We ran over 600 image generations to compare AI image models

Model aesthetics, behavior, and quirks

• OpenAI’s model is repeatedly described as instantly recognizable: strong yellow/orange/“nicotine” cast, “Ghibli-fied” look, and aggressive stylistic changes.
• Multiple commenters note it often alters faces (head shape, eyes, pose), even when asked not to, and corrupts fine details (Newton UI text/icons, background trees).
• Some see this as an architectural consequence of a unified token-based latent space: images are semantically re-encoded and regenerated, not edited at pixel level.
• Others argue it still “fails” many prompts (e.g., bokeh behavior, kaleidoscope symmetry, specific filters) or is too heavy‑handed for precise edits.
• Gemini/NanoBanana are seen as more conservative and photorealistic but often refuse to change images at all, especially with people; they may still claim success in the UI.
• Seedream is viewed as a capable middle ground: fewer outright failures than Gemini in some tasks, supports higher resolution, but tends to globally shift color balance and is uncensored.

Validity of comparisons and reliability concerns

• People disagree on how to judge success: strict prompt adherence vs. aesthetic quality vs. “not failing badly.”
• Some think the experiment’s prompt set is arbitrary and not very informative about reproducible success for others.
• There’s broad concern that “effect-only” edits (filters, bokeh, style transfer) often also change objects and faces, forcing tedious manual verification.

Local vs cloud models and tooling

• Several are disappointed local models weren’t included, but others note cloud unit economics are currently better for small products.
• Local generation on consumer GPUs is already fast enough for many, but tooling (Python scripts, fragmented UIs) is seen as chaotic; ComfyUI vs AUTOMATIC1111/Invoke is debated.
• DIY local setups with SDXL/Flux + LoRAs are said to outperform many SaaS models for niche or uncensored tasks, though generalization about models is hard given their diversity.

Impact on artists and creative work

• Views range from “illustrators/graphic designers largely redundant within decades” to “this is just another tool like photography or Photoshop.”
• Many predict: fewer low‑end illustration jobs, but more artists overall and higher productivity for those who integrate AI into workflows.
• Others argue AI images are mostly “junk food” aesthetics; they’ll dominate cheap mass content but not replace expressive, original art or invention of new styles.
• Stock photography is widely seen as a prime, legitimate casualty: custom AI visuals are already replacing magazine covers and thumbnails.