acceptodds
Under review as a conference paper at ICLR 2027

FreeID: Calibrating Attention for Identity Preservation in Image-conditioned T2I Models

Abstract

Text-to-Image (T2I) diffusion models have driven rapid progress in personalized image generation, yet preserving subject identity from a reference image remains a central challenge. Conventional pipelines add identity through per-subject optimization or learned identity-conditioning modules trained on large-scale data. More recently, image-conditioned T2I (IT2I) models condition on reference images natively, processing text, reference images, and generation tokens within a unified attention framework. Yet this native conditioning still fails in demanding cases such as cross-view generation or small-scale subjects. Our analysis traces this failure to a specific mechanistic gap: IT2I models identify which reference tokens to attend to, but lack a mechanism to calibrate how much attention identity-critical features should receive, under-weighting them relative to peripheral ones; boosting facial attention alone recovers identity. Motivated by findings from human face recognition that internal facial regions carry disproportionate identity weight, we propose FreeID, a training-free framework that closes this gap at inference time. FreeID exploits the model's own attention responses to match reference facial keypoint tokens to their corresponding generation tokens, then injects sparse attention biases that amplify these identity-critical links, with magnitudes set by a closed-form bound on attention-output drift that adapts per token and step. Applied to state-of-the-art IT2I models, FreeID improves face similarity by up to 17.2% with negligible quality degradation, and further improves learned identity modules and image-to-video (I2V) models.

Primary area:
generative models
Keywords:
Text-to-Image (T2I) diffusion models, Identity Preservation, Calibrating Attention

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_eFVtOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes FreeID, a training-free method to improve identity preservation in image-conditioned T2I (IT2I) models such as FLUX.2. The authors observe that IT2I models know where to attend in the reference image, but allocate less attention to identity-critical facial features than to peripheral regions such as hair and clothing, and argue that this causes identity drift in cross-view and small-face generation.

FreeID matches reference facial keypoints to generation tokens using the model's own attention maps, and injects sparse attention biases whose magnitudes are bounded by the induced attention-output drift. Experiments on FLUX.2 and LongCat show improved InsightFace and AdaFace similarity with little change in CLIP-IQA and CLIP-T.

Critical strengths
  • The paper studies a practical problem, especially the "cross-view generation problem". Native IT2I models are becoming mainstream, and the observation that they "know where to attend, but not how much" is interesting.
  • The method is training-free and plug-and-play, with acceptable overhead (+15.5% latency, no extra memory), and it can also be applied on top of PuLID, DreamO, and I2V models.
Critical weaknesses
  • Insufficient evaluation. The results are backed by evaluations on FLUX.2 and LongCat on a test set constructed from 25 samples from Unsplash and 25 samples from CelebA. It is not enough to persuade me by testing 2 models on simple testset, and it is hard to reproduce the results based on the information revealed. There are plenty of benchmarks like MultiID-Bench [1], MultiHuman-TestBench [2], and plenty of models the community may be interested in like Qwen-Image-Edit [3], and adapter-based methods other than PuLID and DreamO like InfiniteYou [4], UMO [5], and WithAnyone [1].
  • Not enough results to back the claims.
    • In the qualitative results shown in the appendix, there are only a few cross-view generation cases. Considering cherry-picking is a widely adopted practice, it is hard to back the claims about "cross-view generation" in the abstract.
    • What's worse, I noticed that in Table 3, InsightFace similarity is improved to above 0.8. According to my practical experience, for InsightFace, only faces with the same view/angle and expression can reach such a high similarity, and cross-view similarity is only about 0.5 or even lower. This, on the one hand, may be caused by the simplicity of the CelebA test set. On the other hand, it is possible that FreeID even makes the cross-view generation problem worse.

