IE-Critic-R1: Human-Aligned Metric for Text-Driven Editing

• IE-Critic-R1 is a multimodal critic model that uses Chain-of-Thought reasoning and RLVR to score text-driven image edits based on image, prompt, and human opinion scores.
• It integrates a vision transformer and text encoder within a fused transformer backbone to generate transparent scores that outperform traditional metrics like CLIP, LPIPS, and SSIM.
• Its RLVR training framework and reward shaping via linear policies ensure reliable human alignment and robust generalization across diverse image editing tasks.

IE-Critic-R1 is a learned, explainable, and human-aligned quality metric for text-driven image editing tasks. It introduces a multimodal “critic” model that evaluates the perceptual quality of edited images based on both an input source image and a text prompt, leveraging a Chain-of-Thought (CoT) reasoning trace and reinforcement learning from verifiable rewards (RLVR) to maximize agreement with human mean opinion scores (MOS). IE-Critic-R1 is part of a new paradigm of critic-based metrics that provide explanatory assessment, outperforming existing approaches such as CLIP-score, LPIPS, and SSIM on human-alignment for text-driven image editing (Qu et al., 22 Nov 2025).

1. Problem Setting and Motivation

The central task addressed by IE-Critic-R1 is automated scoring of text-driven image editing outputs. Given a source image $S$, a natural language edit instruction $P$, and an edited image $I$, the goal is to score $I$ on three axes: (a) text-alignment (faithfulness to $P$); (b) fidelity (appropriate preservation or transformation relative to $S$); and (c) overall perceptual quality as perceived by humans.

Traditional evaluation metrics such as CLIP-score (for text-image alignment), LPIPS and SSIM (for image similarity) are limited, as each covers only a single aspect and does not adequately track human subjective judgments on perceptual quality for complex editing tasks. IE-Critic-R1 overcomes this by building on the IE-Bench dataset, which contains approximately 4,000 annotated samples (triplets $(S, P, I)$) with MOS from 15 raters and explicit subscores for text-alignment, fidelity, and overall quality. The critic model is trained to score edits both accurately and explainably, closing the gap between automated and human evaluation (Qu et al., 22 Nov 2025).

2. Mathematical Formulation

IE-Critic-R1 is formalized as a sequence-to-reward problem:

• Inputs: $(S, I, P)$
• Critic Output: Textual reasoning $y$ consisting of a CoT and a scalar score $s_{\text{pred}} \in [1,5]$ wrapped in tagged output.
• Target: Human-rated MOS $s_{\text{gt}} \in [1,5]$ (z-normalized, then mapped to $[1,5]$).

The reward function for model output $y$ is:

$$r(y; S, I, P) = r_{\text{acc}}(s_{\text{pred}}, s_{\text{gt}}) + \lambda \cdot r_{\text{fmt}}(y)$$

• $r_{\text{acc}}$ measures proximity of $s_{\text{pred}}$ to $s_{\text{gt}}$; four families of reward shaping were considered:
  • Linear ($\ell_1$): $$r_{\text{acc}} = \max(1 - \alpha |s_{\text{pred}} - s_{\text{gt}}|,\ r_{\min})$$
  • Quadratic ($\ell_2$): $$r_{\text{acc}} = \max(1 - \alpha (s_{\text{pred}} - s_{\text{gt}})^2,\ r_{\min})$$
  • Laplacian
  • Gaussian

The linear scheme yielded the best empirical results.

• $r_{\text{fmt}}$ is a formatting reward (1 if strict output tags are present, 0 otherwise), with $\lambda=1$.

At inference, $s_{\text{pred}}$ is parsed from the output to serve as the metric’s quality score $R_\theta(S, I, P)$.

