acceptodds
Under review as a conference paper at ICLR 2027

FaceRoute: Selective Reference Reweighting for Multi-Person Identity Preservation

Abstract

Multi-person image editing must preserve several facial identities while using full-image reference context. Frozen diffusion-transformer editors already encode the reference images, yet their generated faces can lose distinctive features or take on attributes of another person. We ask (i) whether stronger access to reference evidence improves identity preservation and (ii) whether this strengthening should be spatially selective. We propose FaceRoute, a family of training-free routing variants that add a bias from target-image queries to selected reference VAE keys, without changing backbone weights or adding a conditioning branch. On the two-person and three-/four-person splits of MultiID-Bench, whole-image routing (FaceRoute-W) at improves reference and ground-truth identity similarity for both Qwen-Image-Edit-2511 and FLUX.2-klein-base-9B, with all eight paired bootstrap intervals above zero. Face-local routing (FaceRoute-F) also improves over both base editors in paired tests, and its two-person gains replicate across three seeds. Both variants also raise copy-paste preference and slightly lower aggregate aesthetic scores relative to the base editors; for whole-image routing, both costs grow when is raised from to . The answer to (ii) depends on reference face scale: on ordinary references, whole-image routing has higher paired identity similarity, and a mass-matched control links this advantage to the larger injected reference attention mass; on a small-face stress test, face-local routing scores higher on both identity metrics, with higher aesthetic scores and lower copy-paste preference.

Primary area:
generative models
Keywords:
reference-guided image editing, multi-person identity preservation, training-free inference, attention routing, diffusion transformers, subject consistency, personalized image generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1JUcOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper answers "whether stronger attention to references improves id-preservation" through a training-free intervention that adds a constant log α to the attention logits from all target-image queries to a selected set of reference VAE keys. On MultiID-Bench with Qwen-Image-Edit-2511 and FLUX.2-klein-base-9B, α = 1.5 improves both reference and ground-truth face similarity on every backbone–split combination, at the cost of higher copy-paste preference and slightly lower aesthetic scores; whole-image biasing is slightly stronger than face-local biasing on ordinary references, an edge that vanishes once injected attention mass is matched, whereas face-local biasing scores higher on a 49-example small-face set. The appendix contains extensive audits on copy-paste geometry, identity blending, completeness, a text-key pathway control, XVerseBench transfer, and a video extension.

Critical strengths
  • Extensive experiments almost answer all the concerns I raised while reading the article. The statistical reporting is unusually careful for this area.
  • Simple yet effective plug and play method.
  • Some experimental insights are rather intriguing.
Critical weaknesses
  • Practical method, yet no practical problem. My major concern is, the main problems answered in this paper are kind of, well, predictable. By FaceRoute, I thought there would be some routing method to solve the practical problems like blending issues or layout issues, but it seems that the proposed methods are too simple for an ICLR main track paper.
  • No same-backbone trade-off comparison. The only comparisons are routed vs. unrouted and whole vs. face-local. Without other ways of increasing reference reliance on the same backbone (e.g., attention temperature, reference resolution or scale, identity emphasis in the prompt), plotted on the same (CP, simGT) or (preference, simGT) plane, it is impossible to tell whether FaceRoute is a better knob or simply a knob. The system-level comparison does not resolve this: WithAnyone is better than Qwen + FaceRoute-F (α = 2.5) on simref, simGT and CP simultaneously (Table 1).
Questions for authors
  • Just out of curiosity, have you tried applying this to adapter-based methods like PuLID or WithAnyone? It seems simple enough to transfer this method to those cross-attention-based ID-preserving models. You can just answer no to this question since I agree that it is more valuable to orient your time to base models like Qwen and FLUX.
  • Can you provide more discussion about M.8? Since many previous works adopted attention masks for isolation, I found this insight quite surprising.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 1JUc

Thank you for the encouraging review. Brief answers below; full details are in the new Appendix Z.

Q1. PuLID / WithAnyone. We tried WithAnyone (PuLID needs its own trained adapter and did not fit in the rebuttal period). WithAnyone has no reference VAE tokens in joint attention, so we added the same bias (α=1.5) to its SigLIP face branch. ΔSimGT: P3 −0.0002 [−0.0020, 0.0016]; P2 +0.0003 [−0.0007, 0.0014]. Nothing changes, likely because an identity-trained model already exploits its identity branch, whereas frozen editors under-use their reference tokens, which is where FaceRoute helps (App. Z.5).

