acceptodds
Under review as a conference paper at ICLR 2027

Right Identities, Wrong Places: Reference Addressing in Multi-Reference Image Generation

Abstract

Multi-reference image generation integrates multiple visual references under textual instructions. When the prompt specifies spatial relations among these references, the model must ensure not only identity fidelity, but also reference addressing: assigning each reference to its designated spatial slot. Existing methods and evaluations, however, largely overlook reference addressing. Through a controlled counterfactual audit, we show that this is a distinct capability gap: models can preserve the requested identities yet fail to execute their prompt-specified reference-to-slot assignment, independent of identity loss, reference count, and physical input-order dependence. We introduce AddrBind, an inference-time executor that aligns an explicit reference-assignment plan with the model's emerging layout and redistributes, rather than amplifies, existing reference-conditioning influence, softly abstaining where the assignment is ambiguous. Across five reference-conditioning configurations and two benchmarks, AddrBind improves slot and exact-assignment accuracy in every configuration, with relative gains of and on our primary CogCanvas backbone.

Primary area:
generative models
Keywords:
Multi-reference image generation, Controllable image generation, Reference addressing

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_etqDOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper notes that when prompts use numerical identifiers to refer to reference images (e.g., "image 2 on the left"), multi-reference image generators often correctly identify the subject but place it in the wrong location. The authors quantified this issue across five backbones by conducting counterfactual audits using cyclic redistribution and input permutation. They proposed a training-free method called AddrBind: an MLLM planner determines which reference image belongs in which slot, and the attention mechanism among reference patches is redistributed according to this assignment while keeping the total reference attention constant. Slot accuracy across all five backbones surpassed that of the baseline generation methods.

Critical strengths
  1. The analysis is quite thorough. Separating "who is in the image" from "who is at which location" constitutes a useful protocol (Tables 1, 5).
  2. The performance gains hold true across five vastly different backbones and two benchmarks.
Critical weaknesses
  1. A few simple baselines are missing. Fig. 2(b) shows that appearance-based descriptions remain effective, whereas ID-based references fail; Table 1 demonstrates that, using the FLUX backbone, a canonical input order significantly outperforms a shuffled order. This naturally suggests two simple fixes: having the planner rewrite IDs as descriptions or reordering the inputs. Neither approach was tested. Furthermore, aside from MS-Diffusion and SIGMA-GEN, newer multi-subject models such as UNO and DreamO were neither compared against nor cited.

  2. The layout component has not been proven effective. According to the caption for Table 2, the "Oracle plan" and "static GT-region routing" utilize the same set of ground-truth (GT) regions and assignments. Yet, across both FLUX backbones, static routing performs noticeably better (Table 2: 0.871 vs. 0.800, 0.901 vs. 0.832; Table 4: 0.896 vs. 0.842). While the two methods tie on OmniGen2, the Oracle plan actually performs better on SDXL and Qwen. Consequently, for FLUX, the explanation provided in Sec. 4.2 fails to account for this performance gap; the layout-feedback stage might even be detrimental to performance.

  3. The cost in image quality is understated. On FLUX-4B, the BG-Sim score drops from 0.320 to 0.087 (Table 8), yet Sec. 4.3 describes this decline as "only slight." Across 152 routed MultiBanana tasks, the baseline evaluator's Instruction Alignment score drops by 0.58 on a 1–10 scale—with a confidence interval excluding zero (Table 12)—a fact completely omitted from the main text. Plotting a trade-off curve against the shift bound would make this cost much clearer.

Questions for authors
  1. How does AddrBind compare to prompt rewriting via a planner and to direct input reordering?
  2. On the FLUX backbone, why does the oracle plan underperform static GT routing when both have access to the same GT information?
  3. Are the main results based on a single seed?
5 comments
AuthorsOpenReview ↗

Response to Reviewer etqD (1/2): Baselines

Thank you for recognizing our diagnosis and cross-backbone improvements. We have added the suggested comparisons and shift-bound analysis, and clarified execution strategies.

W1 / Q1: Prompt rewriting, input reordering, and additional models.

We evaluated both input-side baselines on the same 50 four-person and 50 five-person CogCanvas tasks. Rewriting uses planner-generated appearance descriptions of the input references; reordering follows the predicted spatial plan and updates image indices consistently. AddrBind preserves the original prompt and reference order and executes the requested identity–position assignment during generation.

