acceptodds
Under review as a conference paper at ICLR 2027

Trust-to-Latent: Aligning Latent Generation with Multimodal Semantic Judgment

Abstract

The central challenge of long-prompt image generation is to realize entities, attributes, quantities, and relations jointly in a semantically coherent scene. Unified understanding–generation models offer a natural opportunity to address this challenge: their understanding capability can provide a semantic criterion for the content their generation capability seeks to realize. Yet judging completed images does not automatically enable understanding to assess the noisy representations encountered during generation. We introduce Trust-to-Latent, which connects these capabilities by extending reward-trained semantic judgment to intermediate generation states. A prompt-conditioned latent evidence interface makes multi-layer generation features accessible to a reader built on a frozen understanding backbone. Image-level preferences, transferred through shared-noise latent pairs, teach the reader to assess evolving visual evidence against the prompt. The complete evaluator then remains fixed while relative preferences among candidates from a common intermediate state guide generator learning. Understanding thus informs how the generator is trained to realize instructions, without reward evaluation or candidate selection at inference. Experiments across five public benchmarks show improved overall instruction following over the base model and reward-optimization baselines, with gains in attribute binding, spatial relations, and joint satisfaction of evaluated requirements. Interface ablations and paired error analyses further support this connection between semantic judgment and compositional generation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Long-Prompt Alignment;Latent-Space Feedback;Text-to-Image Generation;Preference Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FwS2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses compositional text-to-image generation under long prompts. Its goal is to let the understanding capability of a unified understanding–generation model (BAGEL) provide semantic feedback at intermediate generation states, and to use that feedback to train the generator. On five benchmarks, the method achieves higher overall scores than BAGEL and four reward-optimization baselines.

Critical strengths
  • The problem is valuable. Using semantic judgments on intermediate generation states to guide generator training is a natural and important direction for post-training unified models. Figure 1 intuitively shows that the spatial relation is already largely determined early in denoising.
  • The experimental coverage is broad. The method achieves the best overall score on all five benchmarks. The paper also provides fine-grained analyses, including joint satisfaction rates and per-question repair/regression statistics.
  • The paper is well structured, and the method overview figure is clear and easy to follow.
Critical weaknesses
  • The source of the gains is unclear. Before latent optimization, the generator has already been aligned with the image-level reward, but the performance of this image-aligned starting checkpoint is not reported. As a result, the contribution of image-level alignment cannot be separated from that of latent feedback. In addition, Section 4.1 states that UR and LR start from the same image-aligned initialization, yet the base score in Table 6 (67.50) matches that of the original BAGEL. The starting point of each experiment should be clarified. I suggest adding stage-wise ablations that show the actual contribution of each component beyond the latent reward.
  • Comparisons with closely related work are missing. The paper cites several latent-level reward methods but does not compare against them experimentally. Moreover, some baselines (HPSv3, PickScore) mainly target aesthetics or human preference rather than instruction following, which makes them relatively weak points of comparison.
  • The annotation judges and the evaluation VLMs are not clearly specified.
  • Qualitative results are limited. The qualitative comparison consists of only three examples in Figure 3, without the UR variant or any failure cases. This makes it hard to see what latent feedback adds over image-level rewards.
Questions for authors
  • The reader is trained on t ∈ [0.3, 0.9], but during generator optimization it must score child states at roughly t ≈ 0.92–0.975. Is the reader's judgment reliable in this extrapolated range?
  • Can the method be transferred to other generation tasks, such as image editing?
  • What exactly does "Trust" in the title refer to? It does not seem to be defined anywhere in the paper.
8 comments
AuthorsOpenReview ↗

Response to [W1]


Response to [W1]: Source of the Gains and Initialization

We thank the reviewer for this helpful comment. We first clarify the starting point of each experiment, and then report a stage-wise comparison.

1. Training pipeline and starting points

  • The image-level reward stage does not optimize the generation branch. It trains only the understanding branch with the generation branch frozen, establishing the semantic judgment that later supervises latent feedback. The "image-aligned generator checkpoint" refers to the full model after this stage (hereafter BAGEL-Post-U); via shared attention, it only indirectly affects generation conditioning. All generation optimization experiments, including UR, LR, the four reward baselines, and Tables 5–6, start from BAGEL-Post-U.
  • Base score in Table 6. The 67.50 row reports the original BAGEL as a common reference (consistent with Table A1). BAGEL-Post-U scores 67.32 on the same subset, essentially on par, so the direction and ranking of all update rules remain unchanged.

2. Stage-wise comparison

We evaluate BAGEL-Post-U on all five benchmarks. UR and LR both start from it with matched budgets and inference settings, differing only in whether feedback comes from completed images (UR) or intermediate states (LR).

Table R1. Stage-wise generation performance across five benchmarks (% ↑).

