The company fal has introduced two new versions of the Ideogram 4 model focused on maximum generation speed: Ideogram 4 Instant (8 steps) and Ideogram 4 Fast (20 steps). By utilizing step distillation technology and eliminating the need for Classifier-Free Guidance (CFG) during inference, these new checkpoints enable sub-second image generation without loss of quality.
What Happened
fal has released two optimized checkpoints — Ideogram 4 Instant and Ideogram 4 Fast. Both variants are based on the ideogram-4-fp8 model and contain approximately 9.28 billion parameters. The key technical change is the application of timestep distillation, which has reduced the required number of steps to 8 and 20 respectively, and eliminated the need for CFG during runtime.
Context
Traditional Text-to-Image (T2I) models often require Classifier-Free Guidance (CFG), which implies double passes of the model (conditional and unconditional) for every generation step, significantly increasing latency. The use of specialized quantized weights, such as FP8, combined with distillation methods, is becoming the new standard for creating efficient and scalable models.
Why It Matters for the Industry
Moving toward sub-second generation via distillation and the removal of runtime CFG radically reduces inference costs and makes the technology commercially viable for high-load APIs. This sets a new efficiency standard for heavy T2I models, paving the way for the mass adoption of specialized quantized formats (such as FP4 or QAD) in production environments.
Why It Matters for Users
Users can receive high-quality images in the Ideogram style almost instantaneously, opening up possibilities for creating interactive interfaces with "live" design. It is important to note that when running locally via Diffusers, a structured JSON prompt format must be used, as the automatic prompt expansion feature is only implemented in the fal.ai cloud.
Sources
Author
Look at AI, Editorial Team