{"id":"99c844f8-f2ae-4991-922f-c2685801467e","arxiv_id":"2604.13030","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Hierarchical Binary Quantization plus global refinement AR yields 0.56 rFID reconstruction and 1.81 gFID class-conditional generation on ImageNet, with competitive T2I/T2V at 2B scale.","lead":"GRN pairs hierarchical binary quantization with iterative global token refinement to generate images and videos from discrete latents. It reports new ImageNet records for reconstruction and class-conditional generation while adding entropy-based adaptive compute.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Empirical success of random global refinement is demonstrated only in limited regimes; no analysis confirms it continues to correct rather than re-introduce errors as token count grows.","rationale":"The reader correctly isolates the weakest link: the random-refine schedule is only ablated on small C2I models and lacks any theoretical or large-scale empirical guarantee of continued error correction. That concern is real but does not overturn the reported numbers, which are obtained inside the regimes that were actually trained and evaluated. Reconstruction (0.56 rFID) is unambiguously state-of-the-art among the listed tokenizers; generation (1.81 gFID) is best among pure discrete AR models of comparable size even if continuous/hybrid methods are lower. T2I/T2V results at 2 B are competitive once model scale is controlled. Because the paper releases code and models, the concrete instrumentation above can be run immediately; until it fails, the empirical claims stand and the ACCEPT verdict remains appropriate. No internal inconsistency or circular derivation was found.","tokens_in":23272,"tokens_out":590,"duration_ms":24506,"concrete_test":"On a held-out ImageNet validation batch and a 5-second 480p video, instrument the sampling loop to record, for every token and every step t > 10, whether that token was newly filled, refined (prediction changed), or erased (re-randomized). Compute the late-stage erasure rate and the accuracy of tokens that were already selected by step 15. If late erasures exceed ~15 % or the accuracy of early tokens declines by more than 5 points, the correction claim is undermined; otherwise the mechanism is empirically stable in the tested regimes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GRN upgrades AR generation via a global refinement mechanism (Eqs. 4–6 and Alg. 3) that “progressively perfects and corrects” rests on the training–inference consistency of uniform random selection of tokens. Training mixes ground-truth and random tokens uniformly; inference starts from pure noise and progressively overwrites with model predictions while still allowing random erasures. Table 7 and Appendix E.2 show that pure masking (no erasure/refine of prior tokens) collapses on the 130 M C2I model, establishing that the erase capability is essential. However, the paper supplies no measurement of how often erasures versus refinements actually occur during successful sampling, nor any scaling experiment that increases spatial/temporal token count by the 4–100\times factors required for 1024^{2} images or multi-second video. Without that evidence the mechanism’s claimed error-correction property remains an untested extrapolation beyond the ImageNet 256 and short-video regimes where the headline numbers were obtained.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Generative Refinement Networks (GRN), an autoregressive visual synthesis framework that combines Hierarchical Binary Quantization (HBQ) for near-lossless discrete tokenization with a global fill/refine/erase generation loop and entropy-guided adaptive sampling. HBQ maps continuous VAE features into multi-round binary labels whose reconstruction error decays exponentially (Eqs. 1–3, Alg. 1), yielding discrete image/video tokenizers that match continuous baselines at higher compression (Tables 1–2: 0.56 rFID on ImageNet; near-parity rFVD/PSNR for video). Generation starts from a random token map and iteratively updates a hybrid state Ft (Eq. 4) by predicting all tokens and randomly selecting a growing fraction lt of them (Eqs. 6, 9; Algs. 2–3), allowing both correction of prior mistakes and progressive filling. On ImageNet 256×256 the largest model reports 1.81 gFID; at 2 B scale the same architecture is competitive or superior to same-size AR and diffusion baselines on GenEval (T2I) and VBench (T2V). Code and models are released.","tokens_in":23581,"tokens_out":1093,"duration_ms":10707,"significance":"If the reported numbers hold under public re-implementation, the work supplies two concrete advances: (i) a discrete tokenizer that closes the long-standing reconstruction gap to continuous VAEs without expanding latent channels, and (ii) an AR generation loop that demonstrably mitigates irreversible error accumulation (Table 7). The combination yields new ImageNet records and competitive T2I/T2V results at modest scale, while the open release of tokenizers, models and training code makes the claims immediately falsifiable and extensible. The entropy-aware schedule further offers a practical route to complexity-adaptive compute that diffusion models currently lack. These contributions are of clear interest to the visual generation community.","major_comments":[{"comment":"The central claim that the global refinement loop “progressively perfects and