You know what’s really weird, and I’m not the first to point this out (credit to Peter Gostev): image generation quality seems barely correlated with model price.
With LLMs, a model that costs 5x more is generally better than one that costs 5x less, though prices drop quickly over time for a given quality bar.
With image generation, whatever the best model is at the time seems barely correlated with price. For a while it was GPT Image 1 at $0.20, before that Recraft at $0.04, and now Seedream at $0.04. It just goes up and down; there’s no clear price–quality correlation.
To me this implies we still have a lot of headroom in image quality and cost efficiency.
Within a model family like Flux, you can get multiple tiers that generally follow a given cost–quality curve.
Maybe it’s because there’s less competition in image generation. We’re still in a phase where new architectural insights can translate to big quality gains, like when DiT replaced regular diffusion and with autoregressive models appearing.
Maybe when things consolidate around autoregressive winning (which it seems we’re heading toward), we’ll see a more LLM-like cost tiering where quality climbs and the same quality drops in price over time. 🤞
Or maybe image generation will always be messy, with diffusion staying the most cost-efficient, autoregressive having different strengths, and quality being use-case dependent.
TL;DR: it still feels like the Wild West for image generation, which I wouldn’t have predicted a couple years ago. Watching video generation too.