Q2. Isolation (M.8). With FaceRoute-W on, we add hard masks between references (Qwen P3, α=2):

IsolationSimRef ↑SimGT ↑SimMis ↓CP ↓
None (FaceRoute-W only)0.63450.40900.10670.2761
Face-only0.57700.39020.12920.2294
Full0.51640.35570.15360.2006

Isolation lowers identity and raises wrong-identity similarity. Our reading: prior mask-based works train with the masks, while a frozen editor was trained with all reference tokens mutually visible, so cutting paths at inference removes context it relies on. For frozen editors, re-weighting works better than cutting.

W1. Practical problem; simplicity. The problem is identity dilution in frozen multi-reference editors: faces drift from the references as the number of people grows. Existing fixes train adapters or require layout inputs; FaceRoute is a one-line, training-free bias that works on two backbones. On 50 new five-person groups (public MultiID-Bench P1 inputs, α=1.5 fixed in advance, FLUX.2-klein, CFG 4; SimGT is computed against each identity's P1 ground-truth image; paired 95% CI vs. baseline; paired 95% CI vs. baseline):

MethodSimRef ↑SimGT ↑CLIP-T ↑
Baseline0.42720.30450.2586
FaceRoute-F0.4561 (+0.0290 [0.0162, 0.0412])0.3222 (+0.0177 [0.0080, 0.0266])0.2570
FaceRoute-W0.4683 (+0.0411 [0.0233, 0.0582])0.3258 (+0.0213 [0.0086, 0.0335])0.2561

All 50 groups are shown in App. Z.6.

W2. A better knob or simply a knob? We ran the comparison you proposed: same frozen backbone, references, prompts and seeds, all on one (ΔCP, ΔSimGT) plane (P3, paired vs. the same baseline; Fig. 12, App. Z.1). Temperature divides the attention logits of the target-image queries by τ, over all keys; reference area changes the reference size actually fed to the VAE; the identity-emphasis prompt appends the same identity-preserving instruction to every prompt; K/V scaling multiplies the reference keys or values.

ControlΔCPΔSimGT, 95% CI
FLUX: FaceRoute-F, α=1.5+0.0288+0.0362 [0.0273, 0.0451]
FLUX: FaceRoute-W, α=1.5+0.0480+0.0570 [0.0434, 0.0739]
FLUX: temperature τ=0.8−0.0152−0.0286 [−0.0555, −0.0058]
FLUX: temperature τ=0.9−0.0089−0.0101 [−0.0282, 0.0059]
FLUX: reference area ×0.5−0.0413−0.0383 [−0.0562, −0.0200]
FLUX: reference area ×0.25−0.0643−0.0734 [−0.0943, −0.0510]
FLUX: identity-emphasis prompt−0.0204−0.0135 [−0.0310, 0.0037]
Qwen: FaceRoute-F, α=1.5+0.0500+0.0385 [0.0225, 0.0556]
Qwen: FaceRoute-W, α=1.5+0.0784+0.0371 [0.0202, 0.0571]
Qwen: reference K ×0.8−0.1689−0.2434 [−0.2984, −0.1898]
Qwen: reference K ×1.25−0.1696−0.2066 [−0.2835, −0.1290]
Qwen: reference V ×0.8−0.0365−0.0277 [−0.0499, −0.0058]
Qwen: reference V ×1.25−0.0018−0.0208 [−0.0378, −0.0028]
Qwen: GRAG-style d=0.8−0.1731−0.2436 [−0.3011, −0.1849]
Qwen: GRAG-style d=1.2−0.0554−0.0484 [−0.0881, −0.0099]

The FaceRoute arms are the only tested controls that improve identity while keeping image quality; the alternatives give no identity gain. On Qwen, aesthetics stay at the baseline level for FaceRoute (baseline 5.47, F 5.44, W 5.35) but fall to 3.80 for K ×1.25 and 4.37 for GRAG-style d=1.2. Reference resolution acts as an information bottleneck (the benchmark references are already tightly framed around the person), not as a substitute for routing. On Qwen, F matches W's gain at a lower CP. The WithAnyone row in Table 1 is cross-system context (identity-trained vs. training-free); the same-backbone plane above answers the knob question.

We hope this addresses your concerns and are happy to provide anything further.

Reviewer_1JUcOpenReview ↗

Revierwer Response

Thank you for the clarification. My questions are fully resolved.