Model / StageDetailMasterTIIFLongT2IUniGen++DPG
BAGEL57.1472.1379.7281.7384.51
BAGEL-Post-U (Init.)55.9671.3779.1681.6184.68
UR58.7876.7582.6685.9787.70
LR (Ours)59.9579.4983.9588.6089.18
Δ Init. − BAGEL−1.18−0.76−0.56−0.12+0.17
Δ UR − Init.+2.82+5.38+3.50+4.36+3.02
Δ LR − Init.+3.99+8.12+4.79+6.99+4.50

Δ is reported in percentage points.

The shared initialization is on par with or slightly below BAGEL, so the gains do not come from a stronger starting point. Both methods clearly improve over it, and LR outperforms UR on all five benchmarks; since everything else is matched, this additional gain is attributable to latent-level semantic feedback.

3. Contribution of each component

Under the same initialization, Tables 4–6 in the paper isolate the score readout, the interface components (multi-layer features, QKV adaptation, query memory, shared noise), and the optimization rule. Together with Table R1, they show that the image-level reward stage provides the semantic criterion rather than a stronger generator, and that the gains come from subsequent generation optimization, where latent feedback consistently improves over image-level rewards.


AuthorsOpenReview ↗

Response to [W2]


Response to [W2]: Comparisons with Closely Related Work

We thank the reviewer for the suggestion. We first explain the rationale behind our existing baselines, and then report direct comparisons with more task-specific image-level rewards and related latent-level rewards.

1. Rationale for the original baselines

  • General preference rewards (HPSv3, PickScore). These are the most widely used rewards in RL for text-to-image generation. Although trained on human preference data, both condition on the text prompt and thus account for prompt–image matching.
  • Semantic alignment rewards (UnifiedReward-2.0, VQAScore). These directly target text–image semantic alignment.

The original comparisons thus already cover both types of text–image reward signals.

2. Additional baselines

  • Image-level rewards. FGA-BLIP2 (AAAI 2026) focuses on fine-grained text–image semantic matching; LLaVA-Reward (ICCV 2025) uses a multimodal LLM to evaluate images along multiple dimensions, including text–image alignment.
  • Latent-level rewards. StitchVM (NeurIPS 2026) connects intermediate generator features to a pretrained reward model to learn value estimation on noisy states; LPO (NeurIPS 2025) uses the diffusion model itself as a noise-aware latent reward model. Since both were designed for diffusion models, we adapt them to BAGEL's intermediate generation features.

Table R2. Comparison with additional image-level and latent-level reward methods (% ↑).

MethodDetailMasterTIIFLongT2IUniGen++DPG
FGA-BLIP255.7072.5980.8381.2286.29
LLaVA-Reward57.0274.6581.3182.7085.86
StitchVM (Latent)55.5471.8779.5082.0784.66
LPO (Latent)53.0868.1378.5280.4983.21
UR58.7876.7582.6685.9787.70
LR (Ours)59.9579.4983.9588.6089.18

Trust-to-Latent achieves the best results on all five benchmarks, and even the strongest new image-level reward, LLaVA-Reward, remains clearly behind. The two latent-level methods fail to bring consistent improvements: StitchVM stays roughly on par with BAGEL, and LPO, which learns preferences from generation features without an understanding prior, even falls below it, as both struggle to judge compositional constraints such as attributes and spatial relations. In contrast, our method brings semantically supervised understanding into the evaluation of intermediate states, suggesting that reusing understanding capability is key to effective latent feedback.


AuthorsOpenReview ↗

Response to [W3] & [W4]


Response to [W3] & [W4]: Annotation Judges, Evaluators, and Qualitative Results

We thank the reviewer for these suggestions, which we address below.

1. [W3] Annotation judges and evaluation models

  • Training annotation judges. We use three multimodal models from different families to construct the image-level reward training data: Gemini 3.1 Pro (Google), Kimi K2.6 (Moonshot AI), and GLM-4.6V (Z.ai).
  • Benchmark evaluation models. Each benchmark follows its officially released evaluation pipeline, and all methods on the same benchmark share a fixed evaluation configuration (Table R3).

Table R3. Benchmark evaluation models and checkpoints.

BenchmarkEvaluation Model / PipelineModel Family / Backbone
TIIF-BenchQwen2.5-VL-72B-InstructQwen2.5-VL
DetailMasterQwen2.5-VL-7B-Instruct + YOLOE-11L + BLIP-ITMQwen2.5-VL / YOLOE / BLIP
LongT2IBenchLongT2IExpertQwen2.5-VL-7B-Instruct
UniGenBench++UniGenBench-EvalModel-qwen-72b-v1Qwen2.5-VL-72B
DPG-BenchmPLUG VQA (damo/mplug_visual-question-answering_coco_large_en)mPLUG
  • Model independence. The annotation judges belong to different model families from the test-time evaluators and share no weights or checkpoints with them; they are never invoked at test time.

