acceptodds
Under review as a conference paper at ICLR 2027

Preserving Identity: A Near-Free Lunch in Ultra-Low-Rate Face Compression

Abstract

Generative codecs enable image compression at ultra-low bitrates, but on faces they exhibit identity drift: the reconstruction is perceptually plausible yet depicts a different person, a failure that standard fidelity/perceptual metrics do not capture. We build a face codec by repurposing a one-step diffusion face-restoration prior as the decoder of a compression pipeline. To this, we add an ArcFace-aware identity loss at training time. On a rate–distortion basis — with actual entropy-coded bitstreams — the loss yields a large identity gain (measured by held-out recognition networks) at an almost negligible cost on every other axis: a near-free lunch. On FFHQ, the same codec without the identity loss needs 50.9% more bitrate to match our identity, while the BD-rate on all distortion, perceptual, and distributional metrics stays within ±5%; recent generative codecs need 116–225% more bitrate for our identity. Controlled experiments show the gain is identity-specific structural alignment (drift suppression) rather than generic feature regularization or high-frequency hallucination. The identity gain transfers across codecs and datasets, and the codec extends to a single resolution-agnostic model that keeps its identity lead on in-the-wild unaligned faces, and to face-video compression — a near-free route to identity-faithful face compression. Our code is included in the supplementary material.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
learned image compression, face image compression, identity preservation, diffusion-based compression, rate–distortion–perception tradeoff

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_d9PXOpenReview ↗
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
  1. 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.

  2. 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
  1. Which existing results establish a change from costly to near-free identity preservation as bitrate decreases within the same codec and training setup?
  2. 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
AuthorsOpenReview ↗

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.

λ_idλ_ratebppΔIdentityΔPSNR (dB)ΔLPIPSΔDISTSΔCleanFID
0.05320.00454+0.131−0.073+0.0039+0.0014+0.130
160.00825+0.115−0.075+0.0040+0.0021+0.164
80.01442+0.070−0.045+0.0024+0.0013+0.086
40.02369+0.031+0.001+0.0011+0.0005+0.078
0.1320.00501+0.191−0.200+0.0090+0.0036+0.512
160.00877+0.152−0.143+0.0070+0.0033+0.267
80.01511+0.088−0.071+0.0039+0.0019+0.152
40.02446+0.040−0.011+0.0017+0.0011+0.098
0.2320.00556+0.238−0.369+0.0162+0.0066+0.766
160.00967+0.173−0.293+0.0113+0.0051+0.525
80.01599+0.101−0.143+0.0062+0.0030+0.214
40.02556+0.047−0.035+0.0030+0.0017+0.123

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".

AuthorsOpenReview ↗

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).

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

ModelΔArcFace (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.

Reviewer_b2KaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies identity drift in generative face codecs at ultra-low bitrates, where reconstructions look realistic but depict a different person. The authors build a face codec by combining the AEIC encoder and entropy model with an OSDFace one-step diffusion decoder, and add a face-recognition identity loss during training. Measured by BD-rate on real bitstreams, the loss substantially improves identity on held-out recognizers at a small cost on distortion, perceptual, and distributional metrics, which the authors attribute to identity-specific structural alignment. The gain is further shown to transfer to other codecs, an unseen dataset, in-the-wild faces, and face video.

Critical strengths
  • Thorough and transparent experiments and analysis. Extensive experiments are conducted to evaluate the proposed method and to analyze its potential mechanisms.
  • Solid and practical problem. The problem studied is practical and critical for real-world applications.
  • Simple and practical design. The identity loss is applied only during training, adds no parameters or inference cost, and requires no side information.
Critical weaknesses
  • Limited novelty. Identity losses have been widely studied in both the image generation and face compression communities. This paper appears to be a combination of AEIC, OSDFace, and an identity loss, accompanied by extensive experiments that analyze why it works and why it is nearly a free lunch.
  • Limited experiments and analysis on corner cases. Although the paper focuses on analyzing the potential mechanisms of the identity loss, it lacks experiments and analysis on corner cases such as small faces, multiple faces, and general images. Identity losses have been reported to be devastating for non-face scenes and for scenes with small or multiple faces. The lack of experiments on these corner cases reduces the applicability of the added identity loss in real-world applications and makes the contribution limited and incremental.
Questions for authors

At matched bitrate, how do the PSNR, LPIPS, and CleanFID of the λid = 0.1 model compare with those of the anchor, and how should the claim that "identity and fidelity improve together" be interpreted?

