{"id":"fbd89791-e25f-4335-b6e9-5e62893bdb9f","arxiv_id":"2506.23479","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A learnable initialization network plus short fine-tuning produces 2D Gaussian image representations faster than GaussianImage, with adaptive Gaussian counts per image.","lead":"This paper presents a neural network that quickly creates a coarse set of 2D Gaussians to represent an image, then fine-tunes briefly to match a slower baseline method. It claims up to 10x faster training while adapting the number of Gaussians to image complexity.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 14's GT color sampling may make the 2s-vs-20s speed claim an artifact of reading the target image; the paper never states whether colors become standalone parameters after fine-tuning.","rationale":"I read the paper in good faith. The proposed pipeline is a plausible generalizable initialization and adaptive point-placement scheme, and several components (PPM supervision, Floyd-Steinberg dithering, Delaunay-based scale references) are coherent. The concern is not that using the target image to set initial colors is automatically unfair; many fitting methods use the target image. The issue is that the paper never draws the boundary. Eq. 14 is stated as producing the final color attribute, and the opacity field is learned in place of color, reducing the learned color parameters from three to one. If the final stored Gaussians include the ci values sampled from I, the representation is standalone and the speed comparison is legitimate. If the implementation re-samples I during fine-tuning or rendering, then the same-Gaussian-count comparison is not a fair test: GaussianImage must discover colors, while Instant-GI reads them from ground truth. The released code makes this checkable with a one-line surgical ablation. I agree with the reader's weakest assumption, so the verdict should remain conditional until the model is shown to render without the target image. I do not see a reason to reject outright: the ambiguity is specific and resolvable, and the secondary issues noted by the reader (missing Image-GS comparison, Table 1 parameter-count presentation) do not change this assessment.","tokens_in":17724,"tokens_out":6272,"duration_ms":77669,"concrete_test":"Run the released code on image 0844×2: (1) locate every grid_sample(I) call in the fine-tuning and render paths; (2) after the Net. Init. forward pass, store ci as a constant parameter, delete I from GPU memory, and continue the 2s and 20s fine-tune and final render using only stored Gaussians; (3) if PSNR stays within 0.3 dB and rendering succeeds, the representation is standalone; if PSNR drops substantially or the pipeline crashes, Eq. 14's ground-truth access is load-bearing and the headline speed claim is invalid. Report whether any render-time code path reads I.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 defines the Gaussian color as ci = I(pi)·σ(o) (Eq. 14), sampling the target image at the Gaussian center. This makes color a ground-truth lookup rather than a learned parameter. The headline result in Sec. 4.3 is that after 2s of network initialization plus fine-tuning, PSNR reaches 46.68 on image 0844×2 versus 45.90 for random initialization after 20s, with the same Gaussian count. If Eq. 14 is reevaluated during the fine-tuning loop, or if render-time code still reads colors from I, then the comparison conflates 'fitting Gaussians' with 'reading colors from the ground-truth image.' The resulting representation is not self-contained: rendering still requires I, which contradicts the image-representation and compression claims. The paper never states whether, after the 2s fine-tune, the ci values are detached and stored as ordinary Gaussian color parameters, or whether I is accessed again. The claim that 'all compression techniques for GI can be applied directly' assumes the former, while Eq. 14 and the opacity-only prediction suggest the latter. This ambiguity is load-bearing because the order-of-magnitude speedup rests on that 2s comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Instant-GI, a generalizable initialization method for 2D Gaussian splatting image representation. A ConvNeXt-based UNet extracts features, a Position Field predicts a Gaussian Position Probability Map supervised by a pseudo-PPM computed from a converged GaussianImage model, and Floyd-Steinberg dithering discretizes this map to adaptively select Gaussian locations. A second set of MLP fields predicts position, scaling, rotation, and opacity, and the final Gaussian color is defined in Eq. (14) as the image color sampled at the Gaussian center multiplied by the predicted opacity. After a feed-forward initialization, the representation is fine-tuned with a small number of iterations. Experiments on Kodak and DIV2K report PSNR comparable to or better than GaussianImage with fewer iterations and, for three selected images, show large PSNR gains at 2s compared with a randomly initialized baseline at 20s.","tokens_in":18027,"tokens_out":7407,"duration_ms":79929,"significance":"If the correctness issues are resolved, the paper addresses a real limitation of GaussianImage: slow training from random initialization and a fixed number of Gaussians. The idea of a learned prior for Gaussian placement and a dithering-based adaptive count is practical and the code is promised publicly. The best-case result, matching or exceeding GaussianImage in a fraction of the training time, would be useful for image representation and compression. However, the current