Spectral Progressive Diffusion Speeds Up Ideogram AI Inference By 1.6×

We are working hard to integrate SPEED into various open source frameworks and will release our code soon. Stay tuned!

[4/4]

Check out our original paper post here:

Ideogram4 uses Qwen3-VL-8B-Instruct as its text encoder, unlike CLIP or T5 used in FLUX-family models. Its maximum text token count is 2048, which is significantly longer than FLUX's 512. For diffusion inference speedup methods that only reduce image tokens (like ours), the speedup is inevitably bottlenecked by long prompts with large numbers of text tokens. For future work on efficient diffusion inference, designing methods that jointly reduce both text and image tokens in these unified token streams will be crucial!

[2/4]

The text rendering capabilities of Ideogram 4 are largely retained even after SPEED integration, demonstrating the robustness of our method.

[3/4]

@ideogram_ai 's model Ideogram 4:

@BrianCChao @ideogram_ai Does SPEED work with distilled models like DMD/TDM?

@BrianCChao @ideogram_ai incredible!!