5 comments
AuthorsOpenReview ↗

Response to Reviewer b2Ka (1/3)

We sincerely thank the reviewer for this careful and insightful review. Please find our point-by-point response below.

Weakness 1 — Limited novelty

We agree that the identity loss itself is not new, and the paper cites the prior work in §2 "Related Work". What we offer as our contribution is not a new module 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. With the same architecture and the same training recipe, turning on the identity loss alone improves identity BD-rate by −33.7%, while the remaining six axes (PSNR, MS-SSIM, LPIPS, DISTS, PatchFID, CleanFID) move only within a −1.3 to +5.0% range. Prior work has reported the same loss, at higher bitrates, as a substantive cost; we show that at ultra-low bitrates that cost barely materialises, measured on actual bitstreams and with held-out recognition networks. We analyse the cause with controlled experiments, and we confirm that the effect transfers to GAN, VQ and non-generative MSE codecs, to datasets not seen in training, and to face-video key frames.

The contribution lies not in the novelty of the components but in the observation itself and in the fact that it is not confined to our architecture, our training data, or our evaluation recognition networks. To make the boundary of the contribution clear, in the revision we have explicitly added a sentence stating that we claim no novelty for the components themselves, immediately before the contributions are enumerated in §1 "Introduction".

One clarification on the components may also be helpful. The decoder backbone of the reported model is the general-purpose Stable Diffusion 2.1-base, and the LoRA weights released by OSDFace are not part of it; OSDFace is prior work that showed this backbone to be effective for one-step face generation, and it served as our rationale for choosing the backbone. Our own description of this was not sufficiently clear, and we have corrected the wording in the revision. All models reported in the paper share the same backbone, the same pre-trained checkpoint and the same training recipe, so no comparison or number changes. The components are therefore two — the AEIC-based architecture and the identity loss.

Weakness 2 — Corner cases: multiple faces and non-face regions

We thank the reviewer for highlighting the importance of evaluating multiple-face scenarios. In response, we have conducted an additional experiment. We evaluated the photographs we had excluded from the single-face benchmark — 500 Flickr-2W photographs with 2–5 detected faces (1,670 faces) — at the same five operating points. The result is that the benefit of the identity loss transfers even to faces it does not supervise: on the 1,170 unsupervised faces, to reach our identity the λ_id = 0 model needs 16.9% more bitrate and AEIC-ME 26.0% more in BD-rate terms. Averaged over all faces the figures are 20.5% and 32.4% respectively, and the PSNR and LPIPS cost is within 2%. In other words, supervision applied to a single face does not come at the expense of the remaining faces. That said, per-face bit allocation is still not implemented, so situations in which a particular face must be prioritised are left as a limitation in §7 "Limitations". This experiment appears in the revision as App. S "Photographs with several faces" (Table 18), with the summary numbers in §6.

For non-face regions, our measurements show no adverse effect. Semantically segmenting the FFHQ test set and aggregating only the non-face regions (n = 3,500, our default λ_id = 0.05, the same λ_rate = 16), the identity-loss model is in fact marginally better than the λ_id = 0 model: PSNR 22.848 vs 22.790, LPIPS 0.0892 vs 0.0902, DISTS 0.0714 vs 0.0719. The identity loss therefore does not sacrifice the regions outside the face.

AuthorsOpenReview ↗

Response to Reviewer b2Ka (2/3)

Weakness 2 — Corner cases: small faces and general images

For small faces we measured on CPLFW and CFP. The median effective face resolution there is 54 px on CPLFW and 134 px on CFP, i.e. 1/5 and 1/2 of FFHQ (274 px) (median face bbox measured at the original resolution). Under these conditions we compared against five codecs — AEIC-ME, OneDC, MS-ILLM, GLC and VTM (the remaining competitors were excluded because they cannot process mixed native resolutions in a single pass, which is the same criterion as in the in-the-wild comparison in the main text). Each codec decoded with the backbone it was trained with (SD-Turbo for AEIC-ME). In terms of the held-out average (ArcFace, MagFace), the extra bitrate each codec needs in BD-rate terms to reach our identity is as follows.

