{"id":"a492f185-c962-456b-b613-e16e5732ed47","arxiv_id":"2412.15171","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"Pose-corrective networks for Gaussian avatars are distilled into shared linear layers, giving three full-body avatars real-time animation and rendering on a Quest 3 headset.","lead":"SqueezeMe converts high-quality 3D Gaussian avatars into a much smaller representation that can be animated and rendered on a Meta Quest 3, running three avatars at 72 frames per second. It replaces the heavy neural network that computes pose-dependent changes with a simple linear model and shares corrections between nearby surface points to cut memory and compute.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'minimal degradation' claim rests on average metrics; the paper's own Figure 6 shows the low-frequency corrective-sharing assumption breaks at joints and garment edges, and no per-frame error distribution is reported to show these failures are rare.","rationale":"I read the paper as claiming a practical system: a lightweight avatar representation that preserves the quality of a CNN decoder while running on a Quest 3. For that claim, the key scientific step is the distillation of nonlinear correctives into a 64-dimensional linear map plus 4x4 corrective sharing. The paper's quantitative tables show small average degradation, and the AvatarRex results partially mitigate the internal-dataset concern. The strongest counter-evidence is the paper's own limitation section: the same regions where the low-frequency assumption should fail (joints, sleeve boundaries, pants seat) are where visible artifacts appear. The reader's weakest_assumption identifies exactly this. My concern adds that the current averages cannot distinguish 'rare, acceptable artifacts' from 'systematic degradation concentrated in specific poses,' and the paper provides no error bars or worst-case analysis. A per-frame error distribution with a pre-specified threshold would settle this without requiring code release. I do not see a fatal flaw; the method is plausible and the quantitative means are in line with the claims. The CONDITIONAL verdict is appropriate, so no change is needed; the authors should add the distributional analysis or soften the 'minimal degradation' wording.","tokens_in":15880,"tokens_out":10825,"duration_ms":84984,"concrete_test":"Run the authors' evaluation pipeline on the held-out set and compute, per frame and per identity, the LPIPS of SqueezeMe (GCS+Linearized) and the CNN teacher. Report the 95th percentile and the fraction of frames where the distilled model's LPIPS exceeds the teacher's by more than 0.03, and correlate that fraction with joint-angle distance from the 512 distillation key frames. If high-error frames are rare (<5%) and occur only at extreme poses outside the key-frame hull, the 'intermittent' characterization is supported; if they are common within normal articulation ranges, the central fidelity claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition for the central claim is that the distilled model (64-dim PCA pose basis plus one corrective per 4x4 UV block) stays close to the CNN teacher across the pose distribution. That condition is least secure precisely where the paper's own failure cases appear: Figure 6(b,d,e) shows visible degradation at the armpit, t-shirt sleeve edge, and pants seat with legs apart, i.e., regions where neighboring Gaussians move independently and correctives are not low-frequency in UV space. Table 1 reports only means over 4 identities with no error bars or per-frame breakdown; a mean LPIPS of 0.151 can hide a long tail of high-error frames, and the held-out set may be dominated by easy poses. Because the linear distillation is a single least-squares fit from 64 pose coordinates to per-Gaussian correctives, any nonlinear or high-frequency pose dependence in the teacher is systematically smoothed; the open question is whether the resulting artifacts are truly intermittent or occur often enough to break the 'minimal degradation' claim in normal VR use.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents SqueezeMe, a system for converting high-fidelity Gaussian full-body avatars into a mobile-ready representation. The authors train a CNN decoder that predicts pose-dependent Gaussian correctives on a UV map, then distill it into a linear layer via PCA on poses and least-squares regression on correctives. They further reduce memory by sharing correctives among 4x4 UV neighborhoods (GCS). They implement a Vulkan-based splatting pipeline and report real-time animation and rendering of three avatars at 72 FPS on a Meta Quest 3. Quantitative evaluation on an internal 4-identity dataset and on AvatarRex compares favorably to Animatable Gaussians, with a GCS+Linearized model reducing decoder latency to 0.45 ms.","tokens_in":16087,"tokens_out":4305,"duration_ms":37090,"significance":"The