Inference-Time Scaling in Diffusion (ReflectionFlow)

Iterative refinement and self-reflection in text-to-image diffusion models

Overview

Recent advancements in Large Language Models (like OpenAI’s o1) have proven that scaling compute at inference time (Chain-of-Thought, self-reflection) dramatically improves output quality.

ReflectionFlow applies this exact paradigm to Text-to-Image Diffusion Models (specifically FLUX.1-dev). It allows the diffusion model to iteratively assess its own output, generate an actionable “reflection” on what is wrong with the image, and refine it.

Inference-Time Scaling Axes

ReflectionFlow introduces three scaling axes during image generation:

  1. Noise-level scaling: Optimizing latent initialization via multiple samples.
  2. Prompt-level scaling: Tweaking semantic guidance.
  3. Reflection-level scaling: Iteratively assessing and correcting previous generations based on an explicit “reflection” text. Deeper sequential refinement consistently outperforms simply generating a wider batch of parallel images.

The GenRef Dataset

To achieve this, the researchers constructed GenRef, a dataset of 1 million triplets structured as (flawed image, refined image, textual reflection).

  • The dataset incorporates rule-based challenging prompts, reward-based scoring (using PickScore, CLIP), and high-quality editing pairs.
  • It leverages a dedicated MLLM verifier (e.g., Qwen2.5-VL-7B) to dynamically generate these textual reflections.

Core Benefit

For highly complex scenes (like intricate spatial arrangements or exact text rendering), a single forward pass of a diffusion model often fails. ReflectionFlow allows the model to act like a human artist: draw a draft, look at it, spot the mistakes (reflection), and redraw the complex parts until perfect.

TODO: Add architectural diagram showing the multimodal input loop.