PeopleMethodSlot ↑Exact ↑Cov. ↑
4Native.300.000.620
4Native + rewriting.565.100.685
4Native + reordering.265.000.625
4UNO.185.000.205
4UNO + rewriting.170.000.275
4DreamO v1.1-IP.155.000.375
4AddrBind.665.280.755
5Native.272.000.472
5Native + rewriting.508.040.652
5Native + reordering.264.000.500
5AddrBind.608.120.604


Automatic rewriting helps. AddrBind improves Slot over rewriting by .100 at both cardinalities, with paired 95% intervals above zero (Appendix G.1). Exact point estimates are also higher; the five-person Exact difference interval crosses zero. Rewriting has higher five-person coverage, highlighting the distinction between recovering identities and placing them correctly. The tested reordering rule did not improve mean Slot. 

Native, input-side baselines, and AddrBind use FLUX.2-klein-4B. UNO and DreamO use their released FLUX.1-dev backbones for cross-system comparison, receiving all nine references on the four-person tasks. DreamO uses official person/object IP preprocessing and a full-image background-input adaptation. These comparisons and the formal UNO/DreamO citations are included in Appendix G.1, Table 13, and Section 4.1.

AuthorsOpenReview ↗

Response to Reviewer etqD (2/2): Execution, trade-offs, and seeds

W2 / Q2: Oracle versus static GT routing.

Both controls receive the same GT plan, but their execution differs: Oracle retains scout-based soft spatial responsibilities and gating, while Static directly routes within hard GT regions and preserves the original allocation elsewhere. Their comparison therefore measures the combined effect of these execution choices. In Table 4, Oracle and Static obtain Slot .842 and .896, respectively; we have corrected Section 4.2 to distinguish their execution paths.

The added 40-task comparison shares predicted plans across routed conditions (Appendix G.2, Table 14). Prior-only removes scout evidence while retaining the spatial prior, gating, and executor; Planner-static uses predicted hard regions. Slot is .6875 for Native, .8667 for Prior-only, .8375 for Planner-static, and .8875 for AddrBind. These results support the value of executing reference allocation; the complete executor achieves the highest point estimate, with paired difference intervals touching or crossing zero against the simplified controls.

W3: Content preservation and the shift-bound trade-off. 

We have replaced the “only slight” characterization with a description of backbone-dependent quality trade-offs. On FLUX.2-klein-4B, BG-Sim changes from .320 to .087 (Table 8); among 132 valid pairs from 152 routed MultiBanana tasks, Instruction Alignment changes by −.58 (95% CI [−1.01, −.15]; Table 12). 

The requested trade-off curve is now Figure 5 in Appendix G.3. On 40 fixed two-/three-person tasks with the FLUX.2-klein-4B bounded executor, only D varies:

ConfigurationSlot ↑BG-Sim ↑
Native.5625.3596
AddrBind, D = 2.8208.2380
AddrBind, D = 4.8417.0684
AddrBind, D = 8.8417.0633
AddrBind, default D ≈ 27.63.8417.0637

D = 2 retains most assignment gains over Native while improving background similarity relative to the default, with a slightly lower Slot point estimate. This provides an adjustable assignment–background balance within the same allocation mechanism. The default D = ln(10¹²) follows the original implementation’s numerical scale. 

Q3: Random seeds. 

The main results in the original submission are averaged over three random seeds. 

These clarifications and additions are marked in blue in the revised manuscript, with the new evidence in Tables 13–14 and Figure 5.

We hope these clarifications address your concerns and would be happy to discuss any remaining questions.

AuthorsOpenReview ↗

Paper version update

Dear Reviewer,

We have revised the main text and appendices to address your comments. All changes are highlighted in blue.

Sincerely

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

The paper studies multi-reference image generation where the prompt refers to people by image index and assigns each to a position. A controlled audit shows that current models often preserve the requested identities but place them in the wrong positions. The proposed training-free method uses an MLLM planner to extract the person to reference and to position assignment, combines it with early attention maps to estimate where each person should appear, and reweights attention among the reference images toward the assigned reference at those locations. Experiments on five backbones and two benchmarks show improved placement accuracy.