paper is a solid systems contribution. Its strengths include evaluation on held-out poses and cameras, comparison to Animatable Gaussians on two datasets, a clear ablation of distillation and sharing, and an unusually honest discussion of failure cases. If the quality claims survive closer scrutiny, the result is practically significant for VR. The main weakness is that the 'minimal degradation' claim is supported only by aggregate means over four identities, while the paper's own failure figures show localized artifacts at joints and garment boundaries.","major_comments":[{"comment":"The headline claim that 'quality degradation by the proposed distillation is minimal' (Introduction) is supported only by mean metrics over 4 identities, with no per-identity or per-frame breakdown. Given that Fig. 6(b,d,e) shows visible artifacts in exactly the GCS and linearized models, the averages could hide a long tail of high-error frames. Please report per-identity results, standard deviations, and/or quantiles of per-frame errors, and state how often the failure modes in Fig. 6 occur.","section":"Section 6, Table 1"},{"comment":"The load-bearing assumption of Gaussian corrective sharing is that pose-dependent correctives are low-frequency in UV space. The manuscript's own Fig. 6(b,d,e) demonstrates that this assumption fails at armpits, sleeve edges, and the pants seat with legs apart. Since the memory reduction from 60k to 4k correctives depends on this assumption, the claim that the loss in visual fidelity is 'minimal' requires quantitative evidence about how often and how severely these failure modes occur. Please report, for example, the per-frame error distribution separated by pose difficulty, or a UV-space error map for representative frames.","section":"Section 4.2, Eq. (8)-(9) and Section 6 (Limitations)"},{"comment":"The 'Linear from scratch' baseline is averaged over only 3 identities because one training run diverged. This makes the comparison against the distilled models (averaged over 4 identities) inconsistent and weakens the claim that distillation is necessary for stable training. Please report per-identity results for all rows, including the divergent run (e.g., as 'did not converge'), or use the same set of identities for all rows.","section":"Table 1, footnote"}],"minor_comments":[{"comment":"The word 'quantative' should be 'quantitative'.","section":"Section 6, Quantization"},{"comment":"The sentence 'SqueezeMe (GCS) outperforms AG on L1 and SSIM, and SqueezeMe (GCS) matches AG on L1' is internally contradictory; based on Table 3, GCS matches AG on L1 and outperforms on SSIM.","section":"Appendix C"},{"comment":"The stated baseline of 50 ms for the CNN decoder is not reported in Table 1 or Section 5; please specify the hardware and measurement conditions for this number.","section":"Conclusion"},{"comment":"The phrase 'for the first time' should be softened to 'to the best of our knowledge' unless a comprehensive literature search is provided.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the authors are honest about limitations. The main risk is that the central 'minimal degradation' claim may be overstated; the requested per-frame/per-identity statistics are necessary before acceptance. The comparison to previous mobile avatar work is brief but adequate, and the demo videos are a useful complement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this if you care about deploying Gaussian avatars on mobile. The headline is real: they run three full-body avatars at 72 FPS on a Quest 3, with a 0.45 ms decoder, and the quality drop from the CNN teacher is small in aggregate. That is a good systems contribution.\n\nWhat is new here is not the linear PCA basis per se—Gaussian Blendshapes and Gaussian Eigen Models already do that for heads—but doing it for full-body pose correctives, plus the corrective sharing in UV space that cuts the linear matrix 16x. The combination, with a custom Vulkan splatting pipeline, is what gets the on-device result.\n\nThe evaluation is more careful than most: held-out poses and cameras, comparison to Animatable Gaussians on the internal dome and on AvatarRex, an ablation against training a linear model from scratch, and a frank limitations section with failure images. The 'linear from scratch' divergence is honestly reported, not hidden.\n\nThe main soft spot is that the 'minimal degradation' claim rests on averages over four identities with no error bars or per-frame distribution. The paper's own Figure 6 shows the low-frequency sharing assumption breaks at armpits, sleeve edges, and pants seat—so the concern is not hypothetical. Those artifacts may be intermittent, but we do not know how often they appear in normal use. Also, one identity diverged in the from-scratch baseline and was excluded; that is a smell, though it does not invalidate the distillation claim. No code or binaries are released, so the on-device numbers are hard to reproduce independently.