Reinforcement learning (RL) can improve the prompt following capability of text-to-image (T2I) models, yet obtaining high-quality reward signals remains challenging: CLIP Score is too coarse-grained, while VLM-based reward models (e.g., RewardDance) require costly human-annotated preference data and additional fine-tuning. We propose PromptEcho, a reward construction method that requires no annotation and no reward model training. Given a generated image and a guiding query, PromptEcho computes the token-level cross-entropy loss of a frozen VLM with the original prompt as the label, directly extracting the image-text alignment knowledge encoded during VLM pretraining. The reward is deterministic, computationally efficient, and improves automatically as stronger open-source VLMs become available. For evaluation, we develop DenseAlignBench, a benchmark of concept-rich dense captions for rigorously testing prompt following capability. Experimental results on two state-of-the-art T2I models (Z-Image and QwenImage-2512) demonstrate that PromptEcho achieves substantial improvements on DenseAlignBench (+26.8pp / +16.2pp net win rate), along with consistent gains on GenEval, DPG-Bench, and TIIFBench without any task-specific training. Ablation studies confirm that PromptEcho comprehensively outperforms inference-based scoring with the same VLM, and that reward quality scales with VLM size. We will open-source the trained models and the DenseAlignBench.
Figures & tables
Figure 1: Overview of the PromptEcho training pipeline. (Left, dashed green) Data Preparation : a frozen VLM captions images from a source dataset to build the caption corpus {c1,…,cN} . (Main loop) At each iteration, a caption ci is sampled and fed to the T2I model Gθ , which generates K images. Each image xj enters the PromptEcho Reward Computation (blue box): the image, a fixed query q , and the original caption c are concatenated as input tokens to the same frozen VLM; a single teacher-forcing forward pass produces logits over all positions; the cross-entropy loss is computed only on caption token positions, yielding the reward Rj=−CE(c∣xj,q) . Group advantage A^j=Rj−mean(R1:K) is then used by AWM/GRPO to update Gθ via policy gradient. The same frozen VLM (e.g. Qwen3-VL-32B) is shared across data preparation and reward computation, requiring no fine-tuning or human annotation.
Model
Win Rate
Baseline Win Rate
Tie Rate
Net Advantage ↑
Z-Image: + PromptEcho vs Baseline
61.5%
34.7%
3.8%
+26.8pp
QwenImage-2512: + PromptEcho vs Baseline
53.3%
37.0%
9.7%
+16.2pp
Table 1: DenseAlignBench pairwise evaluation results. Pairwise comparison on the prompt following using Gemini-3-flash-preview.
Model
Single Obj.
Two Obj.
Counting
Colors
Position
Attr. Bind.
Overall ↑
Z-Image (Baseline)
0.99
0.91
0.75
0.86
0.41
0.59
0.75
+ PromptEcho
1.00
0.94
0.85
0.86
0.52
0.73
0.82
QwenImage-2512 (Baseline)
0.99
0.94
0.56
0.87
0.47
0.64
0.74
+ PromptEcho
0.99
0.93
0.68
0.90
0.55
0.70
0.79
Table 2: GenEval evaluation results. Evaluating T2I models’ prompt following capability on structured sub-tasks.
Table 4: TIIFBench evaluation results. Each cell contains two values corresponding to short/long description prompts.
Model
GenEval
DPG-Bench
TIIFBench
Single
Two
Count
Color
Pos.
Attr.
Overall ↑
Global
Entity
Attr.
Rel.
Other
Overall ↑
Short ↑
Long ↑
Baseline
0.99
0.91
0.75
0.86
0.41
0.59
0.75
91.60
91.54
90.32
92.76
91.94
86.90
84.91
83.16
+ PromptEcho
1.00
0.94
0.85
0.86
0.52
0.73
0.82
93.05
92.76
91.87
93.89
89.99
87.92
88.50
88.94
+ InferScore
1.00
0.91
0.79
0.86
0.41
0.59
0.76
78.74
93.17
90.21
92.56
82.49
86.86
82.40
84.64
Table 5: Ablation: PromptEcho vs InferScore on public benchmarks (Z-Image + Qwen3-VL-32B). GenEval reports per-subtask and overall scores; DPG-Bench reports per-category and overall scores; TIIFBench reports overall short/long scores.
Model
Win Rate
Baseline Win Rate
Tie Rate
Net Advantage ↑
+ PromptEcho vs Baseline
61.5%
34.7%
3.8%
+26.8pp
+ InferScore vs Baseline
28.0%
31.3%
40.8%
− 3.3pp
Table 6: Ablation: PromptEcho vs InferScore on DenseAlignBench (Z-Image + Qwen3-VL-32B).
Model
Win Rate
Baseline Win Rate
Tie Rate
Net Advantage ↑
+ PromptEcho (32B Reward) vs Baseline
61.5%
34.7%
3.8%
+26.8pp
+ PromptEcho (8B Reward) vs Baseline
57.6%
39.3%
3.2%
+18.3pp
Table 7: DenseAlignBench reward model size ablation: 32B vs 8B VLM, both based on Z-Image.
Model
GenEval ↑
DPG ↑
TIIF-S ↑
TIIF-L ↑
Z-Image (Baseline)
0.75
86.90
84.91
83.16
+ PromptEcho (32B Reward)
0.82
87.92
88.50
88.94
+ PromptEcho (8B Reward)
0.77
87.15
89.08
87.69
Table 8: Public benchmark reward model size ablation: 32B vs 8B VLM, both based on Z-Image.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Qualitative comparison of 6 representative winning cases. Each row shows the full Prompt , Baseline (QwenImage-2512), and Ours (PromptEcho-trained). Ours demonstrates more faithful adherence to fine-grained details including color, count, texture, and spatial relations.
Figure 3: Text rendering quality comparison before and after PromptEcho training (Section 4.4 ). Each row shows the Input Image , Prompt , Before PromptEcho , and After PromptEcho . After training, the model produces more accurate Chinese characters with improved stroke completeness and layout consistency in e-commerce poster generation.