2. [W4] Qualitative results

We have added a qualitative comparison to the revised manuscript; please see Figure A2 in Appendix L (page 21 of the revised PDF, appendix page 9). It includes BAGEL, UR, and LR with shared prompts and random seeds, as well as a failure case. The differences between UR and LR mainly arise in structural constraints determined early in generation:

  • (a) Object state. The teapot lid should be removed and placed in front; BAGEL and UR keep the lid on and add an extra one, while LR generates an open teapot.
  • (b) Spatial layout. UR violates "bird at the rear left, bowl at the front right"; LR follows it.
  • (c) Hierarchical structure. For a two-tier stand, UR generates three tiers with misplaced strawberries; LR keeps two tiers with macarons on top and strawberries below.
  • (d) Exact counting. For "exactly one" tulip, BAGEL and UR generate two; only LR is correct.

Such constraints are largely fixed early in denoising (Figure 1). Image-level rewards only score the completed image and can hardly attribute feedback to these early decisions, whereas LR judges intermediate states directly and corrects such errors more effectively.

Failure case (e). The prompt requires exactly two strawberries under soft, diffused daylight. BAGEL and UR generate three strawberries; LR satisfies the count but shows noticeable direct sunlight. When content and global style constraints co-occur, LR corrects object-level errors but does not fully follow the lighting style.


Reviewer_AJunOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets long-prompt text-to-image generation, where entities, attribute bindings, counts, and spatial relations must be satisfied jointly. It proposes Trust-to-Latent, which uses the understanding branch of a unified model (BAGEL) to assess intermediate noisy generation states. The understanding branch is first trained as an image-level instruction-following judge, after which its backbone weights are frozen. A prompt-conditioned interface, using multi-layer generator features, learnable queries, and low-rank adapters, is then trained to enable latent scoring, with image-level preferences transferred through pairs corrupted by shared noise. The frozen latent evaluator finally provides group-relative rewards over sibling one-step transitions for PPO-style generator updates in early denoising steps, with no reward evaluation or candidate selection needed at inference. Across five benchmarks (DetailMaster, TIIF, LongT2IBench, UniGenBench++, and DPG-Bench), the method improves overall scores over the base model and BAGEL optimized with existing image-level rewards. Improvements include spatial relations, attribute binding, and joint satisfaction of evaluated requirements. On TIIF, joint satisfaction rises from 35.5% for BAGEL to 39.1% for UR and 42.9% for LR. Interface ablations, optimization-rule comparisons, and paired error analyses provide additional support for the design.

Critical strengths

1. An interesting approach to reusing a unified model’s understanding capability during generator training. Prior work has established latent reward modeling and the transfer of image-level reward computation to noisy states. This paper explores a specific connection within a unified model: a prompt-conditioned interface enables reward-trained understanding layers to read intermediate generation features and provide feedback during denoising. I find this reuse of the model’s understanding capability particularly appealing. Shared-noise preference transfer and comparisons among sibling transitions provide a coherent implementation, with local comparisons avoiding the need to calibrate scores across noise levels. The method also requires no reward evaluation or candidate selection at inference. I consider this the most valuable aspect of the paper and a promising direction for follow-up work.

2. Consistent improvements under matched initialization and training budgets. UR and LR start from the same image-aligned generator checkpoint and use matched training budgets and inference settings. LR achieves higher overall scores on all five benchmarks, with improvements including spatial relations and attribute binding. On TIIF, it increases joint satisfaction from 39.1% for UR to 42.9%, compared with 35.5% for BAGEL. Relative to the same base generator, LR also produces more corrections and fewer regressions than UR. These results support the effectiveness of the proposed latent-feedback pipeline, although they do not isolate feedback location as the only contributing factor.

Critical weaknesses

1. The latent evaluator’s reliability and the specific contribution of reward-trained understanding are not directly established. Table 4 evaluates image-level score readout, while Table 5 measures downstream generation performance rather than latent ranking accuracy. All interface ablations retain the understanding layers, with the implementation using layers 25–28. There is no comparison with a latent reward head trained on the same generator features and shared-noise preference data without these understanding layers. It therefore remains unclear how much of the improvement specifically depends on the understanding-based reader, rather than on generator features combined with a learned latent scorer. Moreover, no pairwise accuracy across noise levels is reported. This is particularly relevant because, according to the reported schedule, the earliest three child-scoring times are approximately 0.975, 0.947, and 0.917, exceeding the evaluator’s preference-training range of [0.3, 0.9]. This is my most consequential concern because it bears directly on the paper’s central contribution.