manuscript leaves open a load-bearing question: whether the final representation is self-contained or still requires the original image to render. Because Eq. (14) defines colors as direct samples from the target image, the headline speed advantage may be partly an artifact of ground-truth access. The pseudo-PPM supervision from a converged GaussianImage model also means the 'self-adaptive' behavior is a distillation of the baseline rather than an independent entropy-based criterion. These issues need to be clarified and re-tested before the central claims can be accepted.","major_comments":[{"comment":"The definition ci = I(pi)·σ(o) means that, as written, the Gaussian color is read from the target image at the Gaussian center. The paper never states whether this expression is evaluated only during the network's feed-forward pass and then detached, or whether the renderer continues to access I during fine-tuning and at inference. The headline result in Sec. 4.3 (2s vs 20s) is only a fair test of an image representation if the final stored Gaussians do not require I for rendering; otherwise the speed advantage is in part an artifact of querying ground-truth colors. Please clarify this explicitly, and if colors are detached, state so and include an ablation that replaces I(pi) with a learned color to quantify the contribution of this initialization.","section":"Sec. 3.4, Eq. (14)"},{"comment":"The pseudo PPM used to supervise the Position Field is computed from a converged GaussianImage model (Supplementary Sec. 2.3). The claim that the method 'adaptively determines the number of Gaussians based on information entropy' is therefore a claim about matching the baseline's own placement, not an independently derived entropy-based criterion. This is a form of self-distillation; it should be acknowledged and tested against the alternative of sampling directly from the pseudo PPM at test time to show that the learned PPM adds robustness or quality beyond the teacher.","section":"Sec. 3.3 and Supplementary Sec. 2.3"},{"comment":"The order-of-magnitude speedup is supported by PSNR measurements on only three images, with no standard deviations or repeated-run statistics. In addition, the '2s' interval must be specified precisely (network forward pass, dithering, Delaunay triangulation, number of fine-tuning iterations) and, given Eq. (14), the reported 2s PSNRs may include the effect of GT color lookup. Please report metrics over a larger set with error bars and an experiment that isolates the contribution of GT color sampling from the contribution of position, scaling, and opacity initialization.","section":"Sec. 4.3, Table 2"}],"minor_comments":[{"comment":"The second GI row appears as 'GI ' rather than 'GI†' in the printed table, and the dagger is not visibly attached to the row. Please ensure the marker matches the caption and report the number of Gaussians for each row so the 'same number of Gaussians' comparison is easy to verify.","section":"Table 1"},{"comment":"The GPU memory column for Net. Init. lists '3038 + 408' but the text says 'approximately 3GB'; clarify what the two numbers represent.","section":"Sec. 4.3, Table 2"},{"comment":"There are several typographical errors: 'hass' in Sec. 2.1, 'availiable' in the abstract, 'upper bounder' in Sec. 4.2, and 'x g 0' in Eq. (11) should be 'x ≥ 0'.","section":"Throughout"},{"comment":"The sentence 'Kodak's lower resolution, leading to a 61% reduction in Gaussian points' should be rephrased: it was not Kodak's resolution per se but the adaptive algorithm that selected fewer Gaussians; please state the comparison baseline for the 61% reduction.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The central ambiguity around Eq. (14) must be resolved before this paper can be considered. If the implementation already detaches the sampled colors into standalone per-Gaussian parameters, the authors should state this prominently and provide an ablation with learned colors to show that the speedup is not solely due to reading the target image. The pseudo-PPM supervision is essentially distillation from the baseline; the paper should present it as such and compare against direct use of the pseudo PPM. A comparison with Image-GS, which is the closest content-adaptive Gaussian image representation, would also help position the novelty. These are substantial but addressable concerns, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine step forward for practical Gaussian image representation, and the speedups in Table 2 are striking. The main thing to check before trusting the headline is whether Eq. 14 makes the speed comparison partly circular, and the paper never says explicitly whether colors become standard learned parameters after fine-tuning. That said, the Color Field ablation suggests the concern is not as load-bearing as it first appears.\n\nThe new combination is real: a ConvNeXt-UNet predicts a position probability map and all Gaussian attributes in one forward pass, Floyd–Steinberg dithering sets point counts adaptively, and Delaunay triangulation gives normalized scale references. This is not in GaussianImage or Image-GS. The ablation study is honest and useful, especially the Color Field row, which replaces the ground-truth color sample with a learned color field and only loses a small PSNR at 2s. That is decent evidence the speedup is not purely an artifact of reading colors from the target image.