# Multi-Turn Image Editing

3 related articles

Subscribe to updates

arXiv Nov 5, 2026
Multi-turn image edits stop smearing: TailSteer steers only the denoising tail, cutting cumulative drift over 15 to 20 turns
In multi-turn image editing, each output becomes the next input, so small unintended changes to untouched content accumulate into visible drift. TailSteer concentrates preservation corrections near the tail of the diffusion denoising trajectory while keeping early dynamics close to the base model, and learns them from the editor's own on-policy rollouts using unchanged source content as reference, without curated editing pairs. Across four backbones in the FLUX, SenseNova and Qwen families, cumulative drift over 15 to 20 editing turns drops substantially while text-to-image and single-turn editing benchmarks stay largely intact.
In multi-turn image editing, each output becomes the next input, so small unintended changes to untouched content accumulate into visible drift. TailSteer concentrates preservation corrections near the tail of the diffusion denoising trajectory while keeping early dynamics close to the base model, and learns them from the editor's own on-policy rollouts using unchanged source content as reference, without curated editing pairs. Across four backbones in the FLUX, SenseNova and Qwen families, cumulative drift over 15 to 20 editing turns drops substantially while text-to-image and single-turn editing benchmarks stay largely intact.
Ruikang Li（First author）
TailSteer Multi-Turn Image Editing Trajectory-Tail Steering Add & Preserve
arXiv Oct 29, 2026
Image-editing agents stop retrying the same tool: routed specialists plus verification records win 78.5% of expert comparisons
EditAgent widens an image-editing agent's action space beyond a general editor to specialized methods, and attaches to each candidate a verification record of change maps, close-ups and measured color and geometry, so an off-the-shelf VLM judge compares candidates in one listwise call. On a 600-set, 30-edit-type expert-annotated diagnostic benchmark, the general editor yields no usable result in 42.7% of sets and specialized methods recover 175 of them; in a blind study on 144 Complex-Edit samples, experts preferred parallel search over sequential refinement in 78.5% of decided comparisons, at no more than half the judge calls.
EditAgent widens an image-editing agent's action space beyond a general editor to specialized methods, and attaches to each candidate a verification record of change maps, close-ups and measured color and geometry, so an off-the-shelf VLM judge compares candidates in one listwise call. On a 600-set, 30-edit-type expert-annotated diagnostic benchmark, the general editor yields no usable result in 42.7% of sets and specialized methods recover 175 of them; in a blind study on 144 Complex-Edit samples, experts preferred parallel search over sequential refinement in 78.5% of decided comparisons, at no more than half the judge calls.
Haiyang Xu（First author）
EditAgent Agentic Image Editing Multi-Turn Image Editing
arXiv Sep 29, 2026
MT-OPSD uses on-policy self-distillation on self-generated states to lift open-source image editors from 0.03–0.15 to 0.38–0.52 ten-turn success
The work finds that instruction-based image editing models degrade rapidly under recursive multi-turn editing and attributes this to a train–test mismatch in the conditioning distribution, since models are trained on clean source images but must repeatedly condition on their own imperfect outputs at inference; it proposes MT-OPSD, which rolls out the current student under an identity instruction to produce self-generated states containing its own errors, then uses the same model conditioned on the clean source as a teacher to transfer clean-condition editing behavior onto those states via sparse velocity matching, without multi-turn annotations; it also builds LME-Bench of 100 ten-turn editing sessions, raising SR@10 to 0.38–0.52 and cutting CR@10 to at most 0.
The work finds that instruction-based image editing models degrade rapidly under recursive multi-turn editing and attributes this to a train–test mismatch in the conditioning distribution, since models are trained on clean source images but must repeatedly condition on their own imperfect outputs at inference; it proposes MT-OPSD, which rolls out the current student under an identity instruction to produce self-generated states containing its own errors, then uses the same model conditioned on the clean source as a teacher to transfer clean-condition editing behavior onto those states via sparse velocity matching, without multi-turn annotations; it also builds LME-Bench of 100 ten-turn editing sessions, raising SR@10 to 0.38–0.52 and cutting CR@10 to at most 0.
MT-OPSD On-Policy Self-Distillation Multi-Turn Image Editing LME-Bench

You’re all caught up

Multi-turn image edits stop smearing: TailSteer steers only the denoising tail, cutting cumulative drift over 15 to 20 turns

Image-editing agents stop retrying the same tool: routed specialists plus verification records win 78.5% of expert comparisons

MT-OPSD uses on-policy self-distillation on self-generated states to lift open-source image editors from 0.03–0.15 to 0.38–0.52 ten-turn success