DatasetOneDCAEIC-MEMS-ILLMGLCVTM
CPLFW (n = 3,000, 256 px)+8.5%+47.3%+60.2%+67.0%+188.0%
CFP (n = 2,136, native)+8.7%+40.3%+80.7%+65.3%+185.3%

We should, however, be candid about the size of our margin in this regime. The large gaps against MS-ILLM, GLC and VTM are maintained, but against the strongest competitor, OneDC, the margin is narrow. Comparing our five operating points (λ_rate = 32 to 2) with OneDC's six, OneDC is in fact slightly ahead at both ends of the ladder (iso-bitrate held-out average difference of −0.016 to +0.032). Our advantage is largest where faces are large and unaligned, and it narrows on small aligned crops — this is in the same direction as the observation reported in the main text, and we believe it is expected behaviour, in that the smaller the face, the less information there is to reconstruct from in the first place.

The contribution of the identity loss itself holds up in this small-face regime as well. Measuring the λ_id = 0 model — the same recipe with only the identity loss switched off — at the same 5 operating points, reaching our identity requires 12.9% more bitrate on CPLFW and 18.0% more on CFP in BD-rate terms. Because both sets consist of tight crops in which the face nearly fills the frame, face detection is unstable, so identity was computed from the 5-point landmark alignment provided by the datasets, identically for all codecs.

The CPLFW and CFP comparisons above have been added to the revision as App. Q "Small aligned faces: CPLFW and CFP".

We also share the reviewer's concern about applicability. We too regard this codec not as a general-purpose codec to be applied directly to arbitrary photographs, but as one targeting inputs in which a face is dominant. Within that restricted domain, however, we do present evidence that reaches a real application — in animation-based video-conferencing coding (GFVC), replacing only the key frame codec with ours (leaving the backbone untouched) saves 72% (LPIPS), 79% (DISTS) and 78% (PatchFID) of the total video bitrate across all 33 CTC sequences. This is because the key frame is the only component of a GFVC clip that is image-compressed, so its identity and texture propagate to every frame. We plan to pursue this direction in follow-up work.

We additionally measured the most extreme corner case — natural images with no faces. Computing BD-rate on Kodak24 against AEIC-ME, which is optimised for natural images, our model spends 14–35% more bitrate for the same quality (PSNR +14.0%, LPIPS +18.8%, DISTS +34.8%, patch-FID +33.5%). Almost all of this gap, however, comes from face-domain finetuning, and the contribution of the identity loss itself is 0.3–2.5% (relative to λ_id = 0: PSNR +0.3%, LPIPS +1.9%, DISTS +2.5%, patch-FID +2.4%). That is, our codec's cost on natural images is the price of having learned a face distribution, and the identity objective adds almost nothing to it. This table appears in the newly added App. R "Out-of-domain behaviour on natural images" (Table 17) of the revision, and we state the scope of applicability in one sentence in §7 "Limitations".

AuthorsOpenReview ↗

Response to Reviewer b2Ka (3/3)

Question — At matched bitrate, how do the PSNR, LPIPS and CleanFID of the λ_id = 0.1 model compare with those of the anchor, and how should the claim that "identity and fidelity improve together" be interpreted?

This is a fair question, and the submission answers only half of it. For the identity axis the bitrate-matched comparison was included in App. C "Mechanism details" (the iso-bpp identity control table, in which the λ_id = 0 model is interpolated to the bpp of the identity-loss model). For the non-identity axes the reviewer asks about (PSNR, LPIPS and CleanFID) it was not: those axes were reported only as BD-rate. We are glad to provide it here.

At each model's bpp we interpolate the λ_id = 0 model along its own rate–quality curve, with the same interpolation scheme used for BD-rate throughout the paper (PCHIP). Identity is the average of the held-out recognition networks (ArcFace, MagFace). In the revision we have recomputed the identity-axis table (Table 8 in App. C) under the same PCHIP convention as well (in the submission it used linear interpolation between the two adjacent points), so the λ_rate = 32 identity numbers in that table and the corresponding row of the table below agree.