I believe the insights presented in this paper are valuable to the academic community. For the record, I would like to note that I would raise my rating to somewhere between “Strong Accept” and “Weak Accept,” but it seems there is no such an option.

Thank you very much for your time and for the supportive assessment. We are glad the additional experiments resolved your questions, and your suggestions made the paper stronger. All new results are included in the revised version.

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

FaceRoute adds a fixed bias to target-to-reference attention logits in frozen image editors. It compares boosting all reference tokens with boosting detected face tokens, reporting identity gains on two backbones and a possible advantage for face-local routing when reference faces are small.

Critical strengths

The intervention is simple and requires no training. Paired comparisons, multiple recognition encoders, and multi-seed tests provide useful evidence for the identity gain. The paper also reports the copy-paste and aesthetic costs rather than hiding them.

Critical weaknesses
  1. The main insight about when selective routing is preferable remains weakly established. On ordinary references, the mass-matched control finds no resolved identity difference between the two scopes. The claimed reversal relies on a separate 49-example small-face set, with a 31-example cross-backbone subset; the primary small-face aggregates use method-specific IQR filtering and have no uncertainty intervals. Differences between datasets are not a controlled intervention on face size. This limits the evidence for the face-scale explanation.

  2. The default strength is selected retrospectively on the evaluation benchmark (Sec. 5.2). The split-half checks are helpful but do not replace independent validation of the operating point. Since stronger routing also increases copy-paste preference and reduces aesthetics, the reported identity improvement alone does not establish a reliable selection rule or overall advantage on new inputs.

Questions for authors
  1. Can you report unfiltered, exact-pair confidence intervals for the small-face scope comparison, and clarify whether a matched-reference face-scale or attention-mass control exists there?
  2. How would scope and strength be selected on unseen data? Are results available with these choices fixed on a separate validation set?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Pfbc

Thank you for the careful review. We ran each analysis you asked for; details are in Appendix Z.2–Z.4.

Q1 / W1. Small-face evidence and controls.

(a) Unfiltered exact-pair intervals, F−W (95% paired bootstrap):

Small-face setΔSimRefΔSimGTΔCP
Qwen, full+0.00265 [−0.01083, 0.01675]+0.00275 [−0.00717, 0.01164]−0.01099 [−0.03119, 0.00863]
Qwen, cross-backbone subset+0.00833 [−0.00830, 0.02730]+0.00344 [−0.00449, 0.01217]−0.00821 [−0.03461, 0.02008]
FLUX, cross-backbone subset+0.00463 [−0.00335, 0.01296]+0.00750 [−0.00069, 0.01689]−0.00635 [−0.01556, 0.00239]

Point estimates agree in direction in every row. On ordinary references W is ahead of F at equal α; on small faces that advantage is gone, and F reaches the same identity while biasing far fewer tokens. The paper now says "favors FaceRoute-F" in place of "reverses". All 49 examples are shown in App. Z.7.

(b) Controlled intervention on face size (FLUX P3): each reference is placed on a neutral canvas at 100/75/50% occupancy, with token budget, prompts and seeds fixed. F keeps its full gain as the same face shrinks (F−B ΔSimGT +0.0280, +0.0255, +0.0322, all intervals above zero), while W's edge over F disappears:

Occupancy, F−W at α=1.5ΔSimRefΔSimMis ↓ΔCP ↓
1.00−0.0296 [−0.0520, −0.0117]−0.0064 [−0.0110, −0.0022]−0.0222 [−0.0382, −0.0080]
0.75−0.0253 [−0.0447, −0.0052]−0.0025 [−0.0093, 0.0050]−0.0199 [−0.0340, −0.0065]
0.50−0.0003 [−0.0180, 0.0169]−0.0127 [−0.0232, −0.0039]+0.0007 [−0.0116, 0.0133]

F's own SimRef gain over the baseline grows as the face shrinks (+0.0418 at 1.00, +0.0621 at 0.50; scale interaction +0.0212 [0.0069, 0.0350]). At 0.50, F matches W in identity and, unlike W, does not raise similarity to the wrong identity (SimMis).

(c) Attention-mass matching at each scale: a whole-reference bias matched to F's injected mass shows no benefit over F on any metric (SimRef, SimGT, SimMis, CP, Aes, CLIP-T) at any of the three scales. Full tables are in the supplementary comment below.

