MIRA: Multimodal Iterative Reasoning Agent for Image Editing

Instruction-guided image editing allows users to modify images with natural language prompts, but existing diffusion-based editors often struggle when instructions are long, compositional, or context dependent. They tend to drift semantically or miss fine-grained constraints.

MIRA is a lightweight multimodal agent that performs editing via an iterative perception–reasoning–action loop rather than a single-shot prompt. At each step, the model observes the original image, the current edited image, and the complex instruction, then predicts the next atomic edit or a stop signal. An external editor executes the atomic instruction, and the updated image is fed back for the next decision.

Paired with open-source editing models such as Flux.1-Kontext, Step1X-Edit, and Qwen-Image-Edit, MIRA significantly boosts semantic consistency and perceptual quality, reaching or surpassing proprietary systems like GPT-Image and Nano-Banana on challenging benchmarks.

Given an original image I₀ and a complex instruction C, MIRA maintains an intermediate edited image Iₜ₋₁. At each iteration:

• The policy π receives (I₀, Iₜ₋₁, C) and outputs an atomic edit instruction uₜ.
• An external image editor E (e.g., Flux.1-Kontext) executes uₜ, producing a new image Iₜ.
• A termination controller decides whether to continue or <STOP>.

This loop continues until the global goal encoded by C is satisfied. Unlike pre-planned, open-loop pipelines, MIRA can detect mis-executed edits and issue corrective instructions, making it naturally robust to backend model errors.

# Pseudocode example of MIRA-style editing loop

I0 = load_image("input.png")
C  = "Change the white stove to black, make the floor wooden, and turn the cabinets brown."

I  = I0
while True:
    u = mira_policy(I0, I, C)      # atomic instruction
    if u == "<STOP>":
        break
    I = editing_model(I, u)        # e.g., Flux.1-Kontext, Step1X-Edit, Qwen-Image-Edit

save(I, "output.png")

• Instruction aggregation: multi-turn editing sessions (e.g., from SeedEdit) are merged into single complex instructions, with permutations to enrich compositional structures.
• Two-level rewriting: both atomic edits and aggregated instructions are paraphrased for linguistic diversity.
• Candidate generation & ranking: strong open-source editors generate multiple trajectories; ViEscore-based models rank them, and the best is kept as supervision.

Each dataset entry includes the input image, a high-quality edited trajectory, the complex instruction, and its variants, all decomposed into start, continue, and stop samples for training the iterative policy.

On a 500-sample benchmark constructed from MagicBrush and CompBench subsets, MIRA-enhanced models narrow the gap to GPT-Image and Nano-Banana, while remaining fully open-source and modular.

Qualitative comparisons in the paper further show that MIRA can:

• Follow long, multi-object instructions faithfully.
• Gradually refine layouts and appearances with visible intermediate states.
• Issue corrective edits when backend models make mistakes (e.g., recoloring a fridge back to white after an unintended change).

@article{zeng2025mira,
  title   = {MIRA: Multimodal Iterative Reasoning Agent for Image Editing},
  author  = {Zeng, Ziyun and Hua, Hang and Luo, Jiebo},
  journal = {arXiv preprint arXiv:YYYY.YYYYY},
  year    = {2025}
}

For questions, implementation details, or collaboration, please reach out to the authors via their institutional email addresses or open an issue on the project repository.