Critical strengths
  • The audit design separates placement from identity recovery. The cyclic-assignment and synchronized-permutation conditions (Sec 2.2, Tab 1, App. Tab 5) fix the scene, references and noise, and Fig 2(b) shows that index-based prompts stay near the 1/K chance level while appearance-based prompts on the same references do not. This makes the failure mode concrete and gives a reusable test that existing benchmarks do not provide.
  • The accuracy gains are consistent and the planner is not the bottleneck. Slot and exact accuracy improve on all five backbones on CogCanvas (Tab 2) and MultiBanana (Tab 3), and on held-out K = 4/5 by +.350 and +.340 on the two FLUX models (Tab 7). The learned plan is within .006 slot accuracy of the oracle plan on FLUX.2-klein-4B (.794 vs .800), so the remaining error lies in generation, not in parsing the prompt.
Critical weaknesses

W1. The quality cost is understated and hard to judge (Sec 4.3, App E-F, Fig 4). Sec 4.3 says the quality drops slight (around L440), but on FLUX.2-klein-4B BG-Sim falls from .320 to .087 and Attr-VQA from .337 to .217 (Tab 8), and object-relation scores fall on all five backbones (Tab 10). On MultiBanana, the benchmark's own judge is reported only for one backbone and scores routed outputs lower (Tab 12). The method may cause this -- G0 splits the whole canvas among the people and only person references receive the redistributed attention, so background and object references lose attention even at background pixels. Fig 4 cannot address the question. It only has three prompts at low resolution, omits the background and object references, shows visible scene changes in (a) and (b), and has no failure cases or baseline images.

W2. Limited novelty and missing simple baselines. The method is a variant of regional attention reweighting (e.g. DenseDiffusion, regional prompting, MAST/LAMA). The added parts get little support: removing the confidence gate costs .013 slot accuracy (Tab 4), static GT-region routing beats AddrBind on three of five backbones (Tab 2), and no ablation uses the language prior alone without the trial-step attention. Plus, cheaper alternatives are not tested. Fig 1(c) and Tab 1 show that models follow a matching input order, so the planner could simply reorder the references; the planner's positions could also be fed as boxes to static routing. Initial-noise methods [1-3] steer layout without overriding attention at every step and may avoid the cost in W1.

References

[1] InitNO: Boosting text-to-image diffusion models via initial noise optimization, Guo et al., CVPR 2024.

[2] ReNO: Enhancing one-step text-to-image models through reward-based noise optimization, Eyring et al., NeurIPS 2024.

[3] Be decisive: Noise-induced layouts for multi-subject generation, Dahary et al., SIGGRAPH 2025.

Questions for authors
  1. Which layers and denoising steps are routed, what is the shift bound D, and at which timestep, layers and heads is the trial-step attention averaged? An ablation over the routed step range would show whether routing at every step is needed, or whether early steps are enough and cost less quality.
  2. Qwen-Image-Edit-2511 is "paired on the tasks where its conserving write passed a pre-registered numerical criterion" (App. E). What is the criterion, and how many tasks were excluded?
  3. The CogCanvas metrics in App. E are reimplemented and differ from the CogCanvas paper's description (DINOv2 vs DINOv3 with SAM 3 masks for BG-Sim, LAION aesthetic v2 vs v2.5, different Attr-VQA judges). Are the numbers comparable to those reported in the CogCanvas paper, and will the reimplementation be released?
  4. How were the Fig 4 examples selected?
  5. Does AddrBind reduce the attention share of the background and object references at background pixels (see W1)? It will be helpful to see a per-reference attention-share plot split by foreground and background.
  6. Please consider the baselines and ablations listed in W2, in particular the planner-driven reordering baseline.
5 comments
AuthorsOpenReview ↗

Response to Reviewer NcXC (1/3)

Thank you for your careful reading and detailed, constructive feedback, and for recognizing our diagnosis of reference-addressing failures and the improvements across generators. Your comments on content preservation, simpler baselines, and evaluation transparency helped us clarify the mechanism and its practical trade-offs. In response, we have added targeted diagnostics and controlled comparisons, expanded the qualitative evidence with complete inputs and paired outputs, and released the metric implementation. We address your concerns point by point below. FLUX-4B/9B denote FLUX.2-klein-4B/9B.

W1 + Q5: Reference allocation and content preservation

Thank you for highlighting the need to make this trade-off clearer. AddrBind aims to generate the requested people in their instructed positions. It translates reference-index–position relations into local allocation targets and corrects deviations of current reference shares from those targets, bringing index instructions into generation beyond planner parsing. Total reference attention is preserved while relative shares change. As requested, Figure 8 (Appendix H.3) reports per-reference shares separately at foreground/background locations on 20 FLUX-4B tasks, 10 each at four/five people. Aggregate person-reference shares increase while object/background shares decrease, including at background locations. Responsibilities express relative allocation among subjects, so confidence gating can also act at background locations. Bars compare the same update before/after routing, averaged equally across tasks; foreground includes people and objects, with regions independently segmented and reviewed from final outputs.