[1] WithAnyone: Towards Controllable and ID-Consistent Image Generation

[2] MultiHuman-Testbench: Benchmarking Image Generation for Multiple Humans

[3] Qwen-Image Technical Report

[4] InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity

[5] UMO: Scaling Multi-Identity Consistency for Image Customization via Matching Reward

Questions for authors
  • How are the prompts of your test set constructed?
  • Is there more evidence showing the correlation between the cross-view generation problem and disproportionate attention?
5 comments
AuthorsOpenReview ↗

Rebuttal by Authors #1

We thank the reviewer for the constructive feedback, and for finding the cross-view generation problem practical, our observation that IT2I models "know where to attend, but not how much" interesting, and our training-free, plug-and-play design acceptable in overhead. We address each concern below.

[W1] "Insufficient evaluation. The results are backed by evaluations on FLUX.2 and LongCat on a test set constructed from 25 samples from Unsplash and 25 samples from CelebA. It is not enough to persuade me by testing 2 models on simple testset, and it is hard to reproduce the results based on the information revealed. There are plenty of benchmarks like MultiID-Bench [1], MultiHuman-TestBench [2], and plenty of models the community may be interested in like Qwen-Image-Edit [3], and adapter-based methods other than PuLID and DreamO like InfiniteYou [4], UMO [5], and WithAnyone [1]."

Following the reviewer's suggestion, we evaluated FreeID on MultiID-Bench [1] and MultiHuman-TestBench [2], and on Qwen-Image-Edit [3], InfiniteYou [4], and UMO [5]. FreeID improves identity similarity for every method on both benchmarks (Tables R1 and R2) and on both of our test sets (Table R3), while the other metrics stay close to the baseline.

With these results, FreeID is tested on six IT2I backbones (FLUX.2-klein, FLUX.2-dev, LongCat-Image-Edit-Turbo, Qwen-Image-Edit, OmniGen v1, and UNO) and four adapter-based methods (PuLID, DreamO, InfiniteYou, and UMO). The original submission also includes a larger test set of 200 Unsplash identities × 10 prompts (2,000 images; Sec. 5.2 and Appendix D.2), where the gain holds across viewpoint, face scale, and reference resolution, with positive 95% confidence intervals (Table A6).

For reproducibility, Sec. 4.2 now states when the bias is applied. Appendix B lists the settings for the three main backbones (denoising steps, matching block, r, and τ in Table A1) and all prompts (Table A2).

Table R1. Evaluation on MultiID-Bench. 206 single-person, 194 two-person, 19 three-person, and 14 four-person cases. Sim(GT) and Sim(Ref) are ArcFace cosine similarities between the generated face and the ground-truth face or the reference face, respectively.

MethodSim(GT) ↑Sim(Ref) ↑CLIP-I ↑CLIP-T ↑
FLUX.2-klein0.2190.3090.7490.329
+ FreeID0.2810.4010.7570.328
UMO0.3090.4720.7450.317
+ FreeID0.3420.5280.7490.317
InfiniteYou0.2700.3960.7210.331
+ FreeID0.2740.4030.7220.331

Table R2. Evaluation on MultiHuman-TestBench. Results are averaged over a fixed 500-case subset with 100 cases each for 1–5 people. The single-person group contains 100 Simple cases; each remaining group contains 80 Simple and 20 Complex cases (420 Simple and 80 Complex in total). ID similarity is the ArcFace cosine similarity between reference and generated faces. HPS is the Human Preference Score.

MethodID similarity ↑HPS ↑
FLUX.2-klein0.2700.305
+ FreeID0.3230.304
UMO0.2820.261
+ FreeID0.3230.259
InfiniteYou0.1890.312
+ FreeID0.1910.312

Table R3. Evaluation on Unsplash and CelebA-HQ. Each dataset contains 25 reference identities with 10 prompts per identity, yielding 250 single-person images per configuration.

DatasetMethodInsightFace ↑AdaFace ↑CLIP-IQA ↑CLIP-T ↑
UnsplashQwen-Image-Edit0.6580.6400.5400.234
+ FreeID0.7170.7020.5330.233
InfiniteYou0.6440.6260.5220.241
+ FreeID0.6500.6330.5220.241
UMO0.6160.5830.5520.218
+ FreeID0.6700.6370.5510.218
CelebA-HQQwen-Image-Edit0.5570.5360.4850.234
+ FreeID0.6670.6450.4740.234
InfiniteYou0.5900.5800.5330.233
+ FreeID0.5950.5860.5330.232
UMO0.5930.5740.5380.216
+ FreeID0.6420.6230.5360.216

We added these results to Appendix D.14 (Tables A17 and A18) and Appendix D.15 (Table A19).

