ChatGPT Images 2.0

Original Article ↗ Hacker News Discussion ↗

Capabilities & Quality

  • Many commenters find GPT‑Image‑2 a clear step up: highly photorealistic, strong layout and typography (fake desktop, magazines, posters, UI mockups), good text rendering in Latin and even Chinese (though with occasional typos).
  • For diagrams, slides, and product/UI mockups, people report it as “finally useful” and often better than prior OpenAI models and Gemini/Nano Banana in fidelity.
  • It can maintain characters across panels (manga, comics) better than older models, but still fails on fine logical details (fractional pizza slices, precise chess positions, QR codes, barcodes, exact color bands on snakes, etc.).
  • “Where’s Waldo”–style crowd scenes impress at a glance but collapse under zoom: distorted faces, missing limbs, odd artifacts.
  • It passes some classic tests (piano keyboard, nine‑pointed star) that previously broke models, but still fails other “model killer” prompts and geometric tasks (cubes, grids, circles).

Pricing & Technical Details

  • API model card and pricing show slightly cheaper per‑pixel cost vs GPT‑Image‑1.5 overall, but confusing per‑size prices led to speculation about typos.
  • Resolutions are more flexible now (reported up to 3840×2160 within a pixel budget).
  • Some confusion over transparent PNG support: UI can do it; API status is unclear.
  • Compared with Gemini’s image model, several commenters note GPT‑Image‑2 is more expensive per high‑res image but arguably higher quality.

Editing & Workflows

  • Editing existing images remains a pain point: common complaints about over‑tuned “tone mapping,” loss of sharpness, or large unintended changes from local edit prompts.
  • Some users chain models with traditional tools (layers, masks, inpainting) to get reliable hyper‑localized edits.
  • Sprite sheets and animations for games are still weak; consistency across frames is hard.

Watermarking, Provenance & Detection

  • System card mentions imperceptible watermarking; OpenAI also embeds C2PA manifests.
  • Commenters note research showing such watermarks can be stripped via regeneration, though they survive casual transforms like compression, crops, screenshots.
  • Debate over camera‑side cryptographic signatures: some say this is the right direction; others argue you can always photograph an AI image or spoof sensors.
  • Concern that social platforms strip metadata, undermining provenance schemes unless regulations force them to preserve it.

Use Cases & “Democratization”

  • Positive use cases cited:
    • Business assets: posters, menus, packaging, manuals, tickets, logos, websites, pitch decks, UI mockups.
    • Education: kid readers and coloring books with favorite characters, personalized learning materials, diagrams and maps.
    • Personal design: gardens, rooms, balconies, front yards.
    • Indie projects: album covers, band posters, game assets (where acceptable).
  • Many see this as “democratizing visual communication,” letting non‑artists prototype and communicate ideas visually.
  • Others push back that basic diagramming was already widely accessible and that “democratization” is mostly about undercutting professionals.

Risks, Ethics & IP

  • Strong anxiety about deepfakes and erosion of visual evidence:
    • Fake political content drowning real scandals.
    • Harassment (non‑consensual nudes, virtual kidnappings) and targeted disinformation at scale.
    • Legal ramifications where photos and CCTV become weak evidence.
  • Repeated complaints about training on copyrighted art, photos, and OSS code without consent or compensation; analogies to “mass IP theft” and exploitation of open culture.
  • Some argue the harms (misinformation, job loss, commodification of art, destruction of “truth”) outweigh mostly decorative benefits; others see it as another historical wave of automation and creative tooling.
  • Debate on regulation vs “socializing” AI ownership; skepticism that copyright law will protect small creators.

Environmental & Economic Concerns

  • Worries about power and especially water usage of GPU data centers; counter‑claims that water use is overblown relative to other sectors.
  • Jevons‑paradox style arguments: cheaper, faster image gen leads to far more total images, so environmental cost may still grow.
  • Some say AI art mainly shifts value from many working artists to a few AI platform owners and their investors.

Community Reception & Comparisons

  • Thread is sharply polarized: some “blown away” and already integrating it into real workflows; others repulsed by “AI slop” and intentionally avoid AI‑generated visuals.
  • Comparisons:
    • Many still rate Midjourney as best for “taste” and style, but GPT‑Image‑2 as superior in prompt adherence, text, UI layouts, and diagrams.
    • Nano Banana/Gemini often wins on some visual fidelity benchmarks but lags on logic‑heavy prompts.
  • Some note a persistent “GPT look” (slight sepia/nostalgia filter), though others feel the generic slop aesthetic has diminished.