The allocation mechanism offers a choice of intervention strength. D bounds each correction, balancing execution of index instructions with retention of native allocation. We fix FLUX-4B, 40 tasks (20 each at two/three people), and other settings, varying only D:

ConfigurationSlot ↑BG-Sim ↑
Native56.25%0.3596
AddrBind, D=282.08%0.2380
AddrBind, D=484.17%0.0684
AddrBind, D=884.17%0.0633
AddrBind, default D≈27.6384.17%0.0637

On this subset, a smaller D retains most assignment gains while improving background similarity relative to the default. Thus, improved index assignment is achievable with weaker intervention; stronger intervention yields the higher Slot shown in the table. D=4 and D=8 actively constrain updates despite matching the default mean Slot. BG-Sim measures whole-image similarity to the background reference. This trade-off study covers two/three-person tasks; the four/five-person diagnostic examines redistribution.

On MultiBanana, official Instruction Alignment for the evaluated backbone changes by −0.58 on average over 132 valid pairs (95% CI [−1.01, −0.15]; Table 12), further characterizing the content-preservation trade-off accompanying improved index assignment.

Section 4.3 now replaces “slight” with a description of disaggregated quality changes and adjustable trade-offs, referring to the corresponding tables and D analysis. Tables 8, 10, and 12 report background, attribute, object-relation, and instruction-following changes. Appendix Figure 5 shows the D curve; Figures 6–7 provide complete references and Native/AddrBind outputs to inspect correct person assignment alongside the associated content preservation.

AuthorsOpenReview ↗

Response to Reviewer NcXC (2/3)

W2 + Q6: Contribution and simple baselines

Thank you for suggesting simpler alternatives. Our contribution is to translate index–position plans into reference allocations executed during generation. The planner specifies who should appear where. AddrBind combines spatial priors and scout evidence into subject responsibilities, maps them through reference indices to allocation targets, and corrects deviations of current shares from those targets. Confidence-modulated, bounded corrections preserve total reference attention. DenseDiffusion (Kim et al., 2023) and regional prompting (Chen et al., 2024) use regional conditions; LAMA in MAST (Kang et al., 2026) uses style–mask pairs. In Table 2's same-backbone FLUX-9B comparison, reference-channel adaptations of DenseDiffusion and regional prompting achieve Slot 0.604 and 0.678, versus 0.826 for AddrBind.

First, we test what input adjustments alone can achieve. On FLUX-4B, 50 tasks each at four/five people use planner-generated appearance descriptions for rewriting, or predicted positions for reference reordering with synchronized indices. All conditions retain the intended identity–position assignment and complete reference set.

MethodFour-person Slot ↑Four-person Exact ↑Five-person Slot ↑Five-person Exact ↑
Native0.3000.0000.2720.000
Native + rewriting0.5650.1000.5080.040
Native + reordering0.2650.0000.2640.000
AddrBind0.6650.2800.6080.120

Appearance rewriting does help. AddrBind retains the original index-based prompt and reference order while achieving higher Slot at both cardinalities, with paired 95% difference intervals above zero. Exact point estimates are also higher; the five-person difference interval crosses zero. Five-person Cov. is 0.652 for rewriting versus 0.604 for AddrBind, showing why identity coverage and correct assignment should be evaluated separately. The tested reordering rule does not improve mean Slot.

Next, we fix predicted plans and compare their execution. On FLUX-9B, 40 tasks (20 each at two/three people) compare Prior-only, which removes scout while retaining the spatial prior, gate, and executor, with Planner-static, which routes directly through predicted hard regions.

MethodSlot ↑Exact ↑Cov. ↑BG-Sim ↑
Native0.68750.47500.83330.3048
Prior-only0.86670.72500.91250.1355
Planner-static0.83750.62500.89580.1797
AddrBind0.88750.75000.93330.1182

Prior-only retains our plan-to-reference allocation mechanism; its gains over Native support this execution pathway. Its comparison with the full method further tests scout's additional contribution. Full AddrBind has the highest Slot, Exact, and Cov. point estimates; Slot difference intervals against Prior-only and Planner-static touch and cross zero, respectively. These controls also appear in Appendix G, Tables 13–14.

In Table 2, GT-static converts annotated index–position assignments into fixed hard regions, while AddrBind uses planner-predicted plans. GT-static achieves higher Slot on both FLUX backbones and OmniGen2. Oracle shares its GT plans but retains scout, soft responsibilities, and gating, so their comparison reflects differences in execution strategy. The added Planner-static control shares predicted plans with AddrBind, testing how the same plan is translated into reference allocations. The original gate ablation also yields a Cov. gain of 0.023 (95% CI [0.002, 0.048]).