AuthorsOpenReview ↗

Rebuttal by Authors #2

[W2] "Not enough results to back the claims. In the qualitative results shown in the appendix, there are only a few cross-view generation cases. Considering cherry-picking is a widely adopted practice, it is hard to back the claims about "cross-view generation" in the abstract."

The cross-view claim does not rest on cherry-picked samples. Over all 2,000 paired generations, with no sample removed, FreeID improves identity at every degree of view change, and the gain is largest under large view changes.

We grouped the 2,000 paired generations (200 identities × 10 prompts, FLUX.2-klein) by the head-yaw difference between the generated face and the reference (Table R4). In every group, FreeID scores higher than the baseline in 89–94% of pairs.

Table R4. Identity by view change.

View change (yaw)# ImagesInsightFace (Baseline)InsightFace (+ FreeID)Δ InsightFace [95% CI]
< 15° (same view)13490.5670.642 (+13%)+0.075 [0.072, 0.079]
15–30°2290.4670.550 (+18%)+0.083 [0.073, 0.093]
30–45°1250.3490.474 (+36%)+0.126 [0.110, 0.144]
≥ 45° (cross-view)2970.3470.436 (+25%)+0.088 [0.079, 0.098]

We also added more cross-view qualitative cases, without selection. Fig. A12 shows uncurated results on FLUX.2-klein for three subjects × five seeds, each annotated with its InsightFace similarity and measured head yaw. FreeID improves identity in every pair, while the head yaw stays close to the baseline.

We added these results to Appendix D.11 (Table A13) and Fig. A12.

[W3] "What's worse, I noticed that in Table 3, InsightFace similarity is improved to above 0.8. According to my practical experience, for InsightFace, only faces with the same view/angle and expression can reach such a high similarity, and cross-view similarity is only about 0.5 or even lower. This, on the one hand, may be caused by the simplicity of the CelebA test set. On the other hand, it is possible that FreeID even makes the cross-view generation problem worse."

FreeID does not make the cross-view generation problem worse. Its gain is largest under cross-view (Table R4), and it does not reduce how far the head turns.

Our results agree with the reviewer's experience. Cross-view similarity is low: at ≥ 45°, the baseline scores 0.347 and FreeID 0.436. Scores above 0.8 come mostly from same-view, large faces, which make up most of the test samples (1,349 of 2,000 pairs are within 15°). They do not come from the CelebA test set: every baseline model scores lower on CelebA than on Unsplash (Table 1).

FreeID also does not gain by copying the reference pose. On 175 identities × 10 prompts (FLUX.2-klein), identity improves as τ grows, while for the side-profile and three-quarter prompts the head turns as much as in the baseline (Table R5).

Table R5. Identity and head turn across τ. Head turn is how far the head turns for the side-profile and three-quarter prompts.

τInsightFaceHead turn
0 (baseline)0.50945.1°
0.0050.55444.7°
0.010.57344.6°
0.02 (default)0.59244.6°
0.050.61044.6°
0.10.61444.4°
0.20.58543.6°

We added this analysis to Appendix D.11, and the τ sweep to Appendix D.12 (Table A14).

AuthorsOpenReview ↗

Rebuttal by Authors #3

[Q1] "How are the prompts of your test set constructed?"

The 10 prompts (Table A2) are constructed to cover four types of edits that commonly challenge identity preservation, and the same prompts are used for every identity, model, and method.

  • Expression (2): laughing, serious expression
  • Action (3): talking on a phone, drinking coffee, kicking a soccer ball
  • View angle (3): side profile, three-quarter view, bird's-eye view
  • Framing (2): full-body shots, standing and sitting, where the face appears small

The view-angle and framing prompts target the two settings our paper focuses on: cross-view generation and small faces. The expression and action prompts check that the model still follows edits beyond the reference. We added this description to Appendix B.4.