corrects” (Abstract, §3.2, Alg. 3) rests on training–inference consistency of uniform random selection (Eqs. 4–6). Table 7 shows that pure masking collapses on the 130 M C2I model, establishing that erasure is essential, yet the paper never measures the actual frequency of fill / refine / erase operations during successful sampling, nor does it report any scaling experiment that multiplies spatial/temporal token count by the 4–100× factors required for 1024^{2} images or multi-second video. Without that evidence the error-correction property remains an extrapolation beyond the regimes where the headline numbers were obtained; a short diagnostic (e.g., fraction of tokens overwritten after step t0, or FID vs. token count) would make the claim load-bearing rather than anecdotal.","section":null},{"comment":"T2I results (Table 4) rely on a proprietary high-quality fine-tune set whose size and composition are not disclosed (§4.3). While the pre-training corpus is public, the final GenEval score of 0.76 cannot be reproduced from the released materials alone. Either release a public fine-tune subset or report the same metrics after pre-training only, so that the contribution of the proprietary stage can be isolated.","section":null}],"minor_comments":[{"comment":"Eq. (9) for the entropy schedule introduces free parameters k, b, t0, α whose sensitivity is only partially ablated (Table 11). A one-sentence statement of the default values used for all main tables would improve reproducibility.","section":null},{"comment":"Figure 2 and the surrounding text claim “theoretically near-lossless” quantization; the exponential bound is correct, yet the practical gap to continuous reconstruction still depends on M and λ_GAN (Table 2, Appendix E.1). Soften the wording to “exponentially vanishing quantization error.”","section":null},{"comment":"In Table 3 the IS of GRN-G (299.0) is lower than several smaller models; a brief remark on the FID–IS trade-off under the chosen CFG schedule would help readers interpret the ranking.","section":null},{"comment":"Appendix E.2 (confidence vs. random selection) is important but buried; a one-sentence pointer in the main text of §3.2 would strengthen the design justification for random St.","section":null},{"comment":"Typographical: “V AEs”, “V AR”, “re-writer” appear inconsistently; standardize to “VAEs”, “VAR”, “rewriter”.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The proprietary fine-tune data is the only material obstacle to full reproducibility; if the authors cannot release it, the GenEval claim should be clearly caveated. Otherwise the empirical package (tables, ablations, code release) is unusually complete for a visual-generation submission and the paper is a solid fit for a top venue after the two major points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The two things worth knowing: HBQ finally gets discrete latents to continuous reconstruction quality without blowing up channel count, and the generation loop is not next-token or next-scale or mask-and-freeze—it randomly fills, refines, and erases the whole map at every step. That combination is what lets them claim 0.56 rFID and 1.81 gFID on ImageNet 256 while also running adaptive step counts.\n\nHBQ is the cleanest part. Successive binary buckets with exponential error decay is simple, keeps the latent dimension fixed, and the tables (especially the video ones with M=4/6/8) show it really does close the gap to continuous VAEs at higher compression. The refinement training is also straightforward: mix ground-truth and random tokens uniformly, train to predict the full clean map. Inference starts from noise and progressively overwrites while still allowing erasures. The ablation that pure masking collapses (FID 185 → 18 even after retuning) is convincing evidence that the erase capability is doing real work, not just marketing.\n\nSoft spots are real but limited. There is no measurement of how often the model actually erases versus refines during successful runs, and no scaling experiment that multiplies token count by the factors needed for 1024² or long video. So the claim that the loop “progressively perfects and corrects” is demonstrated only inside the regimes they trained (ImageNet 256, short 480p clips). Proprietary high-quality data for the final T2I fine-tune is also a mild asterisk on the GenEval number. Free parameters (M, k/b schedule, CFG interval) exist, but they are the usual ones and they grid-searched them.\n\nThis is for people who care about discrete AR pipelines, tokenizer design, or LLM–vision unification. The math is elementary, the benchmarks are external and public, code and models are released. I would send it to referees; the empirical package is strong enough that the open questions about long-horizon stability can be handled in revision rather than desk rejection.","headline":"Clean discrete tokenizer plus a real global refine/erase AR loop that actually posts new ImageNet numbers and ships code.","tokens_in":24160,"tokens_out":514,"would_cite":true,"duration_ms":10713,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A discrete visual generator that refines like an artist and spends compute only where images are hard.","keywords":["generative refinement networks","hierarchical binary quantization","autoregressive visual generation","adaptive-step sampling","discrete visual tokenizer","text-to-image","text-to-video","ImageNet"],"falsifier":"Train