\n\nThe stress-test note worries that the assumption breaks exactly where it matters. I think that is fair but proportionate: the paper owns those failure cases, and the average metrics still look acceptable. The missing piece is a per-frame error analysis, not a redesigned method.\n\nWho is this for? People doing mobile neural rendering, avatar systems, and 3DGS compression. It will not reshape the field's theory, but it is a meaningful engineering milestone. I would send it to a serious referee, with a request for per-frame error statistics and a bit more transparency on the excluded identity. My recommendation: accept for peer review, likely conditional.","headline":"A solid systems paper that delivers a real mobile VR demo; the distillation is simple and honestly evaluated, though the 'minimal degradation' claim needs per-frame error statistics.","tokens_in":16694,"tokens_out":2564,"would_cite":true,"duration_ms":18265,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SqueezeMe distills a Gaussian avatar's neural pose decoder into a 0.45 ms linear layer, enabling three full-body avatars to be animated and rendered at 72 FPS on a standalone Quest 3 headset.","keywords":["3D Gaussian Splatting","full-body avatar","pose-dependent correctives","linear distillation","corrective sharing","mobile VR rendering","Vulkan splatting","UV-space Gaussian maps"],"falsifier":"Take the distilled 0.45 ms model and the CNN teacher, feed both a held-out set of extreme poses that the 512 distillation key frames do not cover, and measure per-pixel LPIPS and per-Gaussian corrective error separately at joint neighborhoods (shoulders, elbows, hips, knees) versus torso and limbs; if the joint-region error gap between teacher and distilled model grows well beyond the Table 1 average gap, the low-frequency/linear-subspace assumption is falsified rather than merely occasionally stressed.","tokens_in":15666,"feed_emoji":"🥽","tokens_out":6801,"duration_ms":46919,"temperature":0.7,"pith_summary":"This paper claims that the expensive neural decoder that predicts how a photorealistic Gaussian avatar deforms with pose can be replaced by a much cheaper linear model without losing much fidelity. The authors train a CNN decoder that emits pose-dependent correctives on a UV map, then distill that decoder by solving a least-squares regression from a 64-dimensional PCA pose basis to the corrective values. They further share one corrective among each 4x4 block of nearby Gaussians, cutting the number of correctives from roughly 60,000 to 4,000. The result is a decoder that runs in 0.45 ms on a Meta Quest 3's NPU, which is fast enough to animate and render three such avatars simultaneously at 72 FPS in VR. If true, photorealistic, drivable full-body avatars move from desktop GPUs to standalone headsets, which is what real-time social VR and telepresence would need.","feed_headline":"Three Gaussian avatars run at 72 FPS on a Quest 3","feed_subtitle":"A 0.45 ms linear decoder replaces the CNN pose model, making photorealistic full-body avatars fit in a VR headset.","key_machinery":"The carrying mechanism is the PCA-linearized pose corrective: a linear map from a 64-dimensional PCA-compressed pose vector to per-Gaussian displacements in position, rotation, scale, and spherical harmonics, fitted by the normal equations so that the pseudo-inverse $(C^T C)^{-1} C^T$ maps decoder outputs to the linear layer's weights. Corrective sharing is the second half of the machinery: the decoder is trained to emit one corrective per 4x4 neighborhood of the UV map (4096 correctives instead of 65536), and nearest upsampling broadcasts each corrective to its block. The linear layer's dominant cost drops from 5.0 ms to 0.45 ms on the Quest 3's NPU, and the paper attributes the feasibility of this compression to pose-dependent correctives being lower-frequency than the static Gaussian fields.","core_discovery":"On its own terms, the paper's central discovery is that pose-dependent Gaussian correctives are low-frequency, and that this property supports a two-stage distillation that preserves quality while making mobile inference practical. A teacher avatar built from a CNN decoder (60k Gaussians, 60k correctives) already matches the quality of Animatable Gaussians with five times fewer Gaussians. The teacher's decoder is then replaced by a two-stage linear model: a least-squares fit from a 64-dim PCA-compressed pose vector, including a ones column, to the corrective map, followed by an expansion of six spherical-harmonic channels into the full SH coefficients. Sharing correctives among 4x4 UV blocks shrinks the linear matrix roughly 16x, reducing decoder latency from 5.0 ms to 0.45 ms. With a custom Vulkan splatting pipeline, this makes it possible to animate and render three full-body Gaussian avatars simultaneously at 72 FPS on a Meta