[Q2] "Is there more evidence showing the correlation between the cross-view generation problem and disproportionate attention?"

Yes, under large view changes, the generated face gives less attention to the reference face and loses identity (Table R6). FreeID restores this attention, and identity recovers with it.

We measured attention on 1,750 paired generations (175 identities × 10 prompts, same seeds, FLUX.2-klein). Face share is how much of the generated face's attention to the reference image goes to the reference face.

Table R6. Face share by view change.

View change (yaw)# ImagesFace share (Baseline)Face share (+ FreeID)
< 15° (same view)11900.720.83
15–30°1970.730.85
30–45°1040.730.91
≥ 45° (cross-view)2590.620.83

Beyond 45°, the baseline's face share drops, and FreeID raises it above the same-view baseline level. Identity by view change is in Table R4.

To control for other factors, we also compare samples of the same person, prompt, and face size. All three 95% confidence intervals exclude zero (Table R7).

Table R7. Correlations between face share, view change, and identity.

Relationshipr [95% CI]Meaning
Face share ↔ view change−0.32 [−0.40, −0.24]The more the head turns, the less attention goes to the face
Face share ↔ InsightFace (view also controlled)+0.33 [0.27, 0.37]Less face attention, lower identity
FreeID's Δ face share ↔ Δ InsightFace+0.29 [0.23, 0.35]A larger attention gain goes with a larger identity gain

We added this analysis to Appendix E.3 (Tables A20 and A21).

[Ethics] "Test samples from Unsplash may cause some privacy issue."

The Unsplash references are photos that their photographers shared publicly under the Unsplash License, and Unsplash portraits are a common test set in identity-preservation work (e.g., Unsplash-50, used by LCM-Lookahead [6] and PuLID [7]). We use them only for evaluation: no model is trained on them, and we will not release the dataset.

[1] Xu et al., WithAnyone: Toward Controllable and ID Consistent Image Generation. ICLR, 2026.
[2] Borse et al., MultiHuman-Testbench: Benchmarking Image Generation for Multiple Humans. NeurIPS, 2025.
[3] Wu et al., Qwen-Image Technical Report. arXiv, 2025.
[4] Jiang et al., InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity. ICCV, 2025.
[5] Cheng et al., Scaling Multi-Identity Consistency for Image Customization via Multi-to-Multi Matching Paradigm. CVPR, 2026.
[6] Gal et al., LCM-Lookahead for Encoder-based Text-to-Image Personalization. ECCV, 2024.
[7] Guo et al., PuLID: Pure and Lightning ID Customization via Contrastive Alignment. NeurIPS, 2024.

Reviewer_SLydOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Motivated by the fact that IT2I models have already exhibited strong semantic alignment between reference tokens and generation token. This paper introduces FreeID, a training-free method that supports identity-preserving image generation in TI2I task via attention calibration. It works on mainstream DiT-based models with negligible quality degradation.

Critical strengths
  • They provide insightful analysis of attention allocation in TI2I generation. It demonstrates that generation tokens have already exhibit semantic alignment with reference tokens, while attention is disproportionately allocated to peripheral regions rather than identity-critical facial feature that leads to identity preservation failures.
  • FreeID is a simple and effective training-free framework that can be applied to most of current MMDiT-based models without any identity-specific training data. Furthermore, it is compatible with existing T2I-based personalization methods as shown in Table 3.
  • Both the qualitative and quantitative experiments look promising.
Critical weaknesses
  • Experiments on other representative image editing models, such as Qwen-Image-Edit, would provide stronger evidence of the method's generalizability.
  • Limited evaluation of the trade-off between identity preservation and editability.
Questions for authors
  • Besides of open-sourced base, It would be interesting to compare with state-of-the-art close-sourced models (GPT-Image-2.5, Nano Banana 2.1, etc). These models have stronger reference capabilities. Whether stronger reference capabilities mean a higher correlation between generated tokens and reference tokens worths some discussion.
3 comments
AuthorsOpenReview ↗

Rebuttal by Authors #1

We thank the reviewer for the constructive feedback, and for finding our analysis of attention allocation insightful, our training-free framework simple, effective, and compatible with existing personalization methods, and our experiments promising. We address each concern below.

