cats, zebras
Edges-to-cats turned glitches into a drawing toy.
Conditional GANs became interactive: draw a map, mask, or contour, then get a plausible but unstable image.
Vibe: Cats and facades with local logic and global confusion.
Demo context: Original web demo still responds; old-school browser experience.
The failure grid is the point: cats half-become dogs, seasons drift, labels get confused.
Cycle consistency removed the need for paired examples and made domain transfer a core generative pattern.
Vibe: Texture transfer that sometimes forgets object boundaries.
Demo context: Community Space; may be asleep or slower than newer demos.
CLIP, DALL-E, haunted grids
The avocado chair, radish tutu, and prompt-composition era.
Internet-scale language-image training could compose weird prompts, not just sample a class label.
Vibe: Toylike objects, surreal literalism, charming prompt misunderstandings.
Demo context: Original DALL-E is closed; this links to a public homage from the same nostalgia wave.
The AI art Twitter notebook look: ornate, crunchy, symbolic, over-optimized.
Artists combined CLIP with generators and made promptcraft public before diffusion dominated.
Vibe: Maximal texture, repeated motifs, dream-logic composition.
Demo context: Public HF Space; old CLIP-guided workflows can take time.
Type anything, get nine haunted thumbnails.
It was not the strongest model, but it made text-to-image publicly playable and gave the web a shared broken-AI vocabulary.
Vibe: Blurry grids, scrambled faces, smeared objects, accidental horror lighting.
Demo context: Public Space; this is the nostalgic “weird grid” experience.
SD, DALL-E 2, FLUX
The astronaut-on-a-horse moment where photorealism became the headline.
It pushed text-to-image into a mainstream product race and made inpainting/outpainting feel usable.
Vibe: Clean surreal photography with hand, text, and physics failures around the edges.
Demo context: DALL-E 2 itself is retired from the main flow; ChatGPT is the current OpenAI route.
The open-weight explosion: DreamBooth faces, LoRAs, ControlNet, a thousand UIs.
Latent diffusion made quality image generation cheap enough to run locally and remix widely.
Vibe: Painterly sharpness, over-smooth skin, warped text, community model fingerprints.
Demo context: Community inpainting Space; representative SD 1.5 route, not the only one.
Prompt adherence and readable-ish text started displacing the old glitch aesthetic.
Long natural-language prompts worked better, and image generation moved into chat.
Vibe: Crisp editorial illustration, fewer weird edges, cleaner composition.
Demo context: Current product route; requires access to ChatGPT image generation.
Open models caught up on anatomy, typography, and prompt obedience.
Black Forest Labs reset expectations for open models after the first Stable Diffusion wave.
Vibe: Sharper, more commercial, less obviously old SD.
Demo context: Official HF Space; fast open model demo.
Chat, editing, Nano Banana
Image models became editing partners with better visual memory and text rendering.
OpenAI moved from a separate DALL-E lineage toward image generation inside a native multimodal model.
Vibe: Useful, polished, less nostalgic, harder to spot at a glance.
Demo context: Current product route; availability depends on account access.
Keep the person, change the scene, iterate fast.
Google turned subject-preserving edits into mainstream Gemini behavior with a memorable nickname.
Vibe: Fast literal edits with less identity drift than earlier tools.
Demo context: Current Gemini product route; model availability changes by region/account.
Fast edits plus Pro-style world knowledge and production specs.
Google positions it as Nano Banana Pro capability at Flash iteration speed: text, localization, consistency, and web-grounded knowledge.
Vibe: Controlled, product-ready imagery where weirdness is optional instead of built in.
Demo context: AI Studio preview route; rollout may not be available to every account yet.