Serving sub-second Ideogram v4 without quality loss

TLDR Dev · 1 week ago

FAL reduced Ideogram v4 image generation latency from 2.75 seconds to 0.44 seconds at 1K resolution through FP4 quantization, kernel fusion optimizations, and distillation techniques. The approach involves running the diffusion transformer in FP4 with fused epilogue operations (RMSNorm and gated-SiLU), then using quantization-aware distillation and timestep distillation to maintain quality while reducing computational cost. The optimizations maintain visual parity with the full BF16 model while achieving a 6x speedup across all inference parameters.