# Fal.ai Achieves 6× Speedup Serving Ideogram v4 at Sub-Second Latency

_Research · published 2026-07-09_

Fal.ai has published a technical breakdown of how it reduced Ideogram v4 image generation at 1K resolution from 2.75 seconds down to 0.44 seconds — a 6× speedup — with no visible quality loss. The gains come from attacking both the cost-per-pass and the number of passes in the diffusion transformer: running the model in NVFP4 (4-bit floating point with hardware support on Blackwell GPUs), fusing epilogue operations such as RMSNorm and gated-SiLU directly into the GEMM kernels to eliminate HBM round-trips, and applying timestep distillation to collapse the denoising step count.

Naive FP4 inference introduced visible quality degradation compared to the baseline bf16 model, so the team also developed quantization-aware distillation to recover output fidelity before applying timestep distillation on top. The result stacks all three optimizations — FP4 quantization, kernel fusion, and step-count reduction — to land at the same perceptual quality as the full-precision model at a fraction of the compute.

## Sources
- [fal.ai Blog / News](https://blog.fal.ai/serving-sub-second-ideogram-v4-without-quality-loss/)
- [fal.ai Blog / News](https://blog.fal.ai/how-we-achieved-1000-tok-s-and-16x-throughput-with-dspark-for-ideogram-v4-prompt-expander/)
---
Canonical: https://genbuzz.news/posts/fal-ai-achieves-6x-speedup-serving-ideogram-v4-at-sub-second-latency