← 研踪 PaperTrailarXiv:2611.19955 · v1
Equal Dose, Different Coverage:Tracing Rate and Duration in Image Editing
首次提交 2026-11-15 19:52
版本更新 2026-11-15 19:52
雷达观测 2026-11-17 12:06
摘要
Benchmarks for scientific image editing test whether a model draws the outcome of a described process, leaving open which representations carry its errors and whether correcting them yields the right quantity. We study this question with an airbrush of fixed deposition profile, under which exchanging rate and duration preserves the dose and should preserve the endpoint, yet changes the edited coverage. Matched-norm activation patches across text positions and denoising stages show that, in distilled FLUX with Chinese prompts, representations at unchanged text positions (context) carry this discrepancy: early context transfers 3.51 of the 3.98 percentage-point (pp) signed route difference, against 0.09 pp for equal-norm updates at the two changed digits. Early image latents transmit native (unpatched) route differences across two FLUX configurations and OmniGen, while text-position effects depend on configuration. The trace predicts an intervention: share early context between routes while each retains its own digits. On independent English layouts, this reduces the distilled-FLUX absolute route gap from 1.13 to 0.60 pp, outperforms early digit sharing, and improves agreement more than in base FLUX under directly paired tests. Calibrated evaluation, which prescribes an area for each dose, separates this gain from accuracy: sharing the full prompt encoding produces identical route images with target errors of 8-16 pp, and stating the target percentage directly retains a compressed area response. Localization predicts a consistency gain on new layouts, and calibration measures the quantity error that agreement leaves unresolved.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-15 19:52 · 最新观测版本
Equal Dose, Different Coverage:Tracing Rate and Duration in Image Editing
Benchmarks for scientific image editing test whether a model draws the outcome of a described process, leaving open which representations carry its errors and whether correcting them yields the right quantity. We study this question with an airbrush of fixed deposition profile, under which exchanging rate and duration preserves the dose and should preserve the endpoint, yet changes the edited coverage. Matched-norm activation patches across text positions and denoising stages show that, in distilled FLUX with Chinese prompts, representations at unchanged text positions (context) carry this discrepancy: early context transfers 3.51 of the 3.98 percentage-point (pp) signed route difference, against 0.09 pp for equal-norm updates at the two changed digits. Early image latents transmit native (unpatched) route differences across two FLUX configurations and OmniGen, while text-position effects depend on configuration. The trace predicts an intervention: share early context between routes while each retains its own digits. On independent English layouts, this reduces the distilled-FLUX absolute route gap from 1.13 to 0.60 pp, outperforms early digit sharing, and improves agreement more than in base FLUX under directly paired tests. Calibrated evaluation, which prescribes an area for each dose, separates this gain from accuracy: sharing the full prompt encoding produces identical route images with target errors of 8-16 pp, and stating the target percentage directly retains a compressed area response. Localization predicts a consistency gain on new layouts, and calibration measures the quantity error that agreement leaves unresolved.
查看此版本解读 →查看此版本原文 ↗