(d) Face-local routing supports a higher strength:

Qwen, F(α=2.5) − W(α=1.5)ΔSimRefΔSimGTΔAesΔCLIP-TΔCP
P2+0.0295 [0.0225, 0.0369]+0.0153 [0.0108, 0.0199]+0.0196 [−0.0160, 0.0556]+0.0010 [−0.0005, 0.0025]+0.0292 [0.0220, 0.0367]
P3+0.0396 [0.0212, 0.0616]+0.0141 [0.0023, 0.0255]+0.0035 [−0.0796, 0.0857]−0.0039 [−0.0086, 0.0005]+0.0415 [0.0211, 0.0656]

F at α=2.5 exceeds W at α=1.5 in identity with aesthetics and prompt alignment at the same level.

Q2 / W2. Selecting scope and strength on unseen data. We recommend one fixed setting for all inputs: α=1.5, F by default, W when maximum identity is wanted. No per-image tuning is used anywhere.

(i) Cross-split and cross-backbone: the rule "smallest swept α with mean ΔSimRef ≥ 0.04" is applied on one split and evaluated on the other; both directions select α=1.5.

Selection → evaluation (F−B)ΔSimRefΔSimGT
Qwen P2 → Qwen P3+0.0605 [0.0420, 0.0803]+0.0277 [0.0175, 0.0386]
Qwen P3 → Qwen P2+0.0474 [0.0373, 0.0580]+0.0220 [0.0157, 0.0286]
Qwen-selected → FLUX P2+0.0393 [0.0338, 0.0447]+0.0218 [0.0174, 0.0263]

(ii) New public five-person task: 50 groups built from MultiID-Bench P1 inputs, which are not used elsewhere in the paper; α and scope fixed before generation (FLUX.2-klein). SimGT is computed against each identity's P1 ground-truth image.

vs. baseline, 95% CIΔSimRefΔSimGTΔCLIP-T
CFG 1, F+0.0507 [0.0389, 0.0635]+0.0248 [0.0153, 0.0348]−0.0028 [−0.0056, 0.0004]
CFG 1, W+0.0786 [0.0491, 0.1128]+0.0331 [0.0156, 0.0539]−0.0055 [−0.0112, −0.0005]
CFG 4, F+0.0290 [0.0162, 0.0412]+0.0177 [0.0080, 0.0266]−0.0016 [−0.0035, 0.0001]
CFG 4, W+0.0411 [0.0233, 0.0582]+0.0213 [0.0086, 0.0335]−0.0025 [−0.0066, 0.0011]

(iii) In-house five-person set (not publicly released, not used for selection): SimRef 0.3755 → 0.4259 for F (+0.0504 [0.0397, 0.0608]) and 0.4307 for W (+0.0552 [0.0406, 0.0706]).

The setting chosen on P2/P3 transfers without re-tuning to another split, another backbone, a larger person count, and two datasets outside the selection benchmark, with prompt alignment preserved at CFG 4. CP and aesthetics remain reported next to every identity number, and α stays a single interpretable control. All five-person groups are shown in App. Z.6.

We hope this resolves your concerns and are happy to run further checks.

AuthorsOpenReview ↗

Response to Reviewer Pfbc (supplement): controlled face-scale and attention-mass results

This is the full table for points (b) and (c) of our response to Q1 / W1. On FLUX.2-klein P3, each reference is placed on a neutral canvas at 100/75/50% occupancy, so face size is changed on the same references, prompts and seed. All numbers are paired differences with 95% CI, on the same paired panels as Appendix Z.2.

(1) Equal strength (α=1.5), F − W at each face scale

OccupancyΔSimRefΔSimGTΔSimMis ↓ΔCP ↓ΔAes
1.00−0.0296 [−0.0520, −0.0117]−0.0122 [−0.0240, −0.0025]−0.0064 [−0.0110, −0.0022]−0.0222 [−0.0382, −0.0080]+0.1274 [0.0553, 0.2007]
0.75−0.0253 [−0.0447, −0.0052]−0.0097 [−0.0250, 0.0060]−0.0025 [−0.0093, 0.0050]−0.0199 [−0.0340, −0.0065]+0.1146 [0.0039, 0.2331]
0.50−0.0003 [−0.0180, 0.0169]−0.0019 [−0.0153, 0.0113]−0.0127 [−0.0232, −0.0039]+0.0007 [−0.0116, 0.0133]+0.0127 [−0.0799, 0.1019]

At full scale, F gives lower copy-paste preference, lower similarity to the wrong identity and higher aesthetics than W, for a smaller identity gain. As the face shrinks, W's identity edge disappears (F−W ΔSimRef −0.0296 at 1.00, −0.0003 at 0.50). At 0.50, F matches W on SimRef and SimGT, while W raises wrong-identity similarity over the baseline (+0.0125 [0.0051, 0.0204]) and F does not (−0.0002 [−0.0076, 0.0062]). F's own gain over the baseline grows as the face shrinks (ΔSimRef +0.0418 at 1.00, +0.0621 at 0.50; scale interaction +0.0212 [0.0069, 0.0350]). This is the controlled, same-reference face-size evidence requested in W1.

