- 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
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) | # Images | InsightFace (Baseline) | InsightFace (+ FreeID) | Δ InsightFace [95% CI] |
|---|---|---|---|---|
| < 15° (same view) | 1349 | 0.567 | 0.642 (+13%) | +0.075 [0.072, 0.079] |
| 15–30° | 229 | 0.467 | 0.550 (+18%) | +0.083 [0.073, 0.093] |
| 30–45° | 125 | 0.349 | 0.474 (+36%) | +0.126 [0.110, 0.144] |
| ≥ 45° (cross-view) | 297 | 0.347 | 0.436 (+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.
| τ | InsightFace | Head turn |
|---|---|---|
| 0 (baseline) | 0.509 | 45.1° |
| 0.005 | 0.554 | 44.7° |
| 0.01 | 0.573 | 44.6° |
| 0.02 (default) | 0.592 | 44.6° |
| 0.05 | 0.610 | 44.6° |
| 0.1 | 0.614 | 44.4° |
| 0.2 | 0.585 | 43.6° |
We added this analysis to Appendix D.11, and the τ sweep to Appendix D.12 (Table A14).
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) | # Images | Face share (Baseline) | Face share (+ FreeID) |
|---|---|---|---|
| < 15° (same view) | 1190 | 0.72 | 0.83 |
| 15–30° | 197 | 0.73 | 0.85 |
| 30–45° | 104 | 0.73 | 0.91 |
| ≥ 45° (cross-view) | 259 | 0.62 | 0.83 |
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.
| Relationship | r [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.
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.
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.
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.
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.
We added these results to Appendix D.14 (Tables A17 and A18) and Appendix D.15 (Table A19).