These added controls test input adjustments, scout evidence, and plan execution, providing direct experimental evidence on the gains of the index-assignment mechanism and the roles of its components. Q1 further tests the trade-off between continued routing and background preservation.

AuthorsOpenReview ↗

Response to Reviewer NcXC (3/3)

Q1: Routing configuration, scout collection, and continued execution

Thank you for suggesting the routing-duration analysis. AddrBind applies fixed identity–position targets throughout generation while recalculating corrections from current reference shares at each attention call. Scout-derived spatial responsibilities and confidence remain fixed, while corrections adapt to current shares, allowing assignment targets to remain active during later denoising. Configurations are listed below; FLUX-4B/9B denote FLUX.2-klein-4B/9B.

BackboneRouting/scout modulesHeadsRouted stepsNative scout timestep
FLUX-4BAll 5 dual-stream + 20 single-stream blocks24 per layer1–281000
FLUX-9BAll 8 dual-stream + 24 single-stream blocks32 per layer1–281000
OmniGen232 joint-attention layers; positive text + reference branch21 Q / 7 KV1–500
Qwen-Image-Edit-2511All 60 joint-attention layers24 per layer1–401000
SDXL + IP-Adapter70 cross-attention modules; text branch for scout, reference branch for routing10 modules × 10; 60 modules × 201–501; formal 50-step schedule starts at 981

Scout collects canvas-to-subject-text attention, averaging over heads and then equally across collection layers. OmniGen2 expands KV heads to the query-head count; SDXL uses the conditional branch and bilinear resampling to 64×64. Timesteps follow each scheduler’s native definition: FLUX and OmniGen2 use independent one-step calls, Qwen uses the first step of the formal 40-step schedule, and SDXL main results use the listed configuration.

The explicit bounded executor uses a default bound D=ln(10¹²)≈27.63, inheriting the scale of the original 10⁻¹² numerical floor. D bounds correction magnitude; the actual correction depends on current shares and allocation targets.

We reuse the D study’s fixed 40 tasks (20 each at two/three people), keeping FLUX-4B, default D, and other settings unchanged while varying the routed interval. All conditions complete 28 denoising steps.

Routed stepsSlot ↑BG-Sim ↑
First 70.71250.1109
First 140.80000.0658
All 280.84170.0637

Full-duration routing achieves the highest Slot point estimate, supporting continued index-assignment execution during generation. Compared with the first 14 steps, routing throughout further improves assignment with similar BG-Sim point estimates. Routing only during the first seven steps retains higher background similarity with smaller assignment gains. These results directly support our full-duration configuration in this setting and show how routing duration affects assignment and background preservation.

Q2: Qwen numerical criterion and exclusions

All candidates undergo the same checks: maximum absolute change in total reference attention ≤10⁻⁴ across steps, layers, heads, and canvas positions; finite diagnostics; zero numerical-protection triggers; and successful routing execution. Decisions use numerical execution records.

Paired comparisonCandidatesExcludedExclusion rateValid pairs
Native / AddrBind16263.70%156
Native / Static GT-region routing1623420.99%128

Excluded tasks are removed from both arms; each method is paired with Native on its corresponding valid set.

Q3: Metric comparability and release

All methods share DINOv2 ViT-L/14, LAION aesthetic v2, and fixed VQA judging, supporting internal paired comparisons. Absolute scores are not directly comparable to the official CogCanvas implementations. Appendix E clarifies the metric implementations and comparability. We have released the evaluation reimplementation, scoring configurations, and usage instructions through the anonymous link provided in the paper’s abstract.

Q4: Figure 4 selection and additional examples

For Figure 4, we randomly sampled one example each from the two-, three-, and five-person tasks.

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

The model can keep the right people and still put them in the wrong places. AddrBind tries to fix this without training. InternVL3.5-8B reads which photo goes where. Words like "left" and "right" become a soft mask on the canvas, and that mask stays fixed. The method then moves attention toward the matching reference photo, without adding more attention overall. On CogCanvas with FLUX.2-klein-4B, the original model's slot accuracy is 60.7%, and AddrBind's is 79.4%. The share of images where every person is in the right place is 38.8% for the original model and 58.5% for AddrBind.

Critical strengths

The check is useful. The model often still has the right people, but does not follow "image 1 on the left, image 2 on the right." AddrBind is a training-free try at that problem.