(2) Attention-mass matched, D − F512 at each face scale

D is a whole-reference bias whose injected reference attention mass is matched, per query, head and reference, to that of FaceRoute-F; F512 is FaceRoute-F (α=1.5) run through the same attention implementation as D, so the two arms differ only in where the bias is placed.

OccupancyΔSimRefΔSimGTΔSimMisΔCPΔAesΔCLIP-T
1.00−0.0009 [−0.0046, 0.0027]−0.0014 [−0.0046, 0.0018]−0.0001 [−0.0033, 0.0038]−0.0008 [−0.0058, 0.0043]+0.0102 [−0.0180, 0.0403]+0.0007 [−0.0006, 0.0020]
0.75−0.0005 [−0.0069, 0.0055]−0.0008 [−0.0045, 0.0029]+0.0002 [−0.0026, 0.0035]−0.0007 [−0.0075, 0.0052]−0.0586 [−0.1094, −0.0156]+0.0002 [−0.0017, 0.0020]
0.50−0.0028 [−0.0099, 0.0043]−0.0001 [−0.0054, 0.0049]−0.0014 [−0.0042, 0.0011]−0.0011 [−0.0081, 0.0051]−0.0243 [−0.0394, −0.0081]−0.0013 [−0.0030, 0.0002]

With the injected mass equal, spreading it over the whole reference gives no benefit on any metric at any face scale, and is lower in aesthetics at 0.75 and 0.50. The equal-α gap between W and F therefore comes from the larger attention mass that W injects, and face-local routing reaches the same identity with far fewer tokens. The SimGT and CP columns are those of Appendix Z.2; the other columns extend the same analysis to all metrics.

AuthorsOpenReview ↗

Response to Reviewer Pfbc (supplement): fixed-setting validation on MultiHuman-Testbench

This adds to our response to Q2 / W2. Following Reviewer jD5K's request, we ran the same fixed setting (FLUX.2-klein, α=1.5, no re-tuning) on all 3/4/5-person prompts of the public MultiHuman-Testbench (NeurIPS 2025, arXiv:2506.20879): 1,125 groups, 3,375 images. This benchmark was not used for any selection.

Setup: 375 groups each for 3/4/5 people, CFG=4, official references and prompts, matched seeds. SimRef averages three face encoders (our evaluation protocol, no IQR filtering).

PeopleArmSimRef ↑CLIP-T ↑Aes ↑
3B0.40590.32126.5867
3F0.44300.32066.5888
3W0.47350.31996.5733
4B0.34880.32336.5623
4F0.37930.32346.5453
4W0.40860.32246.5278
5B0.32250.32186.6064
5F0.35730.32146.5858
5W0.39460.32116.5669

Pooled paired differences (1,125 groups), 95% CIs:

MetricF−BW−BF−W
SimRef+0.0341 [0.0314, 0.0368]+0.0665 [0.0630, 0.0700]−0.0324 [−0.0356, −0.0293]
CLIP-T−0.0003 [−0.0008, 0.0001]−0.0010 [−0.0015, −0.0004]+0.0007 [0.0002, 0.0012]
Aes−0.0118 [−0.0205, −0.0032]−0.0291 [−0.0413, −0.0169]+0.0173 [0.0061, 0.0284]

Both scopes improve identity at every person count (all paired CIs exclude zero). W gains more; F keeps higher Aes and CLIP-T than W. The operating point chosen on P2/P3 therefore transfers unchanged to an independent public benchmark, with the same F/W trade-off as in the paper.

Reviewer_jD5KOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper proposes face routing in the attention space.

  • This is a training free enhancement to the attention matrix which conditions the output generation more heavily on input images. It is done by adding a bias term to attention weights contributed to by the input image.

  • There are two proposed variants:

    (1) FaceRoute-W which boosts attention matrix using all the tokens from input images (2) FaceRoute-F which finds the face using detection and only boosts those token's attention scores.