Quest 3, with quantitative quality close to the CNN teacher (L1 0.036 vs 0.039, LPIPS 0.146 vs 0.151 on the combined GCS+linearized model).","pith_inferences":["The low-frequency assumption suggests an immediate extension: use adaptive UV block sizes, small near joints and sleeve/armpit boundaries and large on the torso, to target exactly the artifacts Figure 6 shows in the 4k and linearized models.","The same two-stage recipe (CNN teacher, PCA linearization, spatially shared correctives) should transfer to other animatable Gaussian representations whose bottleneck is pose decoding, such as head avatars or 4D scene models, whenever correctives are smooth in space.","The small gap between the no-decoder and linearized models (0.040 vs 0.039 L1) suggests that after linearization, remaining quality loss comes more from the static template and shared corrective resolution than from the pose model's nonlinearity."],"forward_implications":["A decoder cost of 0.45 ms per avatar is small enough that three full-body avatars can be animated and rendered simultaneously at 72 FPS on a standalone VR headset, meeting the refresh-rate budget VR requires.","The combined GCS+linearized model stays within 0.003 L1 and 0.005 LPIPS of the full CNN teacher on held-out poses and novel views.","Distillation is more reliable than training a linear decoder from scratch, which diverged for one of four identities and scored clearly worse when it did converge.","Because the avatar is driven by on-the-fly pose inputs rather than pre-recorded video, the method supports novel animation, unlike mobile volumetric-video streaming approaches."],"supporting_citations":[{"why":"Defines the 3D Gaussian splatting representation whose pose-dependent attributes SqueezeMe compresses.","marker":"[Kerbl et al. 2023]"},{"why":"Supplies the Animatable Gaussians baseline that the CNN teacher matches with five times fewer Gaussians.","marker":"[Li et al. 2024]"},{"why":"Introduces SMPL and the idea of pose-dependent linear correctives that the distillation replaces the CNN with.","marker":"[Loper et al. 2015b]"},{"why":"Shows linear basis extraction from a pretrained CNN decoder for head avatars, the closest prior step SqueezeMe extends to full bodies.","marker":"[Zielonka et al. 2024]"},{"why":"Provides the compressed-Gaussian rendering acceleration that the custom Vulkan splatting pipeline builds on.","marker":"[Niedermayr et al. 2024]"},{"why":"Represents the mobile volumetric-video streaming alternative that SqueezeMe contrasts with by supporting novel on-the-fly animation.","marker":"[Wang et al. 2024b]"},{"why":"Provides the Sapiens segmentation masks used both as an alpha-loss training signal and for cropping evaluation images.","marker":"[Khirodkar et al. 2024]"},{"why":"Defines the LPIPS perceptual metric used in all quality comparisons.","marker":"[Zhang et al. 2018]"}],"fun_headline_variants":["SqueezeMe: 3 Gaussian avatars at 72 FPS on Quest 3","Mobile Gaussian avatars: real-time on VR headset","Linear decoder compresses Gaussian avatars for mobile VR","Three Gaussian avatars at 72 FPS on a Quest 3","SqueezeMe: Mobile Gaussian avatars run on Quest 3"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that pose-dependent correctives are low-frequency and lie in a low-dimensional linear subspace, so a 64-dim PCA pose basis plus one corrective per 4x4 UV block can reproduce what the CNN decoder learned.","fun_headline_variants_meta":{"raw":{"variants":["SqueezeMe: 3 Gaussian avatars at 72 FPS on Quest 3","Mobile Gaussian avatars: real-time on VR headset","Linear decoder compresses Gaussian avatars for mobile VR","Three Gaussian avatars at 72 FPS on a Quest 3","SqueezeMe: Mobile Gaussian avatars run on Quest 3"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000547,"raw_usage":{"total_tokens":2658,"prompt_tokens":1032,"completion_tokens":1626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":1532}},"tokens_in":648,"tokens_out":1626,"duration_ms":10505,"temperature":1.0,"reasoning_tokens":1532,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:33:59.125173+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the distilled 0.45 ms model and the CNN teacher, feed both a held-out set of extreme poses that the 512 distillation key frames do not cover, and measure per-pixel LPIPS and per-Gaussian corrective error separately at joint neighborhoods (shoulders, elbows, hips, knees) versus torso and limbs; if the joint-region error gap between teacher and distilled model grows well beyond the Table 1 average gap, the low-frequency/linear-subspace assumption is falsified rather than merely occasionally stressed.","supporting_citations":[{"cited_title":"Sapiens: Foundation for human vision models","cited_arxiv_id":null,"evidence_quote":"Provides the Sapiens segmentation masks used both as an alpha-loss training signal and for cropping evaluation images."}],"review_version":1}