Critical weaknesses
  1. The mask comes from a fixed list of words, such as left, right, center, and upper left. It is drawn once and then kept still. With more objects, or with positions outside that list, this will break. The paper also never says how the mask is drawn.

  2. The paper already goes to 5 people on CogCanvas, and to 7 on MultiBanana. On FLUX.2-klein-4B with 5 people, the original model's slot accuracy is 26.4%, and AddrBind's is 57.1%. The share of images where every person is in the right place is 0% for the original model and 10.7% for AddrBind. That score only checks whether a face lands in a named slot. There is no mIoU, and nothing at 8–10 objects. In my experience this kind of training-free method does not put the regions in the right place. Please give a layout with 5 to 10 objects and report mIoU.

  3. Please compare with methods that already take a layout and the reference photos, such as ContextGen (ICLR 2026). Doing better than the original model is not that comparison.

  4. The change inside the denoiser is nothing new. A model first writes a coarse layout, and many methods already do that. The layout is then used to add a number to the attention scores of the matching reference image. Adding a number to the logits is the same as multiplying the weights before softmax. Many methods already boost attention with a mask. There is no new operation inside the denoiser.

These four points are why this is a weak reject. The main thing is to explain how the method still works when there are more instances than position slots. I will also look at how the mask is drawn, the mIoU for 5 to 10 objects, and the ContextGen comparison, and then decide whether to raise my score.

Questions for authors
  1. How do you turn a word such as "left" into the soft mask? Please give the formula, how big the mask is, and how the weight fades out. Your word list only has a few fixed spots, such as left, right, center, and upper left. If there are more people than these spots, or the prompt asks for a place that is not in the list, is that person dropped, or can the method still put them in the right place?
  2. Please give a layout and report mIoU for 5 to 10 objects. mIoU means how much each object's area overlaps the box in that layout. Your slot accuracy at 5 people only checks whether a face falls into a named spot such as left or right. That is a different number, so please do not answer with it.
  3. How does AddrBind compare with ContextGen (ICLR 2026)?
5 comments
AuthorsOpenReview ↗

Response to Reviewer xJyD (1/3): Spatial Priors

Thank you for the careful and detailed review. Your suggestions on spatial priors, larger layouts, and related methods help us clarify the setting and provide more direct evidence. We study whether a generator given text and ordered reference images can place the correct reference identities as instructed. AddrBind improves this capability through inference-time reference allocation, without additional training or changes to the user-facing input interface. Below, we address each concern with implementation details, 5–10-person layout diagnostics, and a ContextGen comparison.

W1 + Q1: Spatial priors, shared locations, and unsupported expressions

The planner parses each subject's reference index and spatial role; a deterministic compiler converts the resulting location field into a soft prior. The fixed rules apply to planner outputs, not the user's complete sentence. Location categories are reusable rather than exclusive subject slots.

Mask construction. On the normalized canvas (x,y)∈[0,1]2(x,y)\in[0,1]^2, a location maps to (cx,cy,w,h)(c_x,c_y,w,h):

m_s(x,y)=b(x;c_{x,s},w_s)b(y;c_{y,s},h_s).$$ Here $\sigma$ is sigmoid, $\tau=1/16$, and $y$ increases downwards. The parameters are: | Axis | Keywords | Center | Width / height | |---|---|---:|---:| | Horizontal | left | 0.20 | 0.42 | | Horizontal | center / centre / middle | 0.50 | 0.42 | | Horizontal | right | 0.80 | 0.42 | | Vertical | top / upper | 0.25 | 0.50 | | Vertical | bottom / lower | 0.75 | 0.50 | | Vertical | front / foreground | 0.70 | 0.60 | | Vertical | background / behind | 0.40 | 0.60 | | Vertical | no matched vertical keyword | 0.55 | 0.70 | For example, left maps to $(0.20,0.55,0.42,0.70)$ and upper left to $(0.20,0.25,0.42,0.50)$. Width and height are canvas fractions; weights decay smoothly around and beyond the nominal boundaries. Masks cover the canvas-token grid: $64\times64$ for the main FLUX setting at $1024\times1024$. **They guide reference influence; the generator determines the final person's silhouette, size, and pose.** **Shared locations.** Each subject retains its own reference index and plan record. For active records $\mathcal A$, the prior and subject-specific scout text attention $L_s(q)$ determine responsibility: $$G^0_{sq}=\frac{m_s(q)}{\max\{1,\sum_{j\in\mathcal A}m_j(q)\}},\qquad \rho_{sq}=\frac{G^0_{sq}L_s(q)}{\sum_{j\in\mathcal A}G^0_{jq}L_j(q)}.$$ A zero responsibility denominator sets responsibility and routing strength to zero, preserving the native allocation there. Two subjects assigned left share a coarse prior but retain separate reference indices and scout responses. **Exceeding the number of location categories does not truncate subjects.** Responsibilities remain fixed after scout; corrections are recomputed from current reference shares at every routed call. **Unsupported expressions.** The compiler matches the planner's location field. A supported horizontal keyword enables a prior; vertical keywords refine it. For example, far left uses the left prior without an additional adjustment for far. Unmappable records receive no active spatial target, while their full prompt, reference inputs, and evaluation entries remain intact. If no records are active, generation follows the native path. **AddrBind turns parsed index assignments into allocation targets executed during generation.** Appendix B.1, “Spatial-prior construction,” now provides these formulas, parameters, boundary behavior, and a mask visualization.
AuthorsOpenReview ↗