  • The paper demnostrates the effect of these methods on two families of architectures on MultiID-Bench and XVerseBench.

Critical strengths
  • Within the limited scope of experiments, the method is simple, cost-effective and also effective in terms of results on recent SOTA generation models. This is a major strength as it means the proposed algorithm could have high impact to the community.

  • The authors have shared code to reproduce their work, which demonstrates their commitment and transparency in making the method reproducible, which could have high impact to the community.

Critical weaknesses

Major:

[1] A major weakness of the work is that the scope of experiments is not vast enough. MultiID-Bench and XVerseBench are very limited in terms of their data. For instance, MultiID-Bench has mostly 2-person samples and only 33 samples of 3 and 4 faces. A recent benchmark from NeurIPS25 <1> has greater coverage on 3,4,5 people scenes. To demonstrate the effectiveness of their approach on multi-face setting, it is imperative that the authors show a larger scope in terms of generalization to 3,4,5 face scenes.

[2] The proposed method is simple to a point that makes it incremental. Apart from being analogous to CFG, the mathematical algorithm of modulating conditioning strength at inference time is very similar to works such as <2,3> which bias the attention(input/output) space for different input conditions. Due to the lack of discussion and comparison with previous methods which proposed updates to the attention space (e.g. some ablation of the base model, base model+Faceroute based attention manipulation, base model+ method A style attention manipulation), the claims are not strong enough.

Minor:

[1] Figure 2 is messy/hard to understand. It needs to be formatted better(even text sizes, etc.) to make it more legible to readers.

[2] Lot's of related work(discussion or empirical comparisons) on multi-human generation missing <1,4,5,6>.

<1> https://arxiv.org/pdf/2506.20879

<2> https://arxiv.org/pdf/2411.02395

<3> https://arxiv.org/pdf/2510.24657

<4> https://arxiv.org/abs/2511.21691

<5> https://arxiv.org/abs/2511.22690

<6> https://arxiv.org/abs/2510.20820

Questions for authors

[1] Did the authors ablate comparing FaceRoute performance with varying input CFG scale for each image?

[2] Instead of scaling attention logits, have the authors tried scaling the input Key/Values and ablating against this?

[3] Wouldn't biasing lead to more copy paste artifacts? Is there a way to quantify this or ablate this?

6 comments
AuthorsOpenReview ↗

Response to Reviewer jD5K

Thank you for the detailed review. We ran every comparison you asked for; details are in Appendix Z.

Q1. CFG ablation; Major 1. Five-person scenes. We built 50 five-person groups from public MultiID-Bench P1 inputs (not used elsewhere in the paper; separate ground-truth photos per person) and ran FLUX.2-klein at CFG 1 and 4 with α=1.5 fixed in advance:

CFGMethodSimRef ↑SimGT ↑CLIP-T ↑
1Baseline0.32520.25050.2659
1FaceRoute-F0.3759 (+0.0507 [0.0389, 0.0635])0.2753 (+0.0248 [0.0153, 0.0348])0.2631
1FaceRoute-W0.4038 (+0.0786 [0.0491, 0.1128])0.2836 (+0.0331 [0.0156, 0.0539])0.2603
4Baseline0.42720.30450.2586
4FaceRoute-F0.4561 (+0.0290 [0.0162, 0.0412])0.3222 (+0.0177 [0.0080, 0.0266])0.2570
4FaceRoute-W0.4683 (+0.0411 [0.0233, 0.0582])0.3258 (+0.0213 [0.0086, 0.0335])0.2561

CFG and FaceRoute are complementary: CFG lifts the baseline and FaceRoute adds a further gain at both settings. On Qwen P2, F−B ΔSimGT is +0.0384 [0.0231, 0.0558] at CFG 4 and +0.0281 [−0.0072, 0.0742] at CFG 1 (smaller subset). An in-house five-person set (not publicly released) shows the same gain: SimRef 0.3755 → 0.4259 (F) and 0.4307 (W). With the existing P3 results, FaceRoute improves identity from two to five people with one fixed setting. All 50 groups are shown in App. Z.6; we also cite MultiHuman-Testbench.

