ADu2021/editthinker-iterative-reasoning

Overview

EditThinker implements a deliberative "Think-while-Edit" cycle where a multimodal language model critiques results, generates reasoning explanations, and produces refined instructions that guide the next iteration. This addresses inherent stochasticity in generation models through multi-turn reasoning rather than attempting perfection in single outputs.

When to Use

• Complex image editing instructions require multiple refinement passes
• You need the model to explain its editing decisions and propose improvements
• Instruction-following quality is critical across various image editing tasks
• Models struggle with ambiguous or multi-step editing requirements

When NOT to Use

• Single-turn editing where one pass is sufficient
• Real-time applications requiring instant results
• Tasks where iteration overhead is unacceptable
• Simple transformations (crop, resize, rotate)

Core Technique

The EditThinker model jointly produces critique scores, reasoning processes, and enhanced instructions:

# Iterative editing framework
class EditThinker:
    def __init__(self, vllm_model, editor):
        self.vllm = vllm_model  # Vision-language model
        self.editor = editor     # Any compatible image editor

    def refine_edit(self, image, instruction, max_iterations=3):
        current_image = image
        for iteration in range(max_iterations):
            # Generate one-step edit
            current_image = self.editor.apply(current_image, instruction)

            # Get critique and reasoning
            critique = self.vllm.critique(image, current_image)
            reasoning = self.vllm.explain_improvements(critique)

            # Generate enhanced instruction
            instruction = self.vllm.refine_instruction(
                instruction, reasoning, critique
            )

            if critique.satisfactory:
                break

        return current_image, instruction

Training uses reinforcement learning to align the model's reasoning with actual editing improvements, generating more targeted instruction modifications.

Key Results

• Significant improvements across four evaluation benchmarks
• Works with multiple image editing model variants
• Instruction-following capability enhanced through multi-turn interaction
• Addresses inherent generation stochasticity

Implementation Notes

• Compatible with any image editor backend
• RL training ensures reasoning correlates with actual improvements
• Critique scores guide iteration depth
• Preserves intermediate reasoning for interpretability

References

• Original paper: https://arxiv.org/abs/2512.05965
• Focus: Instruction-following in image editing
• Domain: Generative modeling, multimodal reasoning