acceptodds
Under review as a conference paper at ICLR 2027

Prompt Reinjection for Spatial Reasoning in MMDiTs: Signal Amount, Depth Allocation, and Content-Specific Repair

Abstract

Multimodal diffusion transformers (MMDiTs) forget prompt semantics as depth grows, and spatial relations suffer most. Prompt Reinjection (PR) counters this by repairing every text token at every layer, but neither the number of repaired positions nor the choice of layers has been examined; we therefore treat PR as a probe of its own redundancy. Repair saturates early: eight real positions match full-scale repair, and the injection weight, not the token count, controls the response; past a bounded window, amplification collapses performance on SD3. Most repaired positions are padding, which joint attention does not mask and which carries no signal. Depth adds a secondary trend at the default weight: with span-matched controls, dense repair over the early half matches full PR while late-only repair is no better than baseline, and several allocations perform comparably. As one instantiation, an early-full/late-sparse schedule (Layer-Adaptive Sparse Prompt Reinjection, LASP) reaches 40.1% on GenEval Position, matches full PR with a third of the token-layer pairs, and remains robust where full PR collapses; we make no optimality claim. The injected content matters—noise, a permutation, or a constant removes the gain—and the pattern transfers to T2I-CompBench and directionally to FLUX.1-Schnell.

Primary area:
generative models
Keywords:
Prompt Reinjection, Multimodal Diffusion Transformers, Spatial Reasoning, Training-free Intervention

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_4mzKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates training-free Prompt Reinjection (PR) for improving spatial-relation generation in text-to-image MMDiTs. It systematically analyzes the effects of repaired token count, layer span, injection weight, and injected feature content. The findings show that repairing a small subset of tokens can match full reinjection, while excessive injection degrades performance and effective repair is concentrated in early-to-middle layers. Based on these observations, the paper proposes Layer-Adaptive Sparse Prompt Reinjection (LASP), which combines dense early-layer repair with sparse late-layer repair to reduce injection cost and improve robustness. Experiments are primarily conducted on SD3-medium, with additional evaluations on FLUX.1-Schnell and T2I-CompBench spatial subsets.

Critical strengths
  1. Systematic analysis of PR's effectiveness and limitations. The paper shows that sparse repair can match full reinjection, while padding repair and feature substitution offer little benefit. It also identifies performance degradation from excessive injection, providing a more detailed analysis of PR on spatial tasks.

  2. A practical application of the findings. LASP achieves performance comparable to full PR with approximately 66% fewer repaired token-layer pairs. Although it offers no inference speedup or clear advantage over volume-matched alternatives, it demonstrates how the findings can guide repair design.

Critical weaknesses
  1. Limited novelty over the original PR method. The original PR[1] already explored layer spans, injection weights, and degradation at higher weights, making the claim that these dimensions were unexplored inaccurate. The main novelty lies in finer-grained analyses of token sparsity, padding, and feature content. LASP shows no clear advantage over full PR or volume-matched alternatives, nor does it improve inference latency, limiting its methodological contribution.

  2. Spatial relation gains are not clearly disentangled from object generation improvements. In the 8-token experiment, object detection improves from 78.9% to 88.7%, while conditional relation accuracy increases only from 37.9% to 40.4%. This suggests that much of the gain comes from better object generation. The main LASP result (40.1%) lacks a similar breakdown, and improvements are concentrated on above/below, with limited gains on left/right.

[1] Yao et al. (2026). Prompt Reinjection: Alleviating Prompt Forgetting in Multimodal Diffusion Transformers.

Questions for authors
  1. Distinction from the original PR. The original PR paper (Yao et al., 2026) already investigates target-layer coverage (Table 4) and injection weights (Tables 10–11), including performance degradation at higher weights. Thus, the claim in Section 2.1 that these dimensions were previously unexplored appears inaccurate.

  2. Duplicated author name in the citation context. The paper contains duplicated author citations, such as “Spingarn et al. Spingarn et al. (2026)” in Section 2.1. These should be corrected.

Reviewer_ai2COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper analyzes the redundancy and operating regime of Prompt Reinjection (PR) in multimodal diffusion transformers. It varies the number of repaired text positions, the allocation of repair across layers, and the injection weight. The experiments suggest that a small subset of real-token positions captures much of full PR’s spatial-generation benefit, while excessive injection can substantially degrade SD3 performance. The paper also introduces LASP, an early-dense/late-sparse schedule, and evaluates transfer to T2I-CompBench and FLUX.1-Schnell.