2. The reliability and scope of the reported gains need further clarification. The generation benchmark evaluations use a single generation seed, 43, without uncertainty estimates or reported variability across repeated training runs. The multimodal judges used for training annotations and the exact models/checkpoints used by the local question-answering evaluation adapters are not clearly specified, making it difficult to assess the independence of the training and evaluation signals. In addition, although the annotation target includes a visual-quality term with coefficient 0.2, no independent visual-quality or aesthetic evaluation is reported. DetailMaster style decreases from 85.6 to 66.4, and four negation-related TIIF categories regress. These results reveal category-specific trade-offs; the style decline indicates reduced style adherence, while its relationship to overall visual quality remains unmeasured.

Questions for authors

Questions that could change my assessment

  1. Understanding-free latent reward baseline. Could you train a lightweight latent reward head, such as an MLP or small transformer, on the same generator features (layers 4/8/12), using the same 26,873 shared-noise preference pairs, prompt/time conditioning, and generator optimization recipe, but without the understanding layers? This would help establish the added value of the understanding-based reader.

  2. Evaluator accuracy across noise levels. What is the evaluator’s pairwise accuracy on held-out prompt/image pairs as a function of noise level? In particular, the earliest three child-scoring times are approximately 0.975, 0.947, and 0.917, outside the preference-training range of [0.3, 0.9]. Is its ranking reliable there, and how much do updates at these earliest parent positions contribute to the gains?

  3. Judges and evaluators. Which models serve as the multimodal judges for training annotations, and which model/checkpoint is used by each local evaluation adapter, especially TIIF? Do any annotation and evaluation models share weights or belong to the same model family? Were the 12,766 training prompts checked for overlap with the benchmark prompts?

  4. Image quality. Could you report an independent visual-quality assessment for BAGEL, UR, and LR, for example a small human study that distinguishes visual quality from instruction following? Could you also explain the DetailMaster style decline from 85.6 to 66.4?

  5. Variance. The benchmark evaluations use generation seed 43. Could you report results across several generation seeds, at least for the UR versus LR comparison on the TIIF subset? Where feasible, variability across independent training runs would also help establish robustness.

  6. Optimization choices. In Table 6, the pairwise variant uses coordinate-summed log odds, whereas the group-relative variants use coordinate-normalized policy surrogates. Could you test coordinate-averaged log odds or retune the inverse temperature to assess whether objective scaling contributes to the pairwise variant’s degradation? Also, since reward-gradient distillation performs best in this design study, was it tested under the final training recipe, and why was group-relative optimization selected?

  7. Compute. What are the update counts and GPU-hours for final LR training compared with UR, and how are the matched training budgets defined?

Suggestions for future work (not required for my assessment)

  • Exploring the method’s applicability to other unified models, such as Janus-Pro and Show-o2, to test generalization beyond BAGEL.
  • Quantifying the Figure 1 observation across all 16 trajectories, or a larger set, to establish how often semantic errors are already observable early in denoising.
11 comments
AuthorsOpenReview ↗

Response to [W1] & [Q1]


Response to [W1] & [Q1]: Understanding-Free Latent Reward Baseline

We thank the reviewer for this constructive suggestion. We first clarify how the requested baseline relates to our original ablations, and then report the controlled comparison.

1. Relation to the original ablations

Our method reuses the semantic judgment of the understanding branch and extends it to intermediate generation states through a latent evidence interface. The original ablations (Table 5) therefore examine how to leverage this understanding capability, e.g., multi-layer features, QKV adaptation, and query memory. The requested baseline complements them by quantifying how much the understanding reader adds over a generic latent scorer learned directly from generation features.

2. Controlled experiment

  • Baselines. We add two latent reward scorers without pretrained understanding layers: an MLP, and a Small Transformer that has the same four-layer depth, learnable queries, text memory, and output head as our reader, but learns latent preference scoring from scratch instead of using pretrained understanding weights.
  • Shared setup. All scorers use the same generator features from layers 4/8/12, prompt/time conditioning, 26,873 shared-noise preference pairs, and noise sampling, and all retain spatial information. Each evaluator is then frozen, and the generator is optimized from the same checkpoint with identical sibling sampling, group-relative optimization, and matched training budgets. The only difference is whether the scorer reuses understanding computation through the latent evidence interface or learns scoring directly from latent preferences.
  • Evaluation. We measure (1) pairwise ranking accuracy (PairAcc) on the same held-out latent preference pairs, which reflects preference judgment; and (2) TIIF after generator optimization, which reflects whether the feedback translates into instruction-following gains.

Table R1. Comparison of understanding-free latent scorers and our understanding-based reader.

Latent Reward ScorerPairAcc (%) ↑TIIF Short (%) ↑TIIF Long (%) ↑TIIF All (%) ↑
MLP (w/o Understanding)55.3968.9169.4569.18
Small Transformer (w/o Understanding)65.3772.6273.0472.83
Trust-to-Latent (Ours)74.5979.5079.4979.49

PairAcc timesteps are sampled from the preference-training time distribution.