Response to Reviewer xJyD (2/3): Larger Layouts and mIoU

W2 + Q2: Layout mIoU with 5–10 people

Following your request, we added a larger-layout diagnostic. The CogCanvas person tasks used here cover up to five people, so we recombined their person-reference assets into nine constructed tasks: three each with 5, 8, and 10 people. After fixing cardinality, we used GPT-6 Astra to draft scene and position prompts following CogCanvas's “Person i from image i” phrasing. Native and unchanged AddrBind use identical text, ordered person references, and generation settings: FLUX.2-klein-9B, 1024×10241024\times1024, and four sampling steps.

Regions and evaluation. Text specifies coarse positions, not exact body dimensions or boundaries. Before generation, we fix evaluation regions consistent with this layout: three upper/two lower for five people, four per band for eight, and five per band for ten. The bands have equal heights and their regions equal widths. These boundaries define a common evaluation convention, not uniquely specified body boxes or additional layout inputs to the generators. Grounding DINO detects person boxes. One-to-one matching maximizes total IoU with target regions, using each detection at most once. Unmatched targets score zero; we average over all requested people, then over the three tasks per group. This identity-agnostic bbox mIoU measures geometric overlap with coarse regions, separately from identity assignment.

PeopleTasks per methodNative mIoU (%) ↑AddrBind mIoU (%) ↑
5329.0129.63
8340.0341.63
10348.4448.74

AddrBind yields small average geometric-overlap improvements at all three cardinalities. The original planner/compiler retains all requested subjects in all nine tasks; eight people share four coarse priors, and ten share six. Location categories therefore impose no subject-count cutoff. Each subject retains a separate reference index. Because region IoU also depends on body size and region occupancy, we focus on paired changes under this common protocol; identity–position correspondence is assessed by the original assignment metrics.

See Appendix H.4, “Supplementary multi-person layout diagnostics,” specifically “Data construction” and “Layout evaluation.” “Constructed examples” shows one case per cardinality; the quantitative evaluation includes all nine.

AuthorsOpenReview ↗

Response to Reviewer xJyD (3/3): ContextGen and Contributions

W3 + Q3: ContextGen comparison and the value of index assignment

Thank you for recommending ContextGen (Xu et al., ICLR 2026), an insightful related work whose layout anchoring and identity-consistency attention provide explicit spatial control for multi-instance generation. We have added it to Related Work and conducted the requested comparison on the same tasks.

ContextGen and AddrBind operate under different settings: these scores do not directly establish the relative merits of their mechanisms under identical conditions. ContextGen receives an explicit visual layout and uses a trained adapter. AddrBind retains text and ordered reference inputs, improving index–position execution through training-free intervention. Their conditioning interfaces, backbones, and training differ.

To address your request, we constructed ContextGen's required layout inputs from the task text. This compares complete systems under their respective interfaces. AddrBind converts textual assignments into reference-allocation targets and corrects current shares; the same-backbone, same-input Native/AddrBind comparison directly tests that mechanism.

Setup. We use 50 four-person and 50 five-person CogCanvas tasks, with the same tasks and complete person, object, and background references across methods, all at 1024×10241024\times1024. Native and AddrBind retain their original inputs. For ContextGen, we directly use the authors' released generation code, model and adapter weights, feeding adapted CogCanvas inputs through its native interface. We construct placement boxes from textual positions, extract people and objects using source-reference foreground masks, preserve person aspect ratios, place objects near their corresponding people by a fixed rule, and use the background reference as the layout base. Adaptation rules are fixed before generation. All methods use our original Slot, Exact, and Cov. evaluator.