[W1] "Experiments on other representative image editing models, such as Qwen-Image-Edit, would provide stronger evidence of the method's generalizability."

Following the reviewer's suggestion, we applied FreeID to Qwen-Image-Edit and FLUX.1-Kontext (Table R1). FreeID improves identity preservation on both models and both test sets, with the largest gains on Qwen-Image-Edit (up to +19.7% InsightFace and +20.3% AdaFace). Image quality drops slightly, and text alignment stays about the same. Each model is evaluated on Unsplash and CelebA-HQ with 25 identities × 10 prompts (250 paired comparisons) per dataset.

Table R1. FreeID on Qwen-Image-Edit and FLUX.1-Kontext.

DatasetMethodInsightFace ↑AdaFace ↑CLIP-IQA ↑CLIP-T ↑
UnsplashQwen-Image-Edit0.6580.6400.5400.234
+ FreeID0.717 (+9.0%)0.702 (+9.7%)0.533 (−1.3%)0.233 (−0.5%)
FLUX.1-Kontext0.7540.7440.6660.231
+ FreeID0.802 (+6.3%)0.794 (+6.7%)0.658 (−1.1%)0.232 (+0.3%)
CelebA-HQQwen-Image-Edit0.5570.5360.4850.234
+ FreeID0.667 (+19.7%)0.645 (+20.3%)0.474 (−2.4%)0.234 (0.0%)
FLUX.1-Kontext0.7380.7370.4610.224
+ FreeID0.807 (+9.4%)0.808 (+9.7%)0.440 (−4.6%)0.224 (+0.1%)

We added these results to Appendix D.15 (Table A19).

[W2] "Limited evaluation of the trade-off between identity preservation and editability."

As requested, we evaluated the trade-off between identity preservation and editability by sweeping τ (Table R2). At the default τ, identity improves by +16%, while the edit success rate stays almost the same (95.5% → 95.3%).

For each τ, we generated 175 identities × 10 prompts with FLUX.2-klein. Fig. 7 already covers identity vs. image quality, so here we measure editability in three ways:

  • Edit success: Qwen3-VL-8B judges whether the prompted edit (expression, action, viewpoint, composition) is applied.
  • Smile kept: how much of the baseline's smile remains for the "laughing" prompt.
  • Head turn: how far the head actually turns for side-profile and three-quarter prompts.

Table R2. Identity and editability across τ.

τInsightFaceEdit successSmile keptHead turn
0 (baseline)0.50995.5%100%45.1°
0.0050.554 (+9%)95.5%99%44.7°
0.010.573 (+13%)95.4%98%44.6°
0.02 (default)0.592 (+16%)95.3%96%44.6°
0.050.610 (+20%)94.7%93%44.6°
0.10.614 (+21%)94.4%86%44.4°
0.20.585 (+15%)94.2%70%43.6°

CLIP-T stays at about 0.226 at every τ. For 65 references with open-mouth smiles, a "serious expression" prompt also closes the mouth as well as the baseline at every τ. At larger τ, the trade-off between identity and expression becomes visible, and τ controls this balance.

We added these results to Sec. 5.3 and Appendix D.12 (Table A14).

AuthorsOpenReview ↗

Rebuttal by Authors #2

[Q1] "Besides of open-sourced base, It would be interesting to compare with state-of-the-art close-sourced models (GPT-Image-2.5, Nano Banana 2.1, etc). These models have stronger reference capabilities. Whether stronger reference capabilities mean a higher correlation between generated tokens and reference tokens worths some discussion."

FreeID cannot be applied to closed-source models, since it needs access to their attention. On open-weight models, stronger reference capability generally goes with a higher correlation between generated and reference tokens (Table R3).

GPT-Image-2.5 and Nano Banana 2.1 are available only through an API that returns images, so their attention cannot be read or biased. We state this in the Limitations (Appendix F). Fig. A6 compares GPT-5.3 with FLUX.2-klein + FreeID on non-celebrity subjects.

