Research map / Large language models / Reasoning and chain of thought

From Scale to Speed: Adaptive Test-Time Scaling for Image Editing

CVPR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ADE-CoT, an adaptive test-time scaling framework for image editing that allocates sampling budgets by edit difficulty, prunes candidates early with edit-specific verifiers, and stops depth-first once intent-aligned results are found, yielding over 2x speedup over Best-of-N at comparable quality.

Problem

Applying Image-CoT test-time scaling designed for open-ended text-to-image generation to goal-directed image editing wastes compute: fixed budgets, unreliable general MLLM scores for early pruning, and redundant correct results from large-scale sampling.

Approach

Replace fixed budgets with a difficulty estimate from one initial sample; prune with a unified score combining general MLLM score, Grounded SAM2-based edited-region localization score, and CLIP caption consistency computed on a one-step clean-latent preview; filter near-duplicates with DINOv2; then run depth-first sequential generation with an instance-specific yes/no question verifier that stops

Key results

• With N=32 on FLUX.1 Kontext (GEdit-Bench), ADE-CoT reaches G_O 6.695 vs 6.641 for BoN while cutting NFE from 896 to 418 (about 2.2x speedup in eta terms).
• On BAGEL, ADE-CoT reports G_O 6.972 vs 6.908 for BoN with 882 vs 1600 NFE, and a 1.8x speedup.
• On Step1X-Edit, ADE-CoT reaches G_O 7.196 vs 7.157 for BoN with NFE 434 vs 896 (2.0x speedup, per the table).

Surprising findings

• General MLLM scores at early denoising stages misjudge about 40% of pruned samples, which end up with high final scores.
• Many edit cases produce more than 8 candidates sharing an identical best score, indicating large redundancy in large-scale editing sampling.

Limitations

• Relies on a proprietary-or-large MLLM verifier (Qwen-VL-MAX) and multiple MLLM queries per edit, adding overhead not fully counted in NFE.
• Gains over BoN in G_O are small (often under 0.1 points), with efficiency being the main benefit.
• Hyperparameters (gamma, thresholds, t_e, t_l, N_high) are tuned per model and benchmark.

Experimental setup

Datasets: GEdit-Bench AnyEdit-Test Reason-Edit

Benchmarks: GEdit-Bench AnyEdit-Test Reason-Edit

Metrics: VIE-Score (G_SC, G_PQ, G_O) CLIP_im CLIP_out DINO similarity PSNR LPIPS CLIP score reasoning efficiency (eta) outcome efficiency (xi) NFE

Baselines: Best-of-N (BoN) PRM PARM TTS-EF

Builds on: Best-of-N TTS-EF PRM PARM VIE-Score

Closest papers by meaning

• Meta-CoT: Enhancing Granularity and Generalization in Image Editing (CVPR 2026 · similarity 0.89)
• Balancing Preservation and Modification: A Region and Semantic Aware Metric for Instruction-Based Image Editing (ICML 2025 · similarity 0.88)
• VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling (NeurIPS 2026 · similarity 0.87)
• Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
• ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.87)
• EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.87)