An image that used to take nearly three seconds now displays in less than half a second

The fal inference platform has optimized the Ideogram v4 image model service to bring it under the one-second mark. At 1K, a generation that previously required 2.75 seconds now runs in 0.44 seconds—an approximately sixfold speedup that fal claims to achieve with no visible degradation in quality compared to the full model.

The reasoning stems from a simple equation: the cost of an image depends on the number of steps executed by the diffusion transformer, multiplied by the cost of each step. fal therefore worked on both fronts. First, making each step cheaper by running the transformer in FP4—a 4-bit format natively supported by recent NVIDIA hardware—and by fusing small operations around each matrix multiplication to avoid costly memory round-trips. Second, reducing the number of steps through distillation: on one hand, by collapsing the classifier-free guidance, which required two passes per step, into a single pass; on the other hand, by reducing the numerous denoising steps to just a handful.

The tricky part was the color: the raw transition to 4-bit desaturated the images, as errors accumulated from one denoising step to the next. fal corrected this through quantization-aware distillation, where a 4-bit model learns to reproduce the outputs of a full-precision model while incorporating its own quantization noise.

Two production tiers rely on this pipeline: Fast, which collapses the guidance branches, and Instant, which additionally reduces the number of steps and proves to be about six times faster than the base version.