Four people (N = 50)

MethodSlot ↑Exact ↑Cov. ↑
Native0.3000.0000.620
AddrBind0.6650.2800.755
ContextGen0.9250.7800.900

Five people (N = 50)

MethodSlot ↑Exact ↑Cov. ↑
Native0.2720.0000.472
AddrBind0.6080.1200.604
ContextGen0.9080.7200.896

Backbones and training. Native and AddrBind use FLUX.2-klein-9B; ContextGen uses its official FLUX.1-Kontext-dev backbone and trained adapter weights. Porting it to our backbone would require additional adaptation training beyond the rebuttal period, so we retain its released model and native layout interface.

ContextGen achieves higher scores with the supplied explicit layouts. Our contribution addresses how an existing generator can better execute textual index assignments without user-supplied layout images or additional training. Under this setting, AddrBind improves Native Slot by 36.5 and 33.6 percentage points, and Exact by 28 and 12 points. These gains under matched conditions demonstrate the value of reference allocation and distinguish our objective from explicit-layout generation.

W4: Research contribution and reference-allocation mechanism

We identify the gap between generating the correct identities and executing their requested assignments, and turn text assignments into inference-time reference allocation. Controlled assignment changes and synchronized reference permutations separate this failure from identity recovery and physical input order, establishing a diagnostic and evaluation of reference addressing.

Executing which reference should influence each location requires both a reference-specific target and a rule for correcting the current allocation. AddrBind constructs targets from the plan, spatial prior, and scout response. At each routed call, a target-to-current log ratio determines the reference-block bias. Thus, a fixed spatial target can yield different corrections as allocation changes.

The rule preserves total reference-attention mass and, in shared attention, relative token weights within each reference block, with confidence and a shift bound regulating intervention. Additive bias and multiplicative reweighting describe equivalent implementations; the technical design lies in the reference-specific targets and the constrained, state-dependent allocation rule.

The same-backbone controls in Table 2 support this design: on FLUX.2-klein-9B, reference-channel adaptations of DenseDiffusion and regional prompting obtain Slot 0.604 and 0.678, versus 0.826 for AddrBind. Our contribution combines problem diagnosis, reference-level targets, and constrained execution to improve existing generators' identity–position assignment without changing their input interface or requiring additional training.

Comments on the paper

2 comments
~Shubhankar_Borse1OpenReview ↗

Missing related works/comparisons

Dear Authors, It seems this paper misses several related works as theoretical comparisons and baselines from NeurIPS25, ICLR26, CVPR26.

<1> https://arxiv.org/abs/2506.20879 <2> https://arxiv.org/abs/2510.14975 <3> https://arxiv.org/html/2510.20820v4 <4> https://arxiv.org/abs/2511.22690

AuthorsOpenReview ↗

General Response and Summary of Revisions

As the discussion period draws to a close, we would like to thank all three reviewers and the area chair for their time, careful consideration, and constructive feedback. The reviews have helped us clarify the paper’s contribution and strengthen the supporting evidence.

The main concerns and corresponding responses are summarized below:

  • Reviewer etqD: The review raised questions about simpler alternatives, the interpretation of execution strategies, and quality trade-offs. The responses provide additional comparisons with prompt rewriting, reference reordering, and related generation systems, clarify the execution controls, and quantify the trade-off between assignment accuracy and content preservation.
  • Reviewer NcXC: The main concerns involved the allocation mechanism, its incremental contribution, evaluation transparency, and qualitative evidence. Additional attention diagnostics, simplified execution baselines, and routing-duration ablations address these points. Implementation and evaluation details have also been clarified, the evaluation code released anonymously, and the qualitative material expanded.
  • Reviewer xJyD: The review focused on spatial-prior construction, larger multi-person layouts, comparison with ContextGen, and technical novelty. The responses explain the mask construction and boundary cases, provide exploratory 5-, 8-, and 10-person layout diagnostics and direct ContextGen comparisons, and clarify the differences in conditioning, setting, and the contribution of reference-level allocation.

Together, these additional experiments and manuscript clarifications strengthen the evidence for our central contribution: improving identity–position assignment in existing multi-reference generators without additional training or user-supplied layouts.

We hope the responses and revisions address the main concerns raised during review. With the discussion deadline approaching today, any remaining questions or feedback would be especially welcome, and we will do our best to provide further clarification before the discussion closes. We respectfully invite the reviewers to consider the additional evidence and revisions when reassessing their scores and recommendations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.