- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper combines a one-step face-restoration prior with a learned codec and adds a training-only identity loss. It reports improved identity retention at ultra-low bitrates with small changes in other quality metrics, and studies transfer across codecs, datasets, resolutions, and face-video key-frame coding.
Critical strengths
The evaluation is substantial: it uses actual coded bitrates, identity networks held out from training, verification tests, and matched identity-loss ablations. Cross-codec and cross-dataset results make the empirical observation useful beyond a single implementation.
Critical weaknesses
-
The methodological novelty is limited, as the paper acknowledges that identity losses have already been used in face codecs. The main new claim is that the trade-off changes specifically in the ultra-low-rate regime. The low-rate results support a favorable operating point, but comparison with higher-rate findings from different prior systems does not isolate a regime change. A controlled account of where the trade-off changes within the same codec would make the contribution more convincing.
-
The mechanism analysis supports identity-sensitive correction but does not uniquely establish the proposed explanation. Wrong-identity targets and random features are weak alternatives to meaningful pretrained feature supervision. Spatial concentration and low-frequency error reduction also do not by themselves show that identity-specific geometry, rather than a useful additional perceptual constraint, explains the gain. The causal interpretation should be narrowed or supported by a better-matched control.
Questions for authors
- Which existing results establish a change from costly to near-free identity preservation as bitrate decreases within the same codec and training setup?
- Is there a matched control using an additional pretrained, non-identity feature loss or structural loss? How do the current controls distinguish that explanation from identity-specific alignment?
4 comments
Response to Reviewer d9PX (2/4)
Weakness 2 — The controls are weak, and spatial and frequency evidence alone does not establish causality
We thank the reviewer for highlighting an important limitation of our original controls. We address the two points separately below.
We agree that the original controls did not fully distinguish identity-specific supervision from general pre-trained feature supervision. The wrong-identity-target and random-network controls go only as far as showing that the gain appears only when correct identity information is present, and a comparison against correctly paired pre-trained non-identity feature supervision was not included. As the reviewer suggests, we have trained a condition-matched control: the same aligned region, the same loss form and the same training recipe, with only the supervising feature space changed to a non-identity one (DINOv3 self-supervised representations). The result was that the perceptual and distributional metrics do improve, but no identity gain is produced — a generic perceptual constraint does not account for the identity gain we report. For the design rationale and the quantitative numbers, please see our answer to Question 2 below.
On the spatial and frequency evidence: although these analyses alone cannot establish causality, we believe they provide complementary evidence on where the observed improvements occur. The quality improvement is not distributed uniformly but concentrates on the parts that carry identity (Table 2) — +1.64 dB on the inner identity region (eyes, eyebrows, nose, mouth) versus +0.04 dB on the background. It is not flat within the face either: inner identity +1.64 dB vs face skin +1.08 dB, and decomposing by part (Fig. 8) puts the parts with high identity discriminability at the top — nose +2.17 dB, upper lip +1.93 dB, eyebrows roughly +1.5 dB. We think a constraint acting uniformly over the whole image would be unlikely to produce part-to-part variation of this kind.
The frequency analysis points the same way. The reduction in error power is confined to the low frequencies and vanishes at high frequencies (low −36.2%, mid −15.7%, high +0.4%). Meanwhile the amount of high-frequency signal actually increases (71.9% → 74.3% relative to the ground truth). That is, the additional high-frequency content is not aligned with the original, and what actually improves is the low-frequency structure. We would add that our reference model is already trained with DISTS (1.0), CLIP (0.1) and a DINOv3-discriminator-based adversarial loss (0.1), so the two analyses above measure the increment from adding or removing identity supervision on top of a model that already receives perceptual supervision.
Even so, we agree that these two analyses alone cannot rule out every possible alternative, and that is precisely why we added the control above. In the revision we have sharpened the causal statement to "this gain does not appear without identity-specific supervision, and the resulting quality improvement concentrates on the identity-bearing facial structures and the low-frequency band", stated at the end of Test 3 the range of alternative explanations that the spatial and frequency analyses alone do not exclude, and added the non-identity feature loss to the list of controls in Test 3 (Fig. 5) and to the matched-bitrate comparison table (Table 8 in App. C).
Response to Reviewer d9PX (3/4)
Question 1 — What existing result establishes that, within the same codec and the same training setup, identity preservation shifts as the bitrate decreases from being costly to being nearly free?
