Rubric-CEPR: Self-Evolving Image Editing via Reward-Verified Self-Distillation

arXiv:2610.12469 · cs.CV · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Rubric-CEPR: Self-Evolving Image Editing via Reward-Verified Self-Distillation".

Jane: The gist The Rubric-CEPR framework improves a pretrained image editor by verifying its own generations with internal representations through a rubric-augmented Contrastive Edit-Preservation Reward (CEPR) without human targets or external…

Tom: First, who's behind it and why it matters.

Paper summary: Tom: To break down what they're doing, the paper proposes this self evolving framework called Rubric-CEPR where a Planner suggests structured edit instructions from images and then the Editor samples several candidate edits based on those instructions.

Jane: After that, a frozen Critic scores each of those candidates using decomposed rubric checks to verify three main things: how well it realizes the edit, how well it removes the old state, and how much content gets preserved.

Lu: And they use noncompensatory gates to reject any candidate that doesn't pass every single applicable check because if one check fails, the entire score is zero.

Meng: That sounds like a rigorous way to filter out bad edits before they even get used for training, which is smart from an engineering standpoint.

Lalam: The reward R they calculate involves G times the square root of E times P, where G is that feasibility indicator that requires every check to pass, and E and P come from various checks.

Tom: So the core of this CEPR reward includes explicit required state, forbidden state, and preservation checks to give the verification a specific focus.

Jane: The goal here is to distill the best verified candidate into the editor through lightweight adapter training using a weighted denoising SFT method instead of needing external targets.

Lu: It’s interesting how they manage to do this self evolving process without any human edited targets or an external training-time reward model, which is what they say is costly.

Meng: The results show that on Qwen-Image-Edit, Rubric-CEPR improves ImgEdit from four point three six to four point six zero, which translates to a plus five point five percent gain and a twenty four point nine percent gain on object isolation specifically <ref:2610.12469#pg2>.

Lalam: They also mentioned that this same procedure improves Step1X-Edit by seven point eight percent on ImgEdit, showing the method works across different versions of the editor.

Tom: That jump from four point three six to four point six zero is significant, especially when you look at how much better object isolation becomes with that method applied in this self evolving way <ref:2610.12469#pg2>.

Jane: It shows that verifying samples internally can lead to real improvements without needing those external supervision pairs or reward models they were trying to avoid initially.

Conclusion: Tom: So, looking at the whole thing, the Rubric-CEPR framework is all about creating a self evolving image editor that learns and improves from its own verified samples using this internal rubric augmented contrastive edit preservation reward.

Jane: The authors are focusing on avoiding those costly human edited training pairs or external reward models by making the AI judge itself through these internal representations.

Lu: The implication is that we can build systems that improve iteratively just by generating their own data and judging it against a set of internal rules, which is a key direction for future self improving models.

Meng: From an engineering standpoint, it’s about creating a closed loop where the model checks its own work before it gets further refined.

Lalam: It suggests that this approach could be a solid baseline for image editors that improve themselves using only their own verified samples from the start.

Tom: So, in simple terms, they’ve proposed a way to make image editing systems better by having them check their own edits against strict internal criteria rather than relying on outside supervision.

Jane: That’s the core of Rubric-CEPR and it moves us toward systems that can evolve their own quality without needing constant external guidance.

Mohamed bin Zayed University of Artificial Intelligence · Aalto University

cs.CV

Submitted: 2026-10-08

Updated: 2026-10-09

Comments: Project Page: $\href{https://riteshthawkar.github.io/Rubric-CEPR/}{\text{this URL}}$

Code: https://github.com/riteshthawkar/Rubric-CEPR

Project page: https://riteshthawkar.github.io/Rubric-CEPR/unlabeled

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: The gist The Rubric-CEPR framework improves a pretrained image editor by verifying its own generations with internal representations through a rubric-augmented Contrastive Edit-Preservation Reward

Key concepts

Planner
The Planner proposes structured edit instructions from images that are not labeled with edits. This instruction includes details like the edit type, entities involved, the specific region to change, and what content must be preserved or forbidden. It acts as a guide for generating meaningful edits.
Editor
The Editor samples multiple potential candidate edits based on the original image and the instructions provided by the Planner. These are different versions of how an edit could be applied to achieve the desired change, providing a set of options for verification.
Critic
The Critic is a frozen component that scores each candidate edit generated by the Editor. It uses decomposed rubric checks to evaluate three main aspects: whether the edit was realized correctly, if the original state was successfully removed, and if important content was preserved. It uses features already available in the editor.
Rubric-Augmented Contrastive Edit-Preservation Reward (CEPR)
CEPR is a reward function that scores an edit based on feasibility and quality checks. It combines a feasibility indicator (G) with metrics derived from rubric checks for edit realization and content preservation. A candidate must pass all required checks to receive a high score, ensuring only plausible edits proceed.