We measured the correlation on six open-weight IT2I models (Table R3).

Table R3. Reference capability and attention correspondence across IT2I backbones. Part correspondence is the fraction of attention from generated facial tokens (eyes, nose, and mouth) that goes to the matching reference regions. Intra-model ρ denotes the Spearman correlation between part correspondence and identity similarity across generated images within each model.

ModelBaseline InsightFace ↑Part correspondence (Baseline) ↑Part correspondence (+ FreeID) ↑Intra-model ρ ↑Δ InsightFace
UNO0.3380.7110.786+0.61+0.242
FLUX.2-klein0.4450.7790.810+0.30+0.108
OmniGen0.5720.4710.606+0.27+0.055
LongCat0.5920.7060.721+0.44+0.070
FLUX.2-dev0.6450.7830.814+0.54+0.036
FLUX.1-Kontext0.7860.9000.914+0.60+0.035

The trend across models is not strictly monotonic. Within each model, higher correspondence goes with better identity preservation (pooled ρ = 0.46), and FreeID improves both on every backbone.

We added this analysis to Appendix E.4 (Table A22).

Dear Reviewer SLyd,

Thank you again for your review. As the discussion period is closing, we wanted to check whether our response addressed your concerns.

We added Qwen-Image-Edit and FLUX.1-Kontext results, an identity-editability trade-off evaluation across τ, and a correlation analysis across six open-weight backbones in place of the closed-source comparison, which FreeID cannot be applied to. Details are in our rebuttal above.

If any part of your concern remains open, we would be glad to answer it while there is still time.

Best regards, The Authors

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

This paper addresses the challenge of single-subject customization in cross-view generation or small-scale subject scenarios. The proposed method not only identifies which reference tokens to focus on, but also leverages the model's own attention responses to match reference facial landmark tokens with their corresponding generation tokens. It then injects sparse attention biases to strengthen these identity-critical connections and adjust the attention magnitude. The method improves facial similarity across multiple T2I models and can also be extended to T2V tasks.

Critical strengths
  1. The proposed method is training-free and introduces no additional parameters. It can be applied not only to text-to-image generation but also to image-to-video generation, demonstrating strong generalizability.
  2. The paper is well-organized, with a clear logical flow and well-motivated methodology.
  3. The experiments are relatively comprehensive. Ablation studies are conducted on most of the parameters and components introduced in the paper. The supplementary material also includes comparisons with open-source identity customization methods.
Critical weaknesses
  1. In cross-view generation, the generated images seem to largely preserve the original facial expressions. If the prompt specifies a different facial expression, can the method still achieve good instruction following? Additionally, since FreeID places significant emphasis on facial identity, could it artificially boost identity similarity scores by limiting variations in pose or facial expression?
  2. According to Fig. 5, FLUX.2-klein and LongCat-Image-Edit-Turbo already demonstrate reasonably good identity preservation capabilities, while FreeID provides further improvements. I wonder whether FreeID remains effective when applied to weaker base models. Can it still achieve substantial improvements in identity preservation?
  3. The abstract and introduction mention that existing customization methods struggle with cross-view generation and small-scale subjects. However, the experiments primarily demonstrate improvements in cross-view generation, while quantitative and qualitative evaluations on small-scale subjects seem to be missing.
Questions for authors
  1. Fig. 4 presents the overall framework, but it is neither referenced nor discussed in sufficient detail throughout the paper. The top part of Fig. 4 seems to suggest that Attention Bias Injection is applied only at specific timesteps. However, I could not find any corresponding explanation or experimental analysis in the main text.
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors #1

We thank the reviewer for the constructive feedback, and for recognizing the training-free design of FreeID and its generalizability to image-to-video generation, the clear organization and motivation of the paper, and the comprehensiveness of our experiments and ablations. We address each concern below.

[W1] "In cross-view generation, the generated images seem to largely preserve the original facial expressions. If the prompt specifies a different facial expression, can the method still achieve good instruction following? Additionally, since FreeID places significant emphasis on facial identity, could it artificially boost identity similarity scores by limiting variations in pose or facial expression?"