λ_idλ_ratebppΔIdentityΔPSNRΔLPIPSΔDISTSΔCleanFID
0.1 (asked about)320.00501+0.191−0.200+0.0090+0.0036+0.512
160.00877+0.152−0.143+0.0070+0.0033+0.267
80.01511+0.088−0.071+0.0039+0.0019+0.152
40.02446+0.040−0.011+0.0017+0.0011+0.098
0.05 (our default)320.00454+0.131−0.073+0.0039+0.0014+0.130
160.00825+0.115−0.075+0.0040+0.0021+0.164
80.01442+0.070−0.045+0.0024+0.0013+0.086
40.02369+0.031+0.001+0.0011+0.0005+0.078

At λ_id = 0.1, identity improves substantially (held-out average +0.191) while the perceptual and distributional metrics are marginally worse (at λ_rate = 32, LPIPS +0.0090 and CleanFID +0.512 — 4.3% and 6.2% of the reference values respectively).

We would add that λ_id = 0.1 is a larger weight than our default (λ_id = 0.05), so to aid the comparison we have included the numbers for the default setting in the table above as well. As stated in §4.2 "Results", we adopt λ_id = 0.05 as the default precisely because the loss on the other axes is negligible at that point. The cost at the default setting is less than half as large (at λ_rate = 32, LPIPS +0.0039 and CleanFID +0.130 — 1.8% and 1.5% respectively). That the identity gain and the perceptual/distributional cost grow together as the weight is increased can also be seen in the identity-weight ablation in §4.2 "Results" (−44.7% at 0.1, −53.8% at 0.2, with a cost of up to +16% for the latter).

On the second part — the interpretation of the phrasing. The phrase "identity and fidelity improve together" referred to the observation in §5 "Why Is Identity Nearly Free?" that, in a comparison at the same λ_rate, the identity-loss model is ahead of the λ_id = 0 model not only in identity but also in fidelity within the face region. In the same-λ_rate comparison the quality improvement is not spread uniformly but concentrates on the regions that carry identity, and the inner identity region (eyes, nose, mouth) improves roughly 6× more than the background in LPIPS and roughly 40× more in PSNR.

We accept, however, that this phrasing could be read as saying that the global metrics improve at matched bitrate too. In the revision we have (i) stated at the beginning of §5 "Why Is Identity Nearly Free?" that the statement concerns a comparison at the same λ_rate, made explicit that at matched bitrate there is a small cost on the non-identity metrics and that we therefore do not claim a simultaneous gain in the rate–distortion sense, and clarified that what this section establishes is where the quality improvement concentrates (the identity-bearing regions and the low frequencies). (ii) The table above appears as Table 9 in App. C "Mechanism details" (the typeset table additionally includes a row for the third ablation weight, λ_id = 0.2), and we cite it at the end of the identity-weight ablation in §4.2 "Results".

We hope the added corner-case experiments address the concern, and we would be glad to run any further check the reviewer considers important.

Reviewer_dT7kOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents IdentiCodec, an ultra-low-bitrate face image compression framework built on AEIC. It adapts the one-step diffusion face-restoration prior OSDFace for decoding and incorporates an identity-preserving loss based on a frozen face recognition network during training. Experimenal results show that this method effectively addresses the issue of identity shift in faces compressed at ultra-low bitrates.

Critical strengths

(1) The paper presents extensive controlled experiments—including wrong-identity and random-network controls, spatial attribution, and spectral analysis—to investigate why identity preservation incurs only a small rate–distortion–perception trade-off at ultra-low bitrates. (2) The proposed method is simple.

Critical weaknesses

Both key components—identity-preserving supervision for face compression and one-step diffusion decoding for generative image compression—have been explored in prior work, limiting the novelty of the proposed method.

Questions for authors

OSDFace is already pretrained with an identity-preserving loss, raising the question of whether the gains from additional identity supervision during compression training depend on an interaction with this identity-aware prior. A controlled comparison using otherwise identical decoders pretrained with and without identity supervision would help clarify this dependence.

2 comments
AuthorsOpenReview ↗

Response to Reviewer dT7k (1/2)

We sincerely thank the reviewer for this careful and insightful review. Please find our point-by-point response below.

Weakness — Both components have been explored in prior work

We agree that both components have been studied before. Identity-preserving supervision has been explored both in face compression and in face generation and restoration, and one-step diffusion decoding is likewise an established approach in generative compression. The paper cites this prior work in §2 "Related Work", and we do not claim either component as a new module. The decoder adopts a composition established in prior work.