Given the same generation features and preference supervision, the understanding-based reader provides clearly more accurate preference judgments and more effective generation feedback, and the three scorers rank identically on PairAcc and TIIF. Notably, the MLP scorer even lowers TIIF below the original BAGEL (72.13), showing that a weak scorer learned only from generation features can mislead optimization. This directly establishes the added value of the understanding-based reader: a latent scorer learned on the same generator features brings at most marginal gains over BAGEL, whereas reusing reward-trained understanding turns the same features and supervision into substantially more accurate preference judgments and clearly stronger instruction following.


AuthorsOpenReview ↗

Response to [W1] & [Q2]


Response to [W1] & [Q2]: Evaluator Accuracy Across Noise Levels

We thank the reviewer for the attention to ranking reliability at high noise levels. Since our latent evaluator only compares sibling states under the same parent state and timestep, it does not require absolute reward calibration across noise levels; it must, however, provide effective local rankings at the timesteps where it is actually used, especially at high-noise positions beyond the preference-training range. We first explain the design motivation, and then examine both the ranking accuracy across noise levels and the actual training contribution of the earliest high-noise updates.

1. Design motivation

  • Why preference training is restricted to t≤0.9t\le0.9. Shared-noise preference pairs satisfy zt+−zt−=(1−t)(z0+−z0−)z_t^{+}-z_t^{-}=(1-t)(z_0^{+}-z_0^{-}): as tt approaches pure noise, the distinguishable difference between candidates diminishes, making it increasingly difficult to learn effective signals from final-image preferences. We therefore concentrate preference training on t≤0.9t\le0.9, where intermediate states carry relatively sufficient semantic evidence.
  • Why ranking signals can persist beyond this range. The reader does not take ztz_t itself as input, but the multi-layer features computed by the generator on (zt,c,t)(z_t,c,t), which already encode the generator's prediction of the final content. As shown in Figure 1, at t=0.938t=0.938 the clean prediction already reveals the spatial-relation error in the final image, suggesting that the evaluator may retain ranking signals at even higher noise levels.
  • Why early feedback matters. Global layout, attribute binding, and object relations take shape early in generation; introducing semantic feedback before these structural decisions are fixed offers a greater opportunity to steer the subsequent trajectory.

2. Ranking accuracy across noise levels

Using the original frozen latent evaluator, we corrupt held-out image preference pairs with shared noise to each online scoring timestep, including the three beyond t=0.9t=0.9 (0.917, 0.947, and 0.975), and measure pairwise ranking accuracy (PairAcc).

Table R2. Latent pairwise ranking accuracy at online scoring timesteps.

Scoring Time tt0.6920.7500.8000.8440.8820.917*0.947*0.975*Average
PairAcc (%) ↑75.376.674.873.371.668.866.765.271.5

Timesteps marked with an asterisk are extrapolated beyond the preference-training range (t≤0.9t\le0.9). Average is the unweighted mean across timesteps.

Ranking accuracy generally decreases as the noise level increases, yet at all three scoring positions beyond the training range it remains well above the 50% chance level. Although semantic judgment is harder on high-noise states, the evaluator thus retains effective relative ranking signals at noise levels not directly covered by preference training.

3. Contribution of the earliest high-noise updates

Keeping the evaluator, generator initialization, and other optimization settings unchanged, we remove the reward-driven updates at the three scoring timesteps beyond t=0.9t=0.9 while maintaining the same total number of updates, and compare this variant with the full LR, with BAGEL and UR listed for reference.

Table R3. Contribution of early high-noise reward updates on TIIF.

MethodTIIF Short (%) ↑TIIF Long (%) ↑TIIF All (%) ↑
BAGEL71.3972.8672.13
Trust-to-Latent (UR)76.5176.9976.75
LR w/o Earliest 3 Updates77.9078.1078.00
Full LR (Ours)79.5079.4979.49

Removing these updates consistently lowers TIIF on both short and long prompts, so despite their lower ranking accuracy, these positions still provide additional gains on top of the remaining latent-feedback updates. This is because the group-relative advantage depends only on the ranking among siblings: as long as the ranking is better than random, the expected update direction remains aligned with the semantic preference, and ranking noise mainly weakens the signal strength, while the clipped policy update and reference KL further limit the magnitude of each update. Meanwhile, the ablated variant still outperforms UR, indicating that latent feedback at the other timesteps is also effective.

Together, these results show that the latent evaluator retains effective semantic preference ranking at extrapolated high-noise positions, and exploiting local feedback at these positions further improves final compositional instruction following.