\n\nSoft spots, in rough order of severity. First, Eq. 14 defines ci = I(pi)·σ(o), and the text calls this the \"initialized\" color. If the fine-tuning loop re-evaluates that sampling, the renderer still needs the original image, which would undermine the representation/compression claim and make the 2s-vs-20s comparison unfair. The paper should state plainly whether ci is detached and optimized as a parameter after the network pass. The Color Field ablation makes me think this is fixable, but the published version needs to settle it. Second, Table 1 is genuinely confusing: two GI rows with different parameter counts, dagger notation only in the footnote, and no clean apples-to-apples line. Third, the headline speed claim rests on three images without error bars or multiple runs. That is fine for a demonstration but not for a benchmark. Fourth, there is no comparison to Image-GS, the closest sibling method; that omission weakens the positioning. The pseudo-PPM being generated from a converged GaussianImage model is a reasonable distillation choice, not a circularity.\n\nWho should read this? Anyone working on Gaussian-based image representation or fast initialization schemes for explicit representations. It deserves a serious referee; the pipeline is original and the experiments mostly support the claims if the color ambiguity is resolved. I would send it to review and ask for a revision that clarifies the color parameterization, adds Image-GS, and cleans up Table 1.","headline":"Network-initialized 2D Gaussians with dithering-based adaptive counts give a real speedup; the color-sampling ambiguity in Eq. 14 needs clarification before the headline claim is fully trusted.","tokens_in":18523,"tokens_out":4611,"would_cite":true,"duration_ms":51438,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A generalizable network initializes 2D Gaussian splats in one forward pass, then minimal fine-tuning matches or beats GaussianImage at a fraction of the training time.","keywords":["2D Gaussian Splatting","Image Representation","Implicit Neural Representation","Generalizable Initialization","Floyd–Steinberg Dithering","Position Probability Map","Self-Adaptive Gaussian Count","Fast Image Rendering"],"falsifier":"Inspect the inference code path after fine-tuning: if rendering a Gaussian representation still requires the original image to supply the color term I(pi) of Eq. (14), then the representation is not self-contained and the training-time comparison is unfair. Alternatively, retrain the fine-tuning stage with color as a free parameter while hiding the source image, and check whether the reported ~10x speed-up and PSNR gains persist.","tokens_in":17512,"feed_emoji":"🖼️","tokens_out":6147,"duration_ms":54437,"temperature":0.7,"pith_summary":"Instant-GI proposes a generalizable image representation based on 2D Gaussian Splatting. A ConvNeXt-based UNet extracts features and predicts a Position Probability Map (PPM), which is discretized by Floyd–Steinberg dithering to place Gaussians adaptively. A set of MLP fields then predicts each Gaussian's position, scale, rotation, and opacity, with colors sampled directly from the input image. After a single feed-forward pass and minimal fine-tuning, the rendered image matches or exceeds the quality of GaussianImage, which trains from random initialization for much longer. The same network also controls the number of Gaussians via a kernel-size parameter, allowing a compression-quality trade-off without retraining.","feed_headline":"Feed-forward seeding cuts 2D Gaussian image training to a tenth","feed_subtitle":"A generalizable network pre-initializes Gaussian splats; brief fine-tuning matches or beats GaussianImage on PSNR.","key_machinery":"The central object is the Position Probability Map (PPM), a per-pixel probability of placing a Gaussian, predicted by a small MLP from a ConvNeXt UNet feature map. Floyd–Steinberg dithering discretizes the PPM, distributing a self-adaptive number of Gaussians that matches local image entropy. Delaunay triangulation on the sampled points yields triangles whose fitted ellipses provide a geometric reference scale and orientation for the Gaussian fields; MLP fields predict barycentric positions, scale offsets, rotation bias, and opacity, while color is obtained by sampling the input image at the Gaussian center and multiplying by a sigmoid-activated opacity. This design turns Gaussian decomposition into a feed-forward operation followed by a short fine-tune.","core_discovery":"The paper's central claim is that a network-initialized Gaussian decomposition can replace random initialization in GaussianImage, achieving comparable or better rendering fidelity in about one tenth of the training time. The key finding is that with the same number of Gaussians, starting from a network's structured prediction—rather than uniform random positions—lets optimization converge to a higher PSNR in far fewer iterations. On the DIV2K ×2 test set the method reports 42.80 dB versus 40.26 dB for GaussianImage with equal Gaussian count; on Kodak it reports 42.92 dB, slightly below GaussianImage's 44.08 dB but with fewer points and more consistent performance across datasets. The paper also demonstrates that rendering quality after 2 seconds of network initialization and fine-tuning can exceed the 20-second result of random initialization.","pith_inferences":["A testable implication is that the same network could amortize initialization across many images in a video sequence, potentially turning per-frame Gaussian fitting into a short residual update.","Because colors are sampled from the source image during initialization, a fair standalone representation requires that fine-tuning converts