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. Concretely, with the same architecture and the same training recipe, turning on the identity loss alone improves identity BD-rate by −33.7%, while the remaining six axes (PSNR, MS-SSIM, LPIPS, DISTS, PatchFID, CleanFID) move only within a −1.3 to +5.0% range. Prior work has reported the same loss, at higher bitrates, as a substantive cost accompanied by a PSNR drop and a rate increase; we show that at ultra-low bitrates that cost barely materialises, measured on actual entropy-coded bitstreams and with held-out recognition networks. We analyse the cause with controlled experiments, and we confirm that the effect transfers to GAN, VQ and non-generative MSE codecs, to datasets not seen in training, and to face-video key frames. The contribution lies not in the novelty of the components but in the observation itself and in the fact that it is not confined to our architecture, our training data, or our evaluation recognition networks.

So that the boundary of the contribution is not misread, in the revision we have explicitly added a sentence stating that we claim no novelty for the components themselves, immediately before the contributions are enumerated in §1 "Introduction" — both the identity loss and the one-step diffusion decoder are taken from prior work, and what is new is the measurement above, together with its cause and its generality. In addition, one phrase in the contribution list read as though the decoder were an achievement of ours, and we have corrected that description as well. The specifics of that correction are given in our answer to the question below.

AuthorsOpenReview ↗

Response to Reviewer dT7k (2/2)

Question — Does the gain depend on an interaction with OSDFace's identity-aware prior?

Before answering, we would like to correct an inaccuracy in our own description of the decoder. The text describes it as "repurposing the OSDFace prior", whereas the backbone of the reported model's decoder is the general-purpose backbone (SD-2.1) that OSDFace showed to be effective for one-step face generation (excluding the OSDFace LoRA); it is not an identity-specialised backbone. (The attached code package loads the official Stable Diffusion 2.1-base as the diffusion backbone as is, and the weights released by OSDFace are a LoRA that is not part of the reported model.) Every model reported in the paper — the λ_id = 0 model and the identity-loss models, at all bitrate points — starts from the same backbone and the same pre-trained checkpoint and uses the same training recipe, so this correction introduces no backbone mismatch between models, and all comparisons and all reported quantitative numbers remain valid as they stand. In the revision we have amended the corresponding statements in Contribution 3 and §3 "Method" accordingly, and we have also corrected the places in the appendix where the OSDFace weights were described as a component.

We also evaluated a configuration that stacks the OSDFace LoRA onto the decoder while keeping all other settings identical. As shown below, it yields only marginal differences, which supports our use of the simpler SD-2.1 backbone.

In the public code accompanying the revision, we have removed the loading of the OSDFace weights, so that the reported configuration is also unambiguous in the code.

With and without stacking — λ_rate = 8, 3,500 FFHQ test images, on the measured rANS bitstream. The bitrates of the two configurations agree to within 0.13%, so they can be compared directly without interpolation. The identity metrics come from held-out networks not used in training (ArcFace, MagFace).

ConfigurationbppArcFaceMagFacePSNRLPIPSCleanFID
Reported model (SD-2.1)0.014420.82080.856825.170.12905.51
+ OSDFace LoRA stacked0.014440.82440.858525.190.12845.40
Difference+0.1%+0.0036+0.0017+0.02−0.0006−0.11

As the table shows, actually stacking OSDFace's identity-supervised prior onto the decoder leaves identity preservation essentially unchanged; the reported identity gain comes not from an interaction with that prior but from the identity term we added to the training objective. This comparison appears as Table 3 in App. A "Codec architecture" of the revision.

We hope these clarifications address the concern. We are grateful for the question on the decoder: it led us to correct a description that was genuinely inaccurate.

Comments on the paper

1 comment
Area_Chair_HxDzOpenReview ↗

Author-Reviewer Discussion

Dear Reviewers,

Thank you for your time and thoughtful reviews. With the author–reviewer discussion period ending on November 18 (AoE), please review the authors' rebuttal and share your updated assessment. Specifically, let us know whether your main concerns have been resolved, and if not, which points remain open.

Your input before the deadline is greatly appreciated!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.