AuthorsOpenReview ↗

Response to [W2] & [Q3]


Response to [W2] & [Q3]: Annotation Judges, Evaluation Models, and Train–Test Independence

We thank the reviewer for the suggestion regarding the transparency of training annotation and evaluation protocols. Below, we specify the annotation judges used during training and the evaluation models for each benchmark, and then clarify the independence between training and evaluation in both models and data.

1. Training annotation judges

We employ three multimodal models from different model families as annotation judges to construct the image-level reward training data: Gemini 3.1 Pro (Google), Kimi K2.6 (Moonshot AI), and GLM-4.6V (Z.ai).

2. Benchmark evaluation models

Each benchmark follows its officially released evaluation pipeline and the corresponding evaluator, and all methods on the same benchmark share a unified, fixed evaluation configuration. In particular, TIIF, which the reviewer specifically asked about, uses Qwen2.5-VL-72B-Instruct as the evaluation model. The evaluation models and their model families for all benchmarks are listed in Table R4.

Table R4. Benchmark evaluation models and checkpoints.

BenchmarkEvaluation Model / PipelineModel Family / Backbone
TIIF-BenchQwen2.5-VL-72B-InstructQwen2.5-VL
DetailMasterQwen2.5-VL-7B-Instruct + YOLOE-11L + BLIP-ITMQwen2.5-VL / YOLOE / BLIP
LongT2IBenchLongT2IExpertQwen2.5-VL-7B-Instruct
UniGenBench++UniGenBench-EvalModel-qwen-72b-v1Qwen2.5-VL-72B
DPG-BenchmPLUG VQA (damo/mplug_visual-question-answering_coco_large_en)mPLUG

3. Independence between training and evaluation

  • Model independence. The Gemini, Kimi, and GLM models used for training annotation belong to different model families from the Qwen2.5-VL, YOLOE, BLIP, and mPLUG evaluators used at test time, and share no model weights or checkpoints with them. The training annotation judges are never invoked at test time, nor are their annotations used in evaluation. Moreover, all compared methods use the same benchmark evaluation models and scoring protocols, ensuring consistent and fair comparisons.
  • Data independence. The 12,766 prompts used for image-level reward training are all independently constructed and are not drawn from any benchmark test set. We further performed an exact, prompt-by-prompt matching check between the training prompts and all test prompts of the five benchmarks, and found no duplicates.

Therefore, the training and evaluation signals are independent in both models and data, and the reported gains do not stem from shared judges or overlapping prompts.


Reviewer_9gzUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Trust-to-Latent, a framework that uses reward-trained multimodal understanding to guide intermediate image generation. A prompt-conditioned interface exposes multi-layer generation features to a reader built on a frozen understanding backbone. The latent evaluator is trained using image-preference pairs corrupted with shared noise and is then frozen during online generator optimization, where it compares sibling transitions from a common parent state. Experiments on five instruction-following benchmarks report improvements in compositional alignment, with interface ablations and paired analyses of constraint satisfaction. The trained generator requires no reward evaluation or candidate selection at inference.

Critical strengths
  1. A well-motivated connection between semantic understanding and intermediate generation. The learned evidence interface offers a concrete way to reuse multimodal semantic judgment during generator training. The resulting model does not require additional reward evaluation or candidate selection at inference, which makes the approach practically appealing.

  2. Consistent empirical gains with useful diagnostic analyses. The LR variant improves all five reported overall benchmark scores over the base generator and the UR variant. On TIIF, accuracy increases from 72.13 to 79.49, while joint constraint satisfaction increases from 35.54% to 42.93%. The interface ablations and repair/regression analyses provide additional evidence that the design helps with compositional instruction following.

Critical weaknesses
  1. The transfer from synthetically noised images to online generation states is not directly validated. The latent evaluator is trained on completed images corrupted with noise, but is used to rank intermediate states produced by the current generator. These distributions may differ even at the same timestep. The reported downstream improvements are encouraging, but the paper does not directly report ranking accuracy on actual generation trajectories or quantify the gap relative to synthetically noised states. This leaves uncertainty about a central assumption behind the online feedback mechanism.

  2. The specific benefit of a unified understanding-generation backbone is not sufficiently isolated. The framework is motivated by unification, but it remains unclear which components require a unified model and how much benefit comes specifically from this choice. A controlled comparison with a separate generator and multimodal reward model using a comparable latent interface would make the contribution of unification clearer.

Questions for authors
  1. Generalization from synthetically noised images to actual generation states.

In Stage 2, the latent evaluator is trained on states obtained by adding noise to completed images, whereas in Stage 3 it evaluates intermediate states actually produced by the current generator. These two state distributions may differ even at the same timestep. Have you validated the latent evaluator's ranking accuracy on actual generation trajectories? How does its ranking performance on these states compare with its performance on synthetically noised image states?

