Ideogram 4 TurboTime Cuts Image Generation to 4 Steps Without Distillation
A new LoRA by ostris slashes Ideogram 4 inference from 20+ steps to just 2-4, using a novel non-distillation training technique still being refined.
Takeaways
- Ostris released Ideogram 4 TurboTime, a free LoRA that cuts Ideogram 4 inference from 20+ steps to just 2-4 steps.
- It eliminates the need for CFG (Classifier-Free Guidance), removing the double-pass compute cost on top of the step reduction.
- Built with a novel non-distillation continuous turbo training method -- full details to be published once the technique is polished.
- Ideogram 4 is the #1 open-weight image model on DesignArena, known for best-in-class text rendering and 2K output.
- Drop-in compatible with the standard diffusers pipeline; load it with
pipe.load_lora_weights("ostris/ideogram_4_turbotime_lora"). - No formal quality benchmarks published yet; the LoRA will receive updated versions as the training method improves.
Ideogram 4 TurboTime is a LoRA (a small adapter file that modifies a model's behavior without replacing its weights) released by community researcher ostris. It plugs into Ideogram 4 and cuts inference from the typical 20+ denoising steps down to just 2-4, with no meaningful change to your pipeline beyond loading the adapter. That is a potential 5-10x wall-clock speedup, and it is available right now for free on Hugging Face.
The model it sits on top of
Ideogram 4 is Ideogram's first open-weight text-to-image model, trained from scratch and not a fine-tune of any existing model. It introduces a structured JSON prompting interface with best-in-class multilingual text rendering, explicit bounding-box layout and color-palette controls, and native 2K resolution output. It currently holds the #1 spot on DesignArena's third-party open-weight leaderboard. In short, it is one of the most capable open-weight image models available, and it is slow by default.
What TurboTime actually does
This LoRA turns Ideogram 4 into a few-step model, capable of generating good quality images in as little as 2 steps with no CFG and no unconditional model. CFG, or Classifier-Free Guidance, is the technique where a diffusion model runs twice per step -- once with your prompt, once without -- and amplifies the difference to steer the output toward what you asked for. Eliminating it halves the compute per step on top of the step count reduction.
Built on the diffusers library, this LoRA operates as an adapter layer on top of Ideogram 4, enabling dramatic speed improvements -- from typical 20+ step inference down to 2-8 steps -- while maintaining visual quality. The sweet spot ostris recommends is 4-8 steps for the best quality-to-speed tradeoff, with 2 steps available when you need maximum throughput.