those colors into independent parameters; if the image remains needed at inference, the reported speed gains over GaussianImage are partly an artifact of information leakage.","The kernel-size control could serve as an explicit rate-distortion knob for progressive compression, but the available experiments only demonstrate it on kernel sizes 3 and 4.","The PPM supervision depends on a pre-trained GaussianImage decomposition; a different teacher model would likely shift both the point distribution and the final rendering quality."],"forward_implications":["Training time for a usable Gaussian image representation drops from tens of seconds to a few seconds: on image 0844, the method reaches 46.68 dB PSNR after 2 s including initialization, versus 45.90 dB after 20 s with random initialization.","The number of Gaussians can be adjusted at test time by changing the dithering kernel size k (e.g., from 3 to 4), trading quality for compression without retraining the network.","Because the Gaussian model matches GaussianImage, existing compression techniques for GaussianImage apply directly to the method's output.","On the DIV2K ×2 dataset the network-initialized representation reaches 42.80 dB PSNR, outperforming GaussianImage's 40.26 dB under the same Gaussian count.","Network-based initialization also yields smoother spatial distributions than gradient-based point placement, avoiding over- and under-sampling in high- and low-entropy regions."],"supporting_citations":[{"why":"Supplies the GaussianImage baseline, the Gaussian model, and the rendering equation that the method builds on and compares against.","marker":"[48]"},{"why":"Provides the 3D Gaussian Splatting framework, including differentiable rasterization and the Gaussian primitive concept.","marker":"[19]"},{"why":"Provides the Floyd–Steinberg dithering algorithm (GPU-accelerated) used to discretize the Position Probability Map.","marker":"[14]"},{"why":"Provides Delaunay triangulation, used to structure sampled points and derive triangles for geometric features.","marker":"[5]"},{"why":"Provides the conic/ellipse fitting method used to compute the geometric reference ellipses for each triangle.","marker":"[13]"},{"why":"Provides the ConvNeXt architecture used as the backbone of the feature extraction UNet.","marker":"[29]"},{"why":"Supplies the anchor-based offset prediction idea that motivates predicting scale offsets relative to fitted ellipses.","marker":"[31]"},{"why":"Describes an alternative content-adaptive Gaussian sampling method (Image-GS) based on gradients and CDF, compared in the ablation.","marker":"[49]"}],"fun_headline_variants":["Network-seeded Gaussians cut image training time tenfold","Self-adaptive 2D Gaussian splats train 10x faster","Instant GaussianImage: 10x faster with adaptive Gaussian count","Network initialization speeds up Gaussian splatting images","Adaptive Gaussian splats match rendering in a tenth of time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that sampling a Gaussian's color directly from the input image at its center is a legitimate part of the representation; if the original image is still required to render after fine-tuning, the representation is not standalone and the claimed speed advantage over GaussianImage largely reflects querying ground truth rather than true learning.","fun_headline_variants_meta":{"raw":{"variants":["Network-seeded Gaussians cut image training time tenfold","Self-adaptive 2D Gaussian splats train 10x faster","Instant GaussianImage: 10x faster with adaptive Gaussian count","Network initialization speeds up Gaussian splatting images","Adaptive Gaussian splats match rendering in a tenth of time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1186,"prompt_tokens":905,"completion_tokens":281,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":521,"tokens_out":281,"duration_ms":3560,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:41:08.464975+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the inference code path after fine-tuning: if rendering a Gaussian representation still requires the original image to supply the color term I(pi) of Eq. (14), then the representation is not self-contained and the training-time comparison is unfair. Alternatively, retrain the fine-tuning stage with color as a free parameter while hiding the source image, and check whether the reported ~10x speed-up and PSNR gains persist.","supporting_citations":[{"cited_title":"A buyer's guide to conic fitting","cited_arxiv_id":null,"evidence_quote":"Provides the Floyd–Steinberg dithering algorithm (GPU-accelerated) used to discretize the Position Probability Map."},{"cited_title":"Discontinuity-aware 2d neural fields","cited_arxiv_id":null,"evidence_quote":"Provides Delaunay triangulation, used to structure sampled points and derive triangles for geometric features."},{"cited_title":"Image super-resolution using deep convolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides the conic/ellipse fitting method used to compute the geometric reference ellipses for each triangle."},{"cited_title":"Mvsgaussian: Fast generalizable gaussian splatting reconstruction from multi-view stereo","cited_arxiv_id":null,"evidence_quote":"Provides the ConvNeXt architecture used as the backbone of the feature extraction UNet."},{"cited_title":"Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting","cited_arxiv_id":null,"evidence_quote":"Describes an alternative content-adaptive Gaussian sampling method (Image-GS) based on gradients and CDF, compared in the ablation."}],"review_version":1}