When Black Forest Labs released Flux 2 Dev in November 2025, fal followed in late December with Flux 2 Dev Turbo: a distilled LoRA on top of FLUX.2 [dev], built for 8-step generation instead of the 28 steps Flux 2 Dev runs by default (ImageGPT exposes 4-8 steps), while maintaining much of the quality improvement. This puts it in direct competition with Flux 1 Schnell, the original speed-focused model from 2024.
Both models target the same use case: fast generation for real-time applications, rapid iteration, and cost-sensitive workflows. Flux 1 Schnell generates images in roughly one second while Flux 2 Dev Turbo takes about 1.5 seconds. That's roughly 2.7× the cost for about 50% more generation time.
The key question is whether the quality gap justifies that difference. Flux 2 Dev Turbo inherits improvements from the FLUX.2 architecture: better prompt adherence, improved detail rendering, and support for image-to-image workflows. Its ELO score (~1159) sits noticeably higher than Schnell's (~1050), suggesting meaningful quality improvements in blind comparisons.
One significant feature difference: Flux 2 Dev Turbo supports image-to-image generation with adjustable guidance (0-20), letting you use a source image as a starting point. Flux 1 Schnell also accepts an input image on ImageGPT, so this is less of a differentiator than it looks.
NoteFlux 2 Dev Turbo appears in ImageGPT's "quality/fast" route as the last fallback, after the Klein and Schnell options, providing upgraded quality when the fastest options are unavailable.