Terminology

Summary

The gist The Rubric-CEPR framework improves a pretrained image editor by verifying its own generations with internal representations through a rubric-augmented Contrastive Edit-Preservation Reward (CEPR) without human targets or external reward models.

How it works

  1. A Planner proposes structured edit instructions from unlabeled images, such as providing an instruction c and a structured specification s containing edit type, entities, region, required / forbidden state, content to preserve <ref:2610.12469#pg6>.

  2. The Editor samples multiple candidate edits yi ∼ Et(· x, c) <ref:2610.12469#pg4>.

  3. A frozen Critic scores each candidate with decomposed rubric checks for edit realization, removal of the old state, and content preservation using features already exposed by the editor <ref:2610.12469#pg5>.

  4. Noncompensatory gates reject infeasible candidates, as every applicable check must pass, else G=0 <ref:2610.12469#pg6>.

  5. The best verified candidate is distilled into the editor through lightweight adapter training using a weighted denoising SFT <ref:2610.12469#pg6>.

Rubric-Augmented Contrastive Edit-Preservation Reward (CEPR)

The reward R is calculated as R = G · √E × P, where G is the feasibility indicator and E and P are derived from various checks <ref:2610.12469#pg4>. The rubric prompts extend the contrastive edit-preservation reward with explicit required-state, forbidden-state, and preservation checks <ref:2610.12469#pg6>. The feasibility check G requires that every applicable check must pass, else G=0 <ref:2610.12469#pg6>.

Key Components of Verification

The Critic evaluates several components separately to ensure quality, including:

**: edit realization <ref:2610.12469#pg4>. The semantic edit checks include the new state appears, the old state is gone, the right edit is made <ref:2610.12469#pg5>. 3. Semantic edit checks (the new state appears, the old state is gone, the right edit is made) <ref:2610.12469#pg6>. The preservation checks compare unrelated content kept and the change stays local <ref:2610.12469#pg5>. 3. Preservation checks compare unrelated content kept and the change stays local <ref:2610.12469#pg6>. The validity term V uses a relative latent-change mask M = 1[zx − zynorm > δ] to estimate changed regions <ref:2610.12469#pg6>. The reward R = G · √E × P, where E = gmean(A, F, Ectr, Etax), P = gmean(PCEP R, Prub) <ref:2610.12469#pg6>. The geometric mean is used for these aggregates <ref:2610.12469#pg6>. The overall reward is calculated as R = G · √E × P <ref:2610.12469#pg6>. 3. A failed gate makes the candidate ineligible for training, as visual plausibility cannot override a failed edit or preservation check <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.22, and 4.45 at K=2, 4, 8 <ref:2610.12469#pg6>. The reward-best selection from K base samples raises the extraction score from 3.41 (K=1) to 3.89, 4.

Improvements for AI systems

  1. The Rubric-CEPR framework allows an editor to improve itself from its own verified samples, specifically by using a propose–edit–verify–distill loop where only the best gate-passing candidate per source becomes a target for weighted denoising SFT. This enables the system to learn from its own internal representations without relying on costly human-edited training pairs or external reward models.

  2. The system achieves significant gains in object isolation, improving it from 3.41 → 4.26 (+24.9%) on Qwen-Image-Edit, indicating it can perform complex segmentation and removal tasks with higher fidelity than the base model alone, especially when the success criterion is internally checkable.

  3. The system learns to distinguish between plausible failures and correct edits by employing noncompensatory gates, where every applicable check must pass before a candidate is selected. This prevents plausible failures that pass a lenient score from becoming supervision, ensuring the training targets are high-quality.

  4. The Planner component can be improved through training with reward-weighted token SFT, which increases the rate of feasible targets by 62.4%, allowing it to propose more effective and grounded structured edit specifications for the Editor.

  5. The Editor adapter is updated using a loss that combines reward-weighted denoising SFT and an anchor term, ensuring that the model learns to generate edits that satisfy both the structural requirements of the instruction and internal preservation constraints simultaneously.

Sources

Related papers