the identical architecture with a pure mask schedule (no re-writing of already-chosen tokens) and measure whether ImageNet FID collapses back toward the 18-point regime shown in the paper’s own ablation; or measure whether the entropy schedule still yields stable videos once clip length exceeds the 2–10 s range used for training.","tokens_in":24174,"feed_emoji":"🎨","tokens_out":892,"duration_ms":7588,"temperature":0.7,"pith_summary":"Diffusion models waste the same number of steps on every image, easy or hard. Standard autoregressive models know which samples are hard via likelihood, but they are held back by lossy tokenizers and by the fact that once a token is written it can never be fixed. This paper introduces Generative Refinement Networks (GRN) to close both gaps. First it replaces ordinary discrete codes with Hierarchical Binary Quantization (HBQ), a multi-round binary scheme whose reconstruction error falls exponentially and matches continuous autoencoders at higher compression. On top of those tokens GRN starts from a random map and, at every step, can fill blank regions, revise earlier tokens, or erase mistakes—exactly as a painter works. An entropy schedule then decides how many steps each sample needs. On ImageNet the method sets new records for both reconstruction and class-conditional generation; at a 2 B scale it also beats same-size diffusion and AR baselines on text-to-image and text-to-video benchmarks.","feed_headline":"Visual AR that rewrites mistakes and spends steps only where needed","feed_subtitle":"Near-lossless binary tokens plus artist-style refinement set new ImageNet records and beat same-size T2I/T2V models","key_machinery":"Hierarchical Binary Quantization (HBQ) plus the generative refinement loop: each latent is repeatedly binary-quantized so error decays as 1/2^j; generation then starts from a random token map and, guided by a rising selection ratio, rewrites every token at every step so earlier mistakes can be corrected.","core_discovery":"A discrete visual generator can simultaneously match continuous reconstruction quality and exceed fixed-step diffusion models if it uses Hierarchical Binary Quantization for near-lossless tokens and then generates by progressive global refinement (fill / revise / erase) under an entropy-guided step schedule.","pith_inferences":["Because HBQ already produces ordered coarse-to-fine bits, the same tokenizer could support progressive decoding or partial-quality streaming without retraining.","The fill–revise–erase loop is formally close to a masked diffusion process; the two families may therefore share the same distillation techniques for further speed-ups.","If the random-selection schedule remains stable at higher resolution, GRN-style models could become the default drop-in replacement for fixed-step diffusion backbones in production pipelines that care about per-sample latency."],"forward_implications":["Discrete visual tokenizers no longer need larger latent channels or huge codebooks to match continuous reconstruction quality.","Autoregressive generators can allocate fewer steps to easy samples and more steps to hard ones without an external controller.","Error accumulation, the classic failure mode of next-token and next-scale AR models, can be mitigated by allowing every token to be rewritten at every step.","A single discrete-token backbone becomes a practical candidate for joint text–image–video modeling inside large language models."],"fun_headline_variants":["AR that rewrites mistakes via progressive refinement and entropy steps","Near-lossless HBQ tokens unlock artist-style AR fill-revise-erase","Complexity-aware visual AR with binary quant and mid-gen corrections","HBQ + global refinement: discrete AR matching continuous quality","GRN: entropy-guided adaptive steps on near-lossless visual tokens"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That the random mixture of real and noise tokens used in training continues to let the model correct, rather than re-introduce, errors when the same random selection is used at inference for longer videos or higher resolutions.","fun_headline_variants_meta":{"raw":{"variants":["AR that rewrites mistakes via progressive refinement and entropy steps","Near-lossless HBQ tokens unlock artist-style AR fill-revise-erase","Complexity-aware visual AR with binary quant and mid-gen corrections","HBQ + global refinement: discrete AR matching continuous quality","GRN: entropy-guided adaptive steps on near-lossless visual tokens"]},"model":"grok-4.5","effort":"low","cost_usd":0.00379,"raw_usage":{"total_tokens":1205,"prompt_tokens":764,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":37900000,"prompt_tokens_details":{"text_tokens":764,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":365,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":764,"tokens_out":76,"duration_ms":3976,"temperature":1.0,"reasoning_tokens":365,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T20:59:23.163688+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the identical architecture with a pure mask schedule (no re-writing of already-chosen tokens) and measure whether ImageNet FID collapses back toward the 18-point regime shown in the paper’s own ablation; or measure whether the entropy schedule still yields stable videos once clip length exceeds the 2–10 s range used for training.","supporting_citations":[],"review_version":2}