  1. The role of the unified multimodal model (UMM).

It is unclear to me why a unified understanding-generation model is necessary for this framework. Which components specifically rely on unification, and could a separate generator paired with a multimodal reward model support the same latent-evaluation interface? Could you provide a controlled comparison or ablation demonstrating the benefit of using a unified model?

  1. Benchmark coverage.

Could you explain the rationale for the benchmark selection and why results on WISE are not reported? Could you provide results on WISE to help assess whether the observed improvements generalize beyond the five benchmarks evaluated in the paper?

4 comments
AuthorsOpenReview ↗

Response to [W1] & [Q1]


Response to [W1] & [Q1]: Generalization from Synthetically Noised Images to Actual Generation States

We thank the reviewer for this important question. We first explain why synthetic noising is a reasonable proxy, and then directly validate the transfer on actual generation trajectories.

1. Why synthetically noised states are a reasonable proxy

  • Consistent with the generative process. We construct noisy states as zt=(1−t)z0+tϵz_t=(1-t)z_0+t\epsilon (Eq. 4), the same conditional noising path used in Flow Matching training. Since the generator is trained so that its marginal distribution at each time tt matches the one induced by this path, synthetically noised states are a principled approximation of intermediate generation states rather than arbitrary perturbations. LPO (NeurIPS 2025) and StitchVM (NeurIPS 2026) adopt similar constructions to learn noisy-state rewards.
  • Online optimization only relies on local rankings. The evaluator does not need to predict absolute values comparable across timesteps or prompts. Online, it only compares sibling states that share the same prompt, parent state, and timestep, and advantages are normalized within each group (Eq. 5). Even if the gap between the two state distributions shifts all sibling scores in common, this shift is canceled by the normalization; what matters for optimization is whether the sibling ranking is preserved.

2. Experiment: validation on actual generation trajectories

We take 1,000 TIIF prompts and use the generator employed in online optimization to produce actual parent states. At each child timestep tt in Table R1, we sample 6 sibling child states from the same parent with the stochastic transition used in training, and record their latent rewards. Each child is then completed with a fixed ODE sampler, and TIIF constraint satisfaction of the final image serves as the preference label. As a control, we re-noise the same final images to the same tt with shared noise (Eq. 4) to obtain synthetic states. The two settings share prompts, final images, and labels, and differ only in the origin of the intermediate states, so the difference in pairwise ranking accuracy (PairAcc) directly measures the transfer gap.

Table R1. Pairwise ranking accuracy on synthetic and actual generation states.

Time ttSynthetic PairAcc (%) ↑Actual PairAcc (%) ↑Δ (pp)
0.88271.869.5−2.3
0.84474.073.3−0.7
0.80073.674.2+0.6
0.75075.476.1+0.7
0.69277.376.5−0.8
Average74.473.9−0.5

Δ = Actual − Synthetic PairAcc in percentage points (pp); Average is the unweighted mean across timesteps.

Across all evaluated timesteps, the evaluator achieves ranking accuracy on actual generation states comparable to that on synthetic states (73.9% vs. 74.4% on average), well above the 50% chance level. The per-timestep gaps are small and of mixed sign, showing no systematic degradation, and accuracy on actual states follows the same trend as on synthetic states, increasing as the noise level decreases. These results show that the preferences learned from synthetically noised supervision transfer to the sibling states actually produced by the generator, so the evaluator provides reliable local feedback for online optimization.


AuthorsOpenReview ↗

Response to [W2] & [Q2]


Response to [W2] & [Q2]: The Role of the Unified Multimodal Model (UMM)

We thank the reviewer for this question. Our goal is to turn the existing multimodal understanding of a UMM into semantic feedback during generation. We first explain why the unified architecture is well suited to this goal, and then report a controlled comparison against a separate reward backbone.

1. Why a unified model

  • Reusing the model's own understanding rather than learning judgment from scratch. Through large-scale multimodal pretraining, the understanding branch of a UMM already holds rich vision–language knowledge, and it interacts with generation representations via shared attention, so it has already been exposed to the generator's own visual representations. We first establish an instruction-satisfaction criterion via image-level reward training and then apply it to intermediate states. Compared with training a latent reward model from scratch, learning can thus focus on reading visual evidence during generation rather than relearning a complete semantic criterion.
  • Native information reuse between understanding and generation. In BAGEL, both capabilities share multimodal pretraining and a compatible Transformer architecture. Prompt-conditioned queries jointly access text memory and multi-layer generation evidence, and the understanding layers relate textual requirements to the emerging visual content, so the model judges attributes, spatial relations, and compositional constraints with its existing understanding computation rather than merely mapping features to a scalar. A separate generator plus an external reward model can realize a similar mechanism, but requires an additional reward backbone and cross-model representation adaptation without shared pretraining; a UMM only needs a lightweight evidence interface.

