Unlocking the Potential of Image Editing via Concept Scaling and Dense Supervision
Long Cui, Xiaoqian Liu, Qi Qin, Yi Xin, Tao Lin, Jianguo Li, Linfeng Zhang
cs.CV
2026-08-18
SJTU and Ant train on 1,000 edit concepts, 12M pairs, and composite local edits. At 5M on Z-Image, ImgEdit overall is 3.75 versus ScaleEdit’s 3.31.
Instruction-based image editing still trains like text-to-image: condition on a source image and a sentence, predict the edited image. Two mismatches sit under that copy. Datasets have scaled source-image diversity while edit concepts stay in a handful of coarse bins such as “change action” or “change style.” Editing is also a sparse objective. Only a patch of pixels actually changes; the rest of the loss is an identity map. Each pair therefore delivers little gradient on the transformation itself.
Shanghai Jiao Tong University and Ant Group split edit operations into more than a thousand leaves, down to “finger heart,” “shrug,” or “anxious.” They also measure what happens when a VLM writes instructions freely: in style transfer, the top five styles take 74.6% of samples and dozens of others sit under 1%. The bottleneck looks like concept coverage, not another round of source-image variety.
A seed taxonomy is expanded by an LLM, then cleaned by people, ending at 1,028 leaf concepts. For each source image a VLM picks compatible concepts, writes instructions, and emits a case-specific VQA checklist. An editing model synthesizes the target; those instance-level questions filter the pair instead of a generic template. A small random-exploration branch remains, so the tree does not freeze the tail.
Dense supervision composes several spatially disjoint local edits into one pair and one joint instruction. A single forward pass then contains multiple active edit regions, and background reconstruction takes a smaller share of the loss. Training mixes ConceptEdit-1000 with composite pairs 1:1. The backbone is Z-Image. Instructions and filters use Qwen3.5-122B-A10B; synthesis uses FLUX.2-klein-9B. Comparisons against UnicEdit and ScaleEdit run at 2M and 5M samples.
On ImgEdit, the full recipe (1,000 concepts plus composition) scores 3.48 at 2M and 3.75 at 5M, beating ScaleEdit by 0.31 and 0.44. Concept granularity has its own ladder: at 2M, 10 / 500 / 1,000 concepts score 3.05 / 3.26 / 3.33. Composition adds another 0.15 overall, and not only on the Compose slice: at 5M, Adjust, Replace, and Action rise by 0.20, 0.26, and 0.25. Matching the composite curve without composites takes about 1.5× samples.
| Setting (5M) | ImgEdit overall | GEdit-EN GO | GEdit-CN GO |
| ScaleEdit | 3.31 | 5.77 | 5.63 |
| ConceptEdit-1000 | 3.60 | 6.40 | 6.36 |
| 1,000 + composition | 3.75 | 6.62 | 6.60 |
Gains on GEdit come mainly from instruction-following GSC. Instance-specific VQA versus generic validation: precision 84% vs 75%, recall 87% vs 57%, F1 86% vs 65%. The appendix ConceptEdit-Bench scores existing models across 1,000 leaves: Nano Banana 2 at 66.19, open-source FireRed-Image-Edit-1.0 at 65.86. Portrait and composition are the steepest drops for almost every model.
Remove is not a clean win. At 2M the full recipe sits 0.41 below ScaleEdit on that category. Scaling concepts does not lift every bin.
For open-source instruction editors, “add another 10 million pairs” splits into two knobs: control the concept distribution, and pack more active supervision into each pair. A library blocks VLM mode collapse. Composition treats sparsity as a compression problem, not a separate multi-edit product task. For a tight training budget, 1.5× sample efficiency is more actionable than another crawl of source images.
ConceptEdit-Bench is most useful when overall scores cluster. Portrait micro-expressions and spatial composition are where current models still fall over. The authors’ own trained model is not on that fine-grained table; the bench currently diagnoses other people’s systems.
Training is Z-Image only. Whether the same data recipe holds for other diffusion backbones or unified multimodal models is untested. The 5M UnicEdit baseline repeats samples because the public release is smaller, so that comparison is not fully clean. Synthesis quality is capped by FLUX.2-klein-9B and the Qwen filter; a finer taxonomy cannot rescue edits the synthesizer systematically botches. The tree keeps cross-category overlap on purpose, so “neon” appears under style, environment, and lighting, which can dilute diagnostic granularity. Main runs stop at 5M; extra gain from the full 12M is unreported. Dataset and code are promised around publication. This write-up has not seen the raw pairs.