This question addresses an important distinction that we did not make sufficiently clear in the original submission. We acknowledge that our experiments do not establish such a transition within the same codec and training setup. As stated in our answer to Weakness 1 above, what we measured and claimed is that the identity loss is nearly free in the ultra-low-bitrate range, and the "costly" side of the contrast is a citation of results reported by prior work. Our paper contains neither a result establishing a transition within the same codec and the same training setup nor a result showing that the trade-off tendency strengthens as the bitrate rises. We acknowledge that our original wording did not sufficiently distinguish these claims, and we have revised the manuscript accordingly.
As the table in that same answer shows, within our range the near-free behaviour does not switch on at a particular point but holds across the entire range tested, and even at the highest-bitrate operating point where rate matching is possible (λ_rate = 4) the cost stays below 0.04 dB in PSNR and below 3.3% in relative perceptual and distributional terms.
The question also lets us state one further limitation more clearly. Whether the difference between the cost reported by prior work and our observation comes from the difference in bitrate regime or from differences in codec, objective and training setup cannot be separated with our data alone, because the systems being compared are different.
In the revision, in addition to the amendments described above, we have restricted the scope of the near-free claim to the bitrate range we measured (roughly 0.004–0.03 bpp) and to the codecs we experimented on, stating this in two places: the "Scope of Contribution 1" paragraph of §1 "Introduction" and §7 "Limitations". In §7 "Limitations" we also state that the bitrate-regime difference and the system difference are not separated, and that separating them would require sweeping the same codec to much higher bitrates.
Response to Reviewer d9PX (4/4)
Question 2 — Is there a condition-matched control using a pre-trained non-identity feature loss or a structural loss, and how do the present controls distinguish that explanation from identity-specific alignment?
The submission did not include such a control. Following the reviewer's comment we have trained a
condition-matched non-identity pre-trained feature loss control, designed as follows. It uses the
same loss form, 1 − cos(F(x̂), F(x)), on the same aligned face region, with the training recipe,
schedule and data all held identical; the one thing that changes is whether the supervising feature
space encodes identity.
We chose DINOv3 as the feature extractor, for two reasons. First, it is a representation obtained by large-scale self-supervised learning without identity labels, so by construction it satisfies the condition of being a meaningful pre-trained feature space while not being aimed at identity discrimination (the ArcFace family, trained with margin-based identity classification, sits at the opposite end). Second, DINOv3 provides a strong general-purpose visual representation, which makes it a meaningful non-identity control complementing the random-network control of the submission.
In addition, so that the comparison would not hinge on a single weight choice, we trained it at the same three weights as the identity loss (0.05 / 0.1 / 0.2) and compared frontier against frontier.
The results are as follows (λ_rate = 32, relative to the λ_id = 0 model, all 3,500 images). The network the identity loss supervises is CurricularFace, while ArcFace and MagFace are held-out networks not used in training — the third column below is included for diagnosis only.
| Model | ΔArcFace (held-out) | ΔMagFace (held-out) | ΔCurricular (supervising network, diagnosis only) |
|---|---|---|---|
| Identity loss (λ_id = 0.1) | +0.233 | +0.257 | +0.274 |
| Non-identity feature loss (w = 0.05) | +0.007 | +0.009 | +0.010 |
| Non-identity feature loss (w = 0.1) | +0.011 | +0.013 | +0.014 |
| Non-identity feature loss (w = 0.2) | +0.018 | +0.021 | +0.022 |
On all three recognition networks, the identity gain of the non-identity feature loss is between 1/29 and 1/13 of that of the identity loss, and quadrupling the weight from 0.05 to 0.2 does not close the gap. Moreover, comparing at matched bitrate makes even this residual gain disappear — interpolating the λ_id = 0 model to each control's bitrate gives +0.002 or less at w = 0.05 and an actual loss at w = 0.1 and 0.2. The small identity increase of the controls was thus an effect of spending more bits (on the same basis the identity loss retains +0.18 to +0.22).
