ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning
Sara Dorfman, Nir Goren, Gal Barak, Daniel Cohen-Or, Or Patashnik
cs.CV, cs.AI, cs.GR
2026-10-16
The student edits its own outputs; the teacher scores from the original image plus the compound instruction. After 10 turns, KRIS consistency rises from 3.03 to 3.85.
A single instruction edit is already usable. In practice the next turn takes the previous output as its input. Training is dominated by clean source photos. Multi-turn inference feeds the editor its own images. The shifts stack: fine texture disappears, color drifts, and regions the instruction never named change too.
Pixel thresholding, high-frequency locks, and low-frequency latent alignment try to stop that at inference, and they cost extra compute. Whatever they pin leaks back when the edit has to change that same region. Labeled multi-turn sequences are a weak substitute. One instruction has many valid images, so once the student leaves the labeled path, later ground-truth frames are not paired targets.
ACEdit carries Self-Forcing from autoregressive video into editing. Self-Forcing means the model is trained on its own previous output, so the inputs it sees in training sit closer to multi-turn inference.
The student is 4-step FLUX.2-klein-9B, with a rank-16 LoRA, 27.8M trainable parameters, about 0.3% of the model. Each rollout runs ten local edits. A turn sees only the previous result and the current instruction. Latents pass forward with no VAE round trip. The gradient goes through the clean estimate at one denoising step of the current turn, and earlier turns are stopped.
With no paired target, supervision is score distillation. The teacher is frozen FLUX.2-klein-base-9B at guidance scale 2. An auxiliary editor of the same size, without guidance, fits the student's current outputs with a flow-matching loss. The clean estimate is noised to a level τ. The student gradient is the gap between the teacher velocity and the auxiliary velocity, and that gap moves the multi-turn result toward the teacher's clean distribution. τ is drawn uniformly from [0, 1).
Conditioning differs on the two sides. The student edits its own previous image. The teacher and the auxiliary see only the original clean image, plus a compound instruction joining every edit up to this turn. Add a bag, make it red, make it leather, add a gold buckle, on the student side, is one teacher prompt on the source: add a red leather bag with a gold buckle.
If the teacher also saw a degraded intermediate, it would treat color cast and blotches as details to preserve, and the accumulated error would enter the loss.
The data are 96 synthetic 1024 images and 960 local instructions: add, remove, replace, resize, or change an attribute. The student is updated 100 times. The auxiliary is updated more often, including 5 warmup steps. One B200 takes about 4.4 GPU-hours. Inference stays 4 unguided steps.
EdiVal-10 has 252 chains of ten local edits on real GEdit-Bench photos, cropped to 1024. Public multi-turn sets mostly stop at three to five turns. Omni IIE Bench averages 4.35 turns, and only 9 of 260 chains reach ten. The edit types are add, remove, replace, color, material, text, count, and position. Each chain reserves one or two anchor objects that no instruction may mention. Claude Sonnet 5.5 writes the instructions.
The table is the mean after turn 10. Gemma-4-31B scores KRIS-Bench quality, consistency, and instruction following from 1 to 5. Anchor distance uses 148 chains. Background distance is LPIPS outside the edited regions.
| Method | Quality | ∆HPSv3 | Consistency | Anchor↓ | Background↓ | Instruction |
| Klein-9B base | 2.73 | −0.78 | 3.03 | 0.314 | 0.351 | 3.31 |
| Emu Edit | 2.90 | −0.15 | 3.32 | 0.196 | 0.264 | 3.29 |
| VAE-LFA | 2.80 | −0.89 | 3.04 | 0.289 | 0.313 | 3.32 |
| FreqEdit | 2.83 | −0.82 | 3.18 | 0.283 | 0.331 | 3.39 |
| ACEdit | 3.19 | −0.17 | 3.85 | 0.203 | 0.254 | 3.38 |
Quality beats every control on the same 4-step backbone. The smallest margin, over Emu Edit, is 0.30±0.08. Consistency is 3.85, against 3.03 for the base model and 3.04 to 3.32 for the training-free patches. Background LPIPS falls from 0.351 to 0.254. Emu Edit sits at 0.264, higher by 0.010±0.005. Anchor distance is 0.203 for ACEdit and 0.196 for Emu Edit, a gap of 0.007±0.012, inside two standard errors. ∆HPSv3 is −0.17 versus −0.15, and that gap is not significant either.
Instruction following is 3.38. FreqEdit, the highest, is 3.39, not a significant difference. The undistilled FLUX.2-klein-base-9B is run at 512 resolution, 30 steps, guidance 4, then upsampled to 1024. Base quality and consistency are 1.42 and 1.08. MT-OPSD reaches 1.95 and 1.29, with instruction following 3.16 against 3.10. That sampler is different, so those rows do not compare directly with the 4-step model. ACEdit's 3.38 is still significantly above 3.16.
On 606 English single edits from GEdit-Bench, HPSv3 drops for every method. ACEdit drops 0.18. The 4-step base drops 0.41. The difference is 0.23±0.02.
After ten deletions in a row, Emu Edit leaves pale blocks, VAE-LFA darkens the frame under a mottled texture, FreqEdit yellows the walls and floor, and MT-OPSD washes the room toward a white blur. ACEdit keeps the walls, the floor, and the workbench. On 15-step and 20-step chains interleaved with keep-the-image-as-is, the base model hardens fur and shifts color. The fine-tune keeps the soft fur.
Ablations use the same 100 student updates. If the teacher also sees the previous output, consistency falls to 2.81. If it sees only the previous image and the current instruction, consistency falls to 2.28, below the base model's 3.03. If the student also sees the source image, consistency rises to 4.08 and instruction following falls to 2.80. A stool turned yellow at turn 3 is blue again at turn 6. Training on 5-edit rollouts reaches consistency 3.29 after ten turns, about a third of the gain to 3.85.
Editing in a chat is multi-turn. The LoRA does not change the 4-step sampler, adds no inference cost, and trains in about 4.4 hours on one B200.
A self-generated trajectory has no single correct intermediate frame. The teacher stays on the clean source and the accumulated instruction. Shown the intermediate image instead, consistency drops below the untuned model.
The improvement is incremental. The hard numbers come from a 10-turn benchmark built for this paper, on one 4-step backbone. Emu Edit still ties anchor distance and ∆HPSv3. For a few-step FLUX.2 Klein in a multi-turn editor, the fine-tune targets the drift those chats produce.
Training uses 96 synthetic images, all local edits. The real-photo benchmark stays inside those eight local types. When edits keep replacing both the subject and the background, current models do not necessarily lose visual quality. Style, weather, and time of day sit outside the main experiment. LME-Bench is left to an appendix, and the quantitative claims here rest on EdiVal-10.
Claude Sonnet 5.5 writes the instructions. Gemma-4-31B assigns the three scores. No human study is reported. The two models can share a preference for images that look tidy, and that can widen the gaps.
The teacher has to perform the compound instruction in one pass. A miss on that pass becomes the training target. That success rate is not measured on its own. The MT-OPSD comparison changes backbone, resolution, and step count together. Beyond ten turns the paper shows qualitative sequences. A single edit still lowers HPSv3 by 0.18. Decay slows, and quality still moves.