Yes, FreeID still follows prompts that change the facial expression, and its identity gain does not come from limiting pose or expression. We extended the evaluation to 40 distinct identities with 10 expression-related prompts each (400 images) as shown in Table R1. In addition, our expanded evaluation of 200 identities (Appendix D.2) already includes two expression-related prompts (laughing with joy and serious expression), yielding 400 images in total.

Table R1. Identity preservation under facial expression changes.

EvaluationMethodInsightFace ↑AdaFace ↑CLIP-IQA ↑CLIP-T ↑
Extended (400)Baseline0.7100.7040.5520.207
+ FreeID0.752 (+5.9%)0.745 (+5.9%)0.552 (0.0%)0.203 (−1.6%)
Original (400)Baseline0.5530.5240.5320.225
+ FreeID0.626 (+13.2%)0.602 (+14.9%)0.523 (-1.7%)0.222 (-1.3%)

FreeID improves identity under expression changes while image quality and text alignment stay close to the baseline. It also does not limit pose or expression: at the default τ, the VLM-judged edit success rate is 95.3% (95.5% without FreeID), and the head turns as much as in the baseline (Appendix D.12, Table A14).

Fig. A13 shows that FreeID follows prompts that change the facial expression while keeping identity, and Fig. 5 and Figs. A1–A5 show the same under viewpoint and pose changes. We added these results to Appendix D.12 (Table A15 and Fig. A13).

[W2] "According to Fig. 5, FLUX.2-klein and LongCat-Image-Edit-Turbo already demonstrate reasonably good identity preservation capabilities, while FreeID provides further improvements. I wonder whether FreeID remains effective when applied to weaker base models. Can it still achieve substantial improvements in identity preservation?"

Yes, FreeID still gives a large gain on weaker base models. We added OmniGen v1 and UNO (FLUX.1-dev-based); UNO has the lowest baseline identity of all models we apply FreeID to, and FreeID improves it the most (InsightFace +71% on Unsplash and +83% on CelebA-HQ) as shown in Table R2.

Table R2. Evaluation on OmniGen v1 and UNO.

DatasetMethodInsightFace ↑AdaFace ↑CLIP-IQA ↑CLIP-T ↑
Unsplash (250)OmniGen v10.5640.5530.5800.224
+ FreeID0.623 (+10.5%)0.621 (+12.3%)0.579 (−0.1%)0.221 (−1.5%)
UNO0.3440.3170.5140.223
+ FreeID0.590 (+71.2%)0.571 (+80.2%)0.508 (−1.1%)0.218 (−2.4%)
CelebA-HQ (250)OmniGen v10.5200.5090.4310.213
+ FreeID0.578 (+11.1%)0.570 (+11.8%)0.398 (−7.7%)0.210 (−1.6%)
UNO0.3020.2870.4880.220
+ FreeID0.550 (+82.5%)0.546 (+90.3%)0.485 (−0.6%)0.214 (−2.6%)

FreeID improves identity on both models, with the largest gain on UNO. This suggests that FreeID helps most when the base model preserves identity poorly. We added these results to Appendix D.13 (Table A16 and Fig. A14).

AuthorsOpenReview ↗

Rebuttal by Authors #2

[W3] "The abstract and introduction mention that existing customization methods struggle with cross-view generation and small-scale subjects. However, the experiments primarily demonstrate improvements in cross-view generation, while quantitative and qualitative evaluations on small-scale subjects seem to be missing."

As requested, we added a quantitative evaluation on small-scale subjects (Table R3). FreeID's gain is largest for small faces: +30%, compared with +13% for large faces. Qualitative examples are in Fig. 3 ("Small-scale subject"), Fig. 5, and Figs. A1–A5.

Table A6 grouped samples by face size in the reference image. Here we group the same 2,000 paired generations (200 identities × 10 prompts, FLUX.2-klein) by face size in the generated image.

