Create and edit images with Gemini 2.0 in preview

Original Article ↗ Hacker News Discussion ↗

Perceived Image Quality & Official Examples

  • Thread sees mixed quality: some outputs “impressive,” others (e.g., polar bear mug, lamp-on-desk, table-with-missing-legs) are called embarrassingly bad for a launch blogpost.
  • Co-drawing/doodle demo is viewed as a fun tech demo but visually rough; some say it looks “vibe coded.”
  • Users report frequent failures on:
    • Precise edits (e.g., changing specific windows to bi-fold doors, modifying clothing in a photo)
    • Correct object placement and scale (lamp vs sofa, room decor, architectural proportions)
    • Understanding stick-figure sketches (inflating them into unintended 3D figures).
  • Some find compositing/editing weaker than OpenAI’s gpt-image-1, though others say Gemini preserves the original image better than GPT-4o when editing.

Speed vs Quality & Cost

  • Strong consensus that Gemini is very fast—often ~5 seconds vs 30+ seconds for OpenAI image models.
  • Several worry Google has over-optimized for speed, yielding “fast but junk” outputs that drive users back to Midjourney or others.
  • Pricing: about $0.039 per image, slightly above Imagen 3, with surprise bills when prompts trigger “many illustrations” and dozens of images in one response.

Prompting, Usability & Workflows

  • System is highly prompt-sensitive; small wording changes cause big quality swings.
  • Conversational interfaces expose limits of users’ ability to describe mental images. Many find it hard to specify clutter, lighting, composition, or technical effects.
  • Suggested strategies:
    • Feed reference images and ask Gemini (or another model) to describe them “in extreme detail,” then adapt that as a prompt.
    • Ramble your intent and have an LLM distill it into a precise prompt; iterate based on results.
    • Chain models: one to analyze texture/layout/typography, another to rewrite into richer visual instructions, then back to Gemini for generation.
  • Co-drawing’s usefulness is questioned if you must describe everything in text anyway.

Model Zoo, Availability & Comparisons

  • Users complain about confusing, fast-changing Gemini variants (Flash, Flash Image Gen March/May, 2.5 Pro/Flash/Live, “IO Edition”), and want a clear capability/price matrix.
  • Some benchmarking suggests Imagen 3 and OpenAI 4o still lead in aesthetic quality and prompt fidelity; Gemini’s main wins are multimodality and speed.
  • Gemini 2.0 Flash image models are unavailable in parts of Europe/EMEA despite earlier access, adding to confusion.

Wider Concerns & “AI Slop”

  • Google’s “product” examples are read as a push toward mass synthetic catalog images and marketing assets.
  • Commenters worry about deceptive e-commerce/real-estate imagery and a coming flood of low-effort “AI slop,” with doubts about long-term consumer tolerance.