Q2. Scaling Keys/Values; Major 2. Comparison with attention-space methods. Same frozen Qwen backbone, inputs and seeds (P3; paired 95% CI vs. baseline):

ControlSimGT ↑Aes ↑ΔSimGT
Baseline0.31135.4669—
FaceRoute-F, α=1.50.35185.4375+0.0385 [0.0225, 0.0556]
FaceRoute-W, α=1.50.34985.3461+0.0371 [0.0202, 0.0571]
Reference K ×0.80.07425.6619−0.2434 [−0.2984, −0.1898]
Reference K ×1.250.11993.8016−0.2066 [−0.2835, −0.1290]
Reference V ×0.80.28265.4716−0.0277 [−0.0499, −0.0058]
Reference V ×1.250.28745.5123−0.0208 [−0.0378, −0.0028]
GRAG-style d=0.80.08615.6165−0.2436 [−0.3011, −0.1849]
GRAG-style d=1.20.25514.3693−0.0484 [−0.0881, −0.0099]

GRAG-style d=1.5 destroys the image (Aes 3.77, CLIP-T 0.195; the requested person count is never produced). On FLUX.2-klein P3 (ΔSimGT): FaceRoute-F +0.0362 [0.0273, 0.0451], FaceRoute-W +0.0570 [0.0434, 0.0739]; temperature τ=0.8 −0.0286, τ=0.9 −0.0101; reference area ×0.5 −0.0383, ×0.25 −0.0734; identity-emphasis prompt −0.0135.

FaceRoute is the only tested control that improves identity while keeping image quality. Scaling K multiplies content-dependent logits and breaks the pretrained matching; scaling V does not change how much the target attends to the references. A constant bias shifts attention mass toward the references and leaves their relative logits intact.

Relation to prior work (now in Section 2): CFG acts on the output prediction and cannot choose which reference tokens are read; Regional Prompting assigns text to regions with masks and needs a layout; GRAG rescales keys around their group mean, which changes the attention pattern. FaceRoute needs one scalar, no training, no layout and no extra branch, and transfers across two architectures. The equally simple alternatives above do not work, so the specific design matters.

Q3. Copy-paste. The paper quantifies it. CP rises moderately (P3, F: +0.0288 on FLUX, +0.0500 on Qwen), but SimGT, measured against photos the model never sees, rises as well, which pure copying would not produce. The geometric audit (App. N) finds no resolved shift toward the reference's landmarks (+0.0121 [−0.0036, 0.0290]) or head pose (+0.0136 [−0.0018, 0.0318]). The cost is controllable by scope (F has lower CP than W at the same α: Qwen +0.0500 vs. +0.0784) and by strength (the α sweep).

Minor 1. Figure 2 is redrawn with uniform text sizes and a cleaner layout.

Minor 2. All six references are now cited in Section 2.

We hope this addresses your concerns and are happy to provide anything further.

Reviewer_jD5KOpenReview ↗

Follow-ups

Dear authors, thank you for your response. I had follow-ups as they did not fully address my concerns.

  1. For the five-person split, what was the rationale to curate your own dataset? Instead of curating your own dataset I would suggest to use an already publicly available benchmark <1> from earlier comment, as it has been constructed for this specific purpose of multi-person generation. The number of samples is simply not enough and the face coverage over demographics is restricted in MultiID-bench.
  2. While the authors mention they are complementary, what is the theoretical difference between proposed and CFG?
  3. While there was a CFG sweep over two values(1 and 4), this ablation is not enough to justify faceroute performance. This should be in a plot, with multiple values of CFG v/s faceroute.
  4. Empirical comparisons on related works(Ar2can, UMO, ID-Patch, Dreamo) are suggested for future, as these works target to solve the same problem as faceroute.
  5. I would also hesitate from the training-free and training-based tags. Training based tag is usually for methods which are trained using low-rank adaptation or other methods for personalization. FLux.2 and Qwen were both trained with personalization labels, like any other method. You can use the term generalist and specialist, but using inference based approaches, faceroute is also making these models specialists in one task.
AuthorsOpenReview ↗

Follow-up response to Reviewer jD5K: MultiHuman-Testbench and CFG

Thank you for the follow-up. We ran both experiments you asked for.

1. MultiHuman-Testbench. All 375 groups each for 3/4/5 people (3,375 images); FLUX.2-klein, CFG=4, α=1.5, official references and prompts, matched seeds. SimRef averages three face encoders (our evaluation protocol, no IQR filtering).