Table R3. Identity by face size in the generated image. InsightFace similarity. 95% confidence intervals are computed with an identity-level cluster bootstrap.

Face size in generated image# ImagesBaseline+ FreeIDΔ [95% CI]
Small (< 1%)4790.3870.504 (+30%)+0.117 [0.109, 0.126]
Medium (1–4%)1220.5210.602 (+16%)+0.082 [0.069, 0.094]
Large (≥ 4%)13990.5500.619 (+13%)+0.069 [0.065, 0.072]
Overall20000.5090.591 (+16%)+0.081

A small face is carried by only a few tokens, so strengthening their attention to the reference face helps the most. We added these results to Appendix D.10 (Table A12).

[Q1] "Fig. 4 presents the overall framework, but it is neither referenced nor discussed in sufficient detail throughout the paper. The top part of Fig. 4 seems to suggest that Attention Bias Injection is applied only at specific timesteps. However, I could not find any corresponding explanation or experimental analysis in the main text."

Attention Bias Injection is applied at every denoising step except the first, which is used only to find the face tokens.

  1. First step, no bias. From the attention at the middle block, we find the generation tokens that match the reference eyes, nose, and mouth (Sec. 4.1). Bias cannot be applied before this, since face token matching relies on the model's own attention.
  2. All remaining steps, bias applied. We inject the adaptive bias into the face tokens found at the previous step (Sec. 4.2), updating them at every step from the new attention.

We added this explanation to Sec. 4.2, with a reference to Fig. 4.

Thank you for your detailed response and for conducting the additional experiments. Your responses have largely addressed the concerns I raised, and I am inclined to maintain my current rating.

Dear Reviewer yjss,

Thank you for reading our response and for your positive assessment. We are glad the additional experiments addressed your concerns. If anything else comes up before the discussion closes, we are happy to answer.

Best regards, The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the Area Chair and all reviewers for their constructive feedback. We briefly summarize the reviewers' noted strengths, their concerns, and our responses, with pointers to the discussion threads.

Strengths noted by the reviewers

  1. An insightful analysis of attention misallocation underlying identity preservation failures, particularly in cross-view generation [SLyd, eFVt].
  2. A simple, training-free, plug-and-play method with broad applicability across image generation, personalization, and video generation models [yjss, SLyd, eFVt].
  3. Promising results, comprehensive ablations, and clear, well-motivated presentation [yjss, SLyd, eFVt].

Concerns and our responses

  1. Evaluation breadth and generalizability [eFVt W1; SLyd W1; yjss W2]. We added MultiID-Bench, MultiHuman-TestBench, and six additional models/methods, including Qwen-Image-Edit, UNO, and InfiniteYou. FreeID consistently improves identity preservation, with gains up to +82.5% on weaker backbones and +19.7% on Qwen-Image-Edit.

  2. Cross-view and small-scale subjects [eFVt W2–3; yjss W1, W3]. Across 2,000 paired generations, FreeID improves identity under all viewpoint changes (+25% beyond 45°) without restricting head rotation. Gains are largest for small faces (+30%). We also added uncurated qualitative results and clarified the high same-view similarity scores.

  3. Identity preservation vs. editability [SLyd W2; yjss W1]. At the default setting, identity improves by +16% while edit success remains nearly unchanged (95.5% → 95.3%) and head rotation is preserved. Additional expression-changing experiments confirm identity gains (+5.9%) with minimal changes in image quality and text alignment.

  4. Attention mechanism [eFVt Q2; SLyd Q1]. We added quantitative evidence linking attention allocation to identity preservation. Under large viewpoint changes, FreeID increases reference-face attention (0.62 → 0.83), with positive correlations between attention and identity similarity (r = 0.33) and between their improvements (r = 0.29).

  5. Implementation and other clarifications [yjss Q1; eFVt W1, Q1, Ethics; SLyd Q1]. We clarified the attention-bias schedule, implementation settings, evaluation prompts, limitations regarding closed-source models, and privacy considerations for Unsplash images.

We sincerely thank the AC and reviewers for their time and consideration. We hope our additional evidence and clarifications have addressed the major concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.