LME-Bench

Long-horizon Multi-turn image Editing benchmark, introduced in MT-OPSD: On-Policy Self-Distillation for Multi-Turn Image Editing.

Existing multi-turn editing benchmarks stop at five turns. LME-Bench has 100 ten-turn sessions in which every instruction is applied to the previous turn's output, to measure whether an editor keeps following instructions and keeps its images intact over long horizons.

• 100 source images, 1024×1024, generated with Z-Image-Turbo; 10 categories × 10 sessions (pet, wild, person, food, vehicle, interior, nature, object, plant, urban).
• 10 instructions per session, written against the actual source image: 6 local edits (add / remove / count / color / material / size / replace / state) and 4 global edits, either hard (art style, black-and-white or sepia grade, weather or time of day) or soft (color temperature, brightness or contrast). Hard globals come early, so later turns edit an already transformed image. Instructions only move forward (no undo / restore).

Files

File	Content
images/<id>.png	source image of session <id> (0001-0100)
metadata.jsonl	one row per session: file_name, id, category, source_prompt, turns (10 instructions), labels (hard / soft / local per turn)
lme_bench.json	the same sessions as one JSON file (used by the evaluation code)

hf download metazlb/LME-Bench --repo-type dataset --local-dir benchmark/data

Evaluation

Sampling and evaluation code: https://github.com/liangbingzhao/MT-OPSD/tree/main/benchmark

Each turn edits the previous output. A GPT-4o judge reports two numbers per turn:

• SR@k (success rate): fraction of sessions whose turns 1..k all followed the instruction (prompt following > 6) while preserving what should be preserved (consistency > 6).
• CR@k (collapse rate): fraction of sessions that have collapsed by turn k: two consecutive turns where content loss + surface corruption ≥ 2 (each scored 0-2), counted from the first such turn onward.

Model	SR@3	SR@5	SR@8	SR@10	CR@3	CR@5	CR@8	CR@10
GPT-Image-1	0.82	0.48	0.09	0.01	0.03	0.35	0.63	0.69
GPT-Image-2	1.00	0.98	0.98	0.96	0.00	0.00	0.00	0.00
Nano Banana	0.67	0.66	0.63	0.56	0.00	0.00	0.00	0.00
Nano Banana Pro	1.00	1.00	0.97	0.91	0.00	0.00	0.00	0.00
Qwen-Image-Edit-2511	0.98	0.43	0.08	0.03	0.00	0.10	0.45	0.55
+ MT-OPSD	1.00	0.91	0.63	0.44	0.00	0.01	0.02	0.02
FireRed-Image-Edit-1.0	0.95	0.57	0.28	0.15	0.01	0.16	0.51	0.61
+ MT-OPSD	0.98	0.88	0.64	0.52	0.00	0.00	0.01	0.03
FLUX.2 [klein] base 9B	0.92	0.52	0.29	0.12	0.00	0.01	0.17	0.25
+ MT-OPSD	0.95	0.76	0.57	0.38	0.00	0.00	0.02	0.04

Citation

@article{zhao2026mtopsd,
  title={MT-OPSD: On-Policy Self-Distillation for Multi-Turn Image Editing},
  author={Zhao, Liangbing and Zhuo, Le and Elhoseiny, Mohamed},
  year={2026}
}