Images 2.6

OpenAI’s new “Images 2.6” model for image generation, which adds native GIF and SVG output and claims faster performance, is drawing mixed reactions from users comparing it to earlier GPT-Image 2.x releases and competitors like Google’s “Nano Banana.” Many find prompt adherence and features like vector export and simple animation promising, but complain about visible watermark artifacts, an over-processed “AI look,” and, in some workflows, a quality regression versus 2.5 despite higher effective costs due to increased token usage. The conversation broadens into concerns that image generation progress has plateaued, that commercial incentives favor speed and cheapness over visual nuance, and that widespread “AI slop” is eroding both aesthetic standards and the perceived value of human-created art.

Overall reaction

  • Many see Images 2.6 as incremental rather than transformative; “same as the last 1.5 years” comes up.
  • Mixed response: some are impressed by specific prompt adherence and complex compositions; others see regressions vs earlier OpenAI models or competitors (e.g., Google’s “Nano Banana”, Seedream, Midjourney).
  • Several note that image-gen progress feels to have plateaued compared to coding models, or at least is improving more slowly.

Scale and usage

  • OpenAI’s “4 billion images per week” claim sparks skepticism and unease; commenters estimate this implies huge automated or bulk usage that far exceeds human viewing capacity.
  • Individual users report generating many variants per idea, reinforcing that most generations are exploratory “slop” rather than final assets.

Pricing and token usage

  • Official per-token rates are unchanged, but multiple users show that 2.6 uses ~3–4× more output tokens than 2.5 at low/medium quality (e.g., ~196 → ~823 tokens for 1024×1024 low).
  • This is seen as a de facto price hike, with some calling the marketing (“same price”) deceptive.
  • High-quality settings reportedly have similar token usage to 2.5.

New features: SVG and GIF

  • SVG output is implemented via a raster-to-vector pipeline; requires specifying an additional svg.model and an LLM, which some find underwhelming versus dedicated vector models.
  • Example SVGs in the docs are considered basic; some hoped for more complex showcases.
  • Native GIF generation works but is judged low-quality: splotchy watermarks, inconsistent motion, “slop” artifacts, and weak frame coherence.
  • Animated GIFs are seen as fine for memes/throwaway content, not professional use. No animated SVG support found.

Image quality and model comparisons

  • Several users prefer 2.5 over 2.6, especially for UI/design work and text-in-image layout; 2.6 is described as simpler, more zoomed-in, and less detailed.
  • Others like 2.6’s reduced tendency to add unwanted inspirational text.
  • Benchmarks and anecdotal tests suggest:
    • OpenAI leads on prompt adherence.
    • Google’s models and various closed/open alternatives may lead on realism or style in some domains.
  • Some prompts show 2.6 worse than even 2.0 on nuanced photography (hands, reflections, water caustics).

Artifacts, “AI look,” and aesthetics

  • Persistent complaints about:
    • Visible watermarking / mottling.
    • A characteristic “AI look” (overly aesthetic, glossy, yellowish tint, unrealistic lighting).
    • Uncanny or incorrect physics (e.g., stained-glass whale lighting).
  • Debate on whether users actually like this look:
    • One view: average users don’t notice or care; cheapness wins.
    • Counterview: many people dislike AI food images and uncanny visuals, especially in restaurants/ads; “AI slop” criticism is widespread online.
  • Some argue RLHF and guidance methods push models toward this aesthetic; others say models misapply averaged reward signals.

Capabilities and limitations

  • Pixel art: can approximate the style, but true grid-aligned, limited-palette pixel art still needs post-processing or specialized models.
  • Consistent sprites / sprite sheets: multiple users report poor consistency; attempts at consistent sprite-sheet art are called a waste of time.
  • Character and layout consistency remain hard, though some external models are praised for doing better.

Editing, control, and tooling

  • Users want finer-grained, reliable editing:
    • Frequent complaint: asking to tweak one small element changes the whole image.
    • Some notice claims of better control in 2.6 but don’t yet see clear improvements.
  • Desire for more explicit tools: seeds/gen IDs, compositional sliders, layers, better outpainting UI, and integration of vector/3D tooling.
  • Some miss older DALL·E-era tooling (e.g., outpainting, stable iteration via IDs).

Ethics and cultural impact

  • Strong ethical criticism: training on massive scraped datasets is described as exploiting or “robbing” human creative labor, even if technically enabled by decades of research.
  • Others counter that these systems are themselves products of human effort and innovation.
  • Broader cultural concern about proliferation of low-quality “AI slop” in menus, ads, and social feeds, and about atrophying human observational and aesthetic skills.