This control does, on the other hand, genuinely improve the perceptual and distributional metrics — at all three weights LPIPS, DISTS and CleanFID are better than the λ_id = 0 model, and CleanFID improves even more than with the identity loss (−0.639 vs −0.279 at w = 0.2). In other words, condition-matched non-identity feature supervision acts as a useful additional perceptual constraint and yet produces no identity gain. These results indicate that the identity gain we report is not explained by the generic perceptual supervision tested here, and support the importance of identity-specific supervision.
On the second part — the discriminative power of the existing controls — the answer is as given in our reply to Weakness 2 above. The wrong-identity-target and random-network controls show only that the gain appears only when correct identity information is present, and they do not distinguish the alternative of useful pre-trained feature supervision. It is the DINOv3 control above that separates that alternative.
In the revision we have added this control as a formal entry to the specificity controls (Fig. 5) of §5 "Why Is Identity Nearly Free?" and to the accompanying text, and we have included the rate-matched comparison in the iso-bpp table (Table 8) of App. C "Mechanism details".
We hope the narrowed claims and the added control address the concern. We thank the reviewer for encouraging us to state the scope of the finding precisely; the paper is more accurate for it.
Response to Reviewer d9PX (1/4)
We sincerely thank the reviewer for this careful and insightful review. Please find our point-by-point response below.
Weakness 1 — A comparison across different prior systems cannot isolate the change of regime
We state in §2 that the identity loss itself is not new, and what we offer as our contribution is not a component but the measurement, and the analysis of its cause, that the trade-off between identity and the other quality axes effectively does not hold at ultra-low bitrates. The reviewer is right that we need to be clearer about the ground on which that claim stands, and we should have written the scope of our claim more carefully.
We would first like to state the scope of the claim precisely. What we measured and claimed is that at ultra-low bitrates the identity gain is far larger than the cost on the other axes. The "usual trade-off" we contrast it with is a citation of results reported by prior work at their own operating bitrates. We did not intend to claim that a trade-off is induced at high bitrates, nor to claim a measurement of the transition point itself, but we agree that the wording in the main text could be read that way.
From the perspective the reviewer raises, we present controlled results from within a single codec. Below, the architecture, the training recipe, the data and the entropy model are all identical and only λ_rate differs across four operating points, each compared with the λ_id = 0 model at matched bitrate (for all three values of λ_id). Operating points that fall outside the range where rate matching against the λ_id = 0 curve is possible are excluded.
ΔIdentity is the average of the held-out recognition networks (ArcFace, MagFace), and the λ_id = 0 model it is compared against is evaluated by PCHIP interpolation along its own rate–quality curve at each model's bpp. A negative ΔPSNR is a cost and a positive ΔIdentity is a gain.
For all three weights, the PSNR cost shrinks as the bitrate rises (λ_rate = 32 → 4: 0.073 → −0.001 dB at λ_id = 0.05, 0.200 → 0.011 dB at 0.1, 0.369 → 0.035 dB at 0.2). At λ_rate = 4 — the highest-bitrate operating point at which rate matching is possible — the PSNR cost is below 0.04 dB, and at our default λ_id = 0.05 the sign flips to a slight gain. The perceptual and distributional costs shrink in the same direction, to 3.3% or less in relative terms at the same operating point (1.8% or less at the default setting). The cost decreases in all 12 combinations of the four cost axes and the three weights. In other words, within our experimental range we observe no tendency for the trade-off to worsen as the bitrate rises.
We should, however, be clear about the limits of this result. The observation above holds within the bitrate range we experimented on (roughly 0.004–0.03 bpp, the range where rate matching is possible), and we have not verified whether it is a phenomenon confined to this range or a property that continues to much higher bitrates. We therefore do not claim to have located the boundary at which the behaviour changes.
In the revision we have (i) stated the above bitrate-dependent cost behaviour in §4.2 "Results" (the cost decreases for all three weights; from λ_rate = 32 → 4 the PSNR cost goes from 0.07/0.20/0.37 → 0.00/0.01/0.04 dB) and included the supporting table as Table 9 in App. C "Mechanism details". (ii) In the "Scope of Contribution 1" paragraph of §1 "Introduction" (immediately after the contributions are enumerated) we have added a scope note saying that this result concerns the bitrate range we measured and the systems we experimented on, and that the contrast with prior work is a comparison that differs not only in bitrate but also in system, and we have revised "reverses the usual tradeoff" in Contribution 2 to "the trade-off is weak in this range". (iii) The limitation above is stated in §7 "Limitations".