Skip to main content
Back

OctoT2I Beats Single-Model Scaling With 90% Faster Image Generation

OctoT2I routes prompts across multiple image models autonomously, matching top-tier quality while cutting inference time by 90% and energy use by 57%

  • OctoT2I is an agentic router that dispatches text-to-image prompts to the best model in a pool, optimizing for both quality and speed.
  • It achieves a GenEval score of 0.96 -- matching or beating top single-model baselines -- with a 90.3% inference speedup and 56.6% energy savings over Flow-GRPO.
  • A Self-Evolving Mechanism builds the model capability knowledge base automatically via a Propose-Solve-Evaluate-Learn (PSEL) loop, requiring zero human annotations.
  • The router uses long-term knowledge + short-term session memory to make multi-round decisions, retrying with a different model if the first attempt underperforms.
  • The framework is model-agnostic: new T2I models can be profiled and added by re-running the PSEL loop, no retraining needed.
  • Code and models will be released by the authors.

Scaling a single text-to-image model has a ceiling. You can pour more compute, more data, and more RL fine-tuning into one model -- and the gains keep shrinking. OctoT2I takes a different approach entirely: instead of making one model smarter, it builds an intelligent router that dispatches each prompt to whichever model in a pool is best suited to handle it. The result is near state-of-the-art image quality at a fraction of the compute cost.

The wall that single-model scaling hits

The text-to-image space has exploded with models of every size -- from massive diffusion transformers like FLUX to real-time lightweight generators like SANA. Each has distinct strengths. A model tuned for photorealism may struggle with precise object counting. A fast, low-parameter model might nail simple scenes but fall apart on complex spatial compositions.

Existing agentic T2I methods suffer from three key challenges: reliance on expensive handcrafted priors or human annotations, rigid single-path decision mechanisms, and a neglect of inference efficiency. Previous multi-model systems typically required humans to manually define which model handles which type of prompt -- a brittle, expensive process that doesn't scale. And even when they worked, they tended to always run the most powerful (and slowest) model, defeating the efficiency argument.

What OctoT2I actually does

OctoT2I is a novel agentic framework that reformulates the T2I task as a joint optimization of generation quality and inference efficiency, implementing a stateful, multi-round routing strategy that adaptively selects the most suitable tool based on its knowledge and memory. Think of it as a dispatcher that has studied every model's strengths and weaknesses in advance -- and keeps a running memory of what's worked in the current session.