3. Critic Model Architecture

The critic model in IE-Critic-R1 is based on Qwen-2.5-VL-7B-Instruct, a multi-modal LLM featuring:

• Image encoder: Vision Transformer (ViT-style) for patch embedding and positional projection; processes both $S$ and $I$ into sequences of visual tokens.
• Text encoder: Standard tokenizer and embedding for the edit prompt $P$.
• Multimodal backbone: Interleaved transformer blocks fuse visual and text tokens with cross-attention, enabling integrated processing of both images and prompt.
• Decoder head: Generates a Chain-of-Thought trace (outputting a “> …” segment) concluding with a final scalar score marked as “<answer> $s_{\text{pred}}$ </answer>”.

By integrating both images and the prompt, the critic assesses the quality of the edit in a causally explainable sequence.

4. RLVR Training Framework

IE-Critic-R1 employs Reinforcement Learning from Verifiable Rewards (RLVR) with Group Relative Policy Optimization (GRPO):

• Initialization: The critic undergoes supervised fine-tuning (SFT) on the IE-Bench dataset using both CoT and direct scoring examples.
• Group sampling: For each input $(S,I,P)$, $G$ trajectories $\{y_i\}_{i=1}^G$ are sampled under the current policy $\pi_\theta$.
• Advantage estimation: Compute trajectory rewards $r_i$; compute group mean $\bar{r}$ and standard deviation $\sigma_r$; the advantage is $A_i = (r_i - \bar{r})/\sigma_r$.
• Policy update: The model is updated with a clipped PPO-like objective (no KL term), directly maximizing the expected group-averaged advantage.
• Hyperparameters: Typical choices include $E=5$ episodes, $G=8$ rollouts, batch size 128, learning rate $10^{-6}$, and full-score formatting enforced.

This RLVR regime enforces alignment with human assessment and encourages verbose, rational CoT traces, preventing reward hacking and ensuring the model maintains a stable scoring policy over successive updates.

5. Experimental Evaluation

IE-Critic-R1 was benchmarked extensively on IE-Bench and generalization benchmarks:

Setting	Baseline (MainScore)	IE-Critic-CoT (SFT)	IE-Critic-R1 (CoT+RLVR)
Full-Context (S,I,P→score)	0.8208	0.8304	0.8661
Edited-Only (I,P→score)	0.6344	—	—
AGIQA-3k (generalization)	0.8971	0.8948	0.9155

• Ablation studies: The two-stage recipe (CoT+direct scoring, then RLVR) is essential; single-stage SFT or RLVR on direct scoring alone leads to lower alignment or reward hacking.
• Reward shaping: Linear ($\ell_1$) shaping achieved maximal MainScore; alternatives were consistently lower or led to collapsed CoTs.
• Generalization: CoT+RLVR with full context provided strong cross-benchmark generalization (IE-Bench to AGIQA-3k).

6. Alignment with Human Perception and Interpretability

IE-Critic-R1 demonstrates strong empirical alignment with human ratings:

• Outperforms single-axis metrics (CLIP, LPIPS, SSIM) by more than 0.3 MainScore on the Edited-Only setting and more than 0.04 on Full-Context.
• Generates explainable CoT traces, making the rationale for each score transparent along axes of text-alignment, fidelity, and perceptual quality.
• RLVR fine-tuning ensures both robust accuracy and a stable response style, maintaining model reliability over time.
• The critic architecture is agnostic to the nature of the text edit (style, semantics, structure), enabling applicability to a wide range of editing tasks and generalizes across benchmarks.

7. Impact and Implications

IE-Critic-R1 establishes a new standard for explainable, human-aligned quality metrics in text-driven image editing. Its design—explicitly modeling the conditional relationship between source and edited images under a prompt, yielding transparent, stepwise reasoning—marks a departure from previous black-box or single-axis approaches. The success of RLVR in optimizing for human agreement without reward hacking suggests reinforcement-trained, sequence-generating critic models may play a central role in future evaluation regimes for generative systems. Its applicability and generalizability position it as a benchmark for the next generation of learned perceptual metrics (Qu et al., 22 Nov 2025).