PeopleArmSimRef ↑CLIP-T ↑Aes ↑Count ↑
3B0.40590.32126.586784.27%
3F0.44300.32066.588884.00%
3W0.47350.31996.573385.07%
4B0.34880.32336.562380.27%
4F0.37930.32346.545380.00%
4W0.40860.32246.527879.73%
5B0.32250.32186.606477.60%
5F0.35730.32146.585875.73%
5W0.39460.32116.566977.33%

Pooled paired differences (1,125 groups), 95% CIs:

MetricF−BW−BF−W
SimRef+0.0341 [0.0314, 0.0368]+0.0665 [0.0630, 0.0700]−0.0324 [−0.0356, −0.0293]
CLIP-T−0.0003 [−0.0008, 0.0001]−0.0010 [−0.0015, −0.0004]+0.0007 [0.0002, 0.0012]
Aes−0.0118 [−0.0205, −0.0032]−0.0291 [−0.0413, −0.0169]+0.0173 [0.0061, 0.0284]

Both scopes improve identity at every person count (all paired CIs exclude zero). W gains more; F keeps higher Aes and CLIP-T than W. Count differences from B include zero. The baseline shows identity dilution (SimRef 0.406 → 0.349 → 0.323 from three to five people); at five people, F (0.357) and W (0.395) both exceed the baseline at four. Our own five-person groups simply extend MultiID-Bench, the benchmark of our main results, to five people.

2. Difference from CFG. CFG mixes two output predictions with one global scalar, v = (1 − s)·u + s·c. In FLUX.2-klein both passes see the same reference tokens and differ only in text, so s scales only what the text changes and cannot choose which reference tokens are read. FaceRoute acts inside the forward pass: adding log α to target-to-reference logits turns reference attention mass m into m′ = αm / (αm + 1 − m), with all ratios inside and outside the references unchanged. This is the minimum-KL change to the pretrained attention for that mass. It is selective (whole references or face tokens only), keeps each row a probability distribution, and needs no extra forward pass. The two are complementary.

3. CFG sweep. Plotted in the revised appendix (Figure 15, App. Z.8). Given limited compute, it uses our 50 five-person P1 groups: FLUX.2-klein, CFG 1–6, α=1.5, same seeds.

CFGArmSimRef ↑CLIP-T ↑Aes ↑
1B0.32520.26596.179
1F0.37590.26316.150
1W0.40380.26035.736
2B0.40380.26166.362
2F0.45990.25906.330
2W0.48460.25606.230
3B0.41890.25886.472
3F0.46830.25936.418
3W0.48180.25706.268
4B0.42720.25866.470
4F0.45610.25706.407
4W0.46830.25616.257
5B0.41740.25836.443
5F0.45860.25746.399
5W0.46620.25596.306
6B0.41510.25716.450
6F0.45190.25616.412
6W0.46530.25596.313

F and W raise SimRef at every CFG value (all 12 paired 95% CIs exclude zero). CFG alone cannot reproduce the gain: the best baseline SimRef is 0.4272 (CFG 4), and FaceRoute-F exceeds it at every CFG ≥ 2, with CLIP-T within 0.006 and Aes within 0.07 of the baseline.

Comments on the paper

1 comment

General response: summary of the revision

We sincerely thank all reviewers for their constructive feedback. We ran the requested experiments and uploaded a revised PDF. The method and all original results are unchanged; each reviewer's points are answered in the individual responses below.

Main text (within the page limit)

  • Figure 2 is redrawn for legibility (jD5K).
  • Section 2 adds the suggested related work on multi-human generation and attention-space guidance (jD5K).
  • Sections 1 and 7: one wording refinement on the small-face result, and a closing paragraph in Section 7 pointing to the new controls.
  • Section 5: one pointer sentence to the new appendix.

New Appendix Z

  • Z.1 Same-backbone reference-strength controls (1JUc, jD5K): attention temperature, reference resolution, identity-emphasis prompt, K/V scaling and GRAG-style modulation, on one (CP, SimGT) plane with FaceRoute.
  • Z.2 Small-face and same-reference scale controls (Pfbc): unfiltered paired intervals, a same-reference face-scale control, and attention-mass matching at each scale.
  • Z.3 Strength selection (Pfbc): cross-split and cross-backbone transfer of the fixed α=1.5.
  • Z.4 Five-person groups (jD5K, Pfbc): 50 new groups from public MultiID-Bench P1 inputs, at CFG 1 and CFG 4.
  • Z.5 WithAnyone (1JUc): SigLIP-branch adaptation.
  • Z.6–Z.7 Galleries: all 50 five-person groups and all 49 small-face examples.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.