2. Controlled comparison

We compare three alternative readers under identical conditions: an MLP and a Small Transformer without pretrained understanding layers, and a UnifiedReward-2.0 Latent Reader built on a separate multimodal reward backbone. The latter corresponds to the reviewer's "separate generator + multimodal reward model" setting: it is reward-trained with the same image-level annotations as ours and reads BAGEL's generation features through the same latent evidence interface. All readers use the same multi-layer BAGEL features and 26,873 shared-noise preference pairs, and are then frozen to optimize the generator from the same initialization with identical sibling sampling, group-relative optimization, and 1,000 updates. PairAcc is evaluated on the same held-out latent preference pairs.

Table R2. Comparison of latent reward reader architectures.

Latent Reward ReaderPairAcc (%) ↑TIIF Short (%) ↑TIIF Long (%) ↑TIIF All (%) ↑
MLP (w/o Understanding)55.3968.9169.4569.18
Small Transformer (w/o Understanding)65.3772.6273.0472.83
UnifiedReward-2.0 Latent Reader68.5473.2573.8973.57
Trust-to-Latent (Ours)74.5979.5079.4979.49

The reader built on BAGEL's own understanding branch performs best in both ranking accuracy and downstream generation, and the four readers rank identically on PairAcc and TIIF, suggesting that better semantic judgment translates into better generation. Under identical interface, supervision, and optimization settings, the separate reward backbone can be plugged into our framework but performs clearly worse. Reusing understanding within a UMM thus lets the reader judge intermediate evidence more accurately and turn it into more effective feedback, making the unified model a more effective path for bringing understanding into generator training.


AuthorsOpenReview ↗

Response to [Q3]


Response to [Q3]: Benchmark Coverage and Generalization to WISE

We thank the reviewer for the suggestion regarding evaluation coverage. We first explain our benchmark selection, and then report additional results on WISE.

1. Rationale for benchmark selection

  • Research focus. Our work targets complex compositional instruction following under long prompts, i.e., the correct binding and joint satisfaction of multiple semantic constraints such as entities, attributes, quantities, and spatial relations.
  • Selected benchmarks. We therefore select five benchmarks that evaluate this capability from complementary perspectives:
    • DetailMaster and LongT2IBench: complete realization of entities, attributes, and relations in long prompts;
    • TIIF: satisfaction of fine-grained instructions and compound constraints;
    • UniGenBench++: semantic test points under long prompts;
    • DPG-Bench: overall alignment with dense prompts.
  • Why WISE was not in the main evaluation. WISE targets world-knowledge-informed generation, requiring models to leverage cultural, spatio-temporal, and scientific knowledge to infer and depict content implied by the prompt. Its core challenge lies in knowledge reasoning, whereas our work focuses on the accurate realization of explicit semantic constraints. We thus prioritized the five benchmarks more directly related to our research question.

2. Additional evaluation on WISE

We agree that WISE provides a valuable complementary evaluation, and evaluate all methods on it, including the reward optimization baselines.

Table R3. Generalization performance on WISE.

MethodWISE (%) ↑
BAGEL36.4
HPSv333.4
PickScore33.9
UnifiedReward-2.036.5
VQAScore37.0
Trust-to-Latent (UR)39.1
Trust-to-Latent (LR)40.0

Evaluated on WISE Verified; all methods share the same evaluation protocol and generation settings.

LR achieves the best WISE score, outperforming BAGEL and all external reward baselines, whereas preference-oriented rewards such as HPSv3 and PickScore even fall below the base model. Notably, we design no dedicated module or objective for the implicit world knowledge targeted by WISE. These results suggest that the gains in semantic instruction following brought by Trust-to-Latent are not limited to the five compositional benchmarks, but also extend to knowledge-driven generation.


AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer 9gzU,

Thank you again for your time and constructive comments on our paper.

In our response, we addressed your comments as follows:

  1. Transfer from synthetically noised states to actual generation states: We directly validated the latent evaluator's ranking accuracy on actual generation trajectories. Average accuracy is 73.9% on actual states and 74.4% on synthetic states, which is comparable, and no timestep shows a systematic drop.
  2. Role of the unified model: We compared our reader with a UnifiedReward-2.0 Latent Reader built on a separate reward backbone, using the same interface, supervision, and optimization settings. Reusing the UMM's own understanding performs clearly better in both ranking accuracy and downstream generation.
  3. Benchmark coverage: We explained why we selected the five benchmarks and added results on WISE. Our method (LR) achieves the best WISE score, so its gains are not limited to compositional instruction following.

As the discussion period is coming to an end, we would greatly appreciate any further feedback. We hope our responses have fully addressed your concerns. If you have any remaining questions before the discussion closes, we would be glad to clarify or provide more details.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.