Critical strengths
  1. This paper addresses a clearly motivated question. Investigating whether full-token, full-depth reinjection is necessary is worthwhile.
  2. The experiments are designed well and the paper introduces them in a paced flow.
Critical weaknesses
  1. the practical use of this paper remains unclear. Although this paper shows that not all token positions' injection is necessary, the LASP introduced in this paper does not produce faster inference than the full PR as suggested in the experiments.
  2. There lacks an experiment for the early-dense/late-sparse design of LASP. An early-sparse/late-dense experiment is needed.
Questions for authors
  1. Given the absence of latency or memory savings, what practical advantage should motivate choosing LASP over a properly tuned full-PR baseline?
  2. Could you compare early-dense/late-sparse and early-sparse/late-dense schedules under matched layer spans, token-layer budgets, weights, and selection policies?
2 comments

Thank you for the additional experiments. The matched reversed-schedule comparison addresses my second concern and supports the early-dense/late-sparse design at higher injection weights.

However, the practical advantage over a properly tuned full-PR baseline remains unclear. The preservation gains at matched weights are useful, but do not establish a better accuracy–preservation trade-off when each method is tuned independently. With no latency or memory savings, this concern remains material to my assessment. I therefore maintain my score.

Thank you for noting that the matched reversed-schedule control addresses your second concern. The remaining question is whether the accuracy–preservation trade-off survives when each schedule is tuned independently. We measured two operating points that bear on it after your comment of 18 November. Over 10 seeds the base model generates 299 of the 1000 images correctly. For each repaired model we count how many of those 299 it keeps correct and how many it makes incorrect. The two columns sum to 299.

Operating pointAccuracy (%)Correct outputs keptCorrect outputs made incorrect
Full PR, w = 0.0538.5249 / 29950
LASP, w = 0.0538.1261 / 29938
LASP, w = 0.0638.6255 / 29944
LASP, w = 0.0739.4250 / 29949
LASP, w = 0.1040.1228 / 29971

Both new points sit above full PR's best point on both axes though the preservation margin at w = 0.07 is a single image. At w = 0.06 LASP reaches 38.6% accuracy and keeps 255 of the 299 correct outputs against 38.5% and 249 for full PR so the count made incorrect falls from 50 to 44. At w = 0.07 accuracy rises to 39.4% while the count kept stays level with full PR at 250.

The picture is not uniform across models. SD3.5-Large at the official preset gives 32.5% versus 31.0% (+1.5pp, p = 0.188) and on FLUX.1-Schnell at the default weight the difference reverses to −0.4pp.

Two qualifications. We added w = 0.06 and w = 0.07 after seeing full PR's accuracy, so the choice of weight is post hoc. With 100 prompts the accuracy interval is about ±3pp and a six-image difference in preservation is within the same stability band, we describe the direction as consistent rather than the size as precise.

Reviewer_v3EvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The paper investigates Prompt Reinjection (PR), which addresses the forgetting of prompt semantics as the depth of Multimodal Diffusion Transformers increases. Specifically, it examines the number of repaired token positions, injection weight, and repair location. The findings show that repairing a small number of real token positions can match full PR, while the injection weight strongly affects performance. The authors also find that early-layer repair is more effective than late-only repair. Based on these findings, the paper introduces Layer-Adaptive Sparse Prompt Reinjection (LASP), an early-full/late-sparse repair schedule that matches full PR on GenEval Position with fewer repaired token-layer pairs. The authors further evaluate the findings on T2I-CompBench and FLUX.1-Schnell.

Critical strengths
  • The paper investigates the effectiveness of Prompt Reinjection in addressing prompt forgetting in Multimodal Diffusion Transformers.
  • The paper provides appropriate background and technical details to help readers understand the method and the questions the authors investigate.
  • The experiments examine key aspects of PR, including token selection, injection weight, and repair location. These findings motivate LASP, which matches full PR on GenEval Position with fewer repaired token-layer pairs and remains more robust at higher injection weights on SD3.
  • The LASP experiments demonstrate its usefulness on GenEval, the main evaluation benchmark, and provide additional evidence of its effectiveness on the spatial subsets of T2I-CompBench.
Critical weaknesses
  • The investigation focuses mainly on spatial relations in GenEval. Although the paper includes T2I-CompBench spatial evaluation and additional GenEval Colors and Color+Attribute results, broader evaluation would help establish how well the findings generalize to other tasks and benchmarks.
  • The paper investigates a limited number of models, primarily SD3-medium and FLUX.1-Schnell. Evaluating additional architectures and model sizes would help clarify whether the findings depend on model scale or architecture.
  • The description of the random-content controls needs clarification. Section 5.3 states that all three substitute-content conditions fall significantly below baseline, but Table 4 reports 30.0% for random permutation and 29.9% for baseline. Could the authors clarify this discrepancy and whether the results are robust across different sampled permutations and noise realizations?
  • The evaluation relies heavily on templated prompts and does not systematically examine instruction complexity. Testing prompts with multiple interacting spatial relations would help establish whether PR remains effective as instruction complexity increases.
Questions for authors

Please address all weaknesses raised above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.