{"id":"3342c82c-ea0a-4de0-86a3-31dbcf7b02a1","arxiv_id":"2505.23660","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"D-AR recasts pixel-space diffusion as vanilla autoregressive next-token prediction using a diffusion-ordered discrete tokenizer, reaching 2.09 FID on ImageNet 256x256 with a 775M Llama backbone.","lead":"The paper introduces D-AR, a way to generate images by turning the diffusion process into a simple next-token prediction over discrete tokens, using an unmodified Llama-style autoregressive model. It reports a 2.09 FID on ImageNet class-conditional generation and shows that partial token generation yields consistent previews and zero-shot layout control.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The schedule sufficiency behind the coarse-to-fine token order is only validated by qualitative previews; without a quantitative consistency check, the central 'diffusion via AR' claim remains under-supported.","rationale":"I read the paper in good faith. The method is well-motivated, the tokenizer design is novel, and the reported FID of 2.09 with a 775M AR backbone is competitive. The central mechanistic claim is that sequential AR generation of discrete tokens directly mirrors the diffusion denoising procedure. The weakest point is indeed the condition schedule: each diffusion step uses only its own group of tokens, and the paper gives no theoretical argument for why this one-group-per-step conditioning is sufficient. I agree with the reader that this is the most load-bearing assumption. However, I would refine the concern: the training objective already imposes a soft pressure on each group to be informative at its assigned timestep, and the good reconstruction rFID suggests the schedule is broadly workable. What is genuinely missing is any quantitative evidence that the advertised 'consistent previews' are consistent at scale. Table 5 reports FID of partial-token previews, but FID does not measure consistency with the final sample; a preview could be a realistic image of a different object and still achieve a reasonable FID. The qualitative figures (Fig. 5, Fig. 11) are suggestive but not sufficient for a central property. This gap is directly testable. If the proposed test shows high semantic agreement between previews and finals, the schedule-sufficiency concern is largely resolved and the central claim stands. If not, the equivalence between AR generation and diffusion, and the preview property, are called into question. The other issues raised by the reader (parameter-count conventions, qualitative layout control, no released code) are real but secondary: the table note explicitly states the tokenizer size, and layout control is a described demonstration rather than a scaled claim. Therefore I do not see a reason to move the verdict; a conditional accept requiring quantitative preview-consistency validation is appropriate.","tokens_in":15489,"tokens_out":19865,"duration_ms":190425,"concrete_test":"Generate 50K images with D-AR-XL using the released model. At the point where 64 tokens (2 groups, 25%) have been generated, save the jump-estimate preview x_hat_1 computed via x_hat_1 = (1-t) v_t + x_t, and also save the final image after all 256 tokens. Then compute (i) average LPIPS between the 25%-token preview and the final sample, and (ii) the percentage of samples for which an off-the-shelf ImageNet classifier (e.g., ViT-L/16) assigns the same top-1 class to the preview and the final sample. If top-1 agreement is below 80% or LPIPS exceeds roughly 0.6, the 'consistent previews' claim fails and the schedule-sufficiency assumption is doubtful. If agreement is high and LPIPS is moderate, the concern is retired.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that D-AR recasts the diffusion process as a vanilla autoregressive procedure hinges on the condition schedule c(t) in Eq. 5 and the fixed 8-step sampler in Eq. 8. These equations assume that each diffusion timestep is sufficiently conditioned by exactly one disjoint group of N/K tokens, consumed once in order. If a later token group carried information needed in an earlier denoising step, the streaming previews and the equivalence between AR token generation and the diffusion trajectory would break. The paper's training loss (Eq. 7) does couple each group to its timestep, so a severely insufficient group would inflate the flow-matching loss; this is an indirect, soft constraint on the learned schedule, not a guarantee. The published evidence for sufficiency is limited to a good reconstruction rFID (1.52) with all tokens and a few qualitative preview examples (Fig. 5, Fig. 11). Table 5 reports FID of partial-token jump estimates, but FID does not measure whether the preview is semantically consistent with the final sample. A preview with good FID could depict a completely different object than the final image. Because 'consistent previews from partial tokens' is one of the three advertised properties in the strongest claim, and because it is the observable consequence of the schedule-sufficiency assumption, the absence of a quantitative consistency metric is the most load-bearing gap. The parameter-count caveat in Table 3 is transparent, and the lack of code is a reproducibility issue, but neither undermines the mechanistic claim as directly as an unvalidated preview consistency would.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D-AR, a framework that bridges diffusion and autoregressive visual generation by introducing a 'sequential diffusion tokenizer.' This tokenizer maps an image into a 1D sequence of 256 discrete tokens and is trained so that successive groups of tokens condition successive timesteps of a flow-matching diffusion decoder on raw pixels. A standard decoder-only Llama-style transformer then performs vanilla next-token prediction on these tokens, and the resulting token generation can be incrementally decoded into diffusion denoising steps. The paper reports a reconstruction rFID of 1.52 for the tokenizer and an ImageNet 256x256 class-conditional generation FID of 2.09 for D-AR-XL (775M AR parameters), along with claimed native support for KV-cache inference, consistent streaming previews from partial tokens, and zero-shot layout-controlled synthesis.","tokens_in":15814,"tokens_out":5860,"duration_ms":57560,"significance":"If the claims are substantiated, D-AR offers a conceptually clean way to reuse standard LLM autoregressive infrastructure for image generation while retaining diffusion-style coarse-to-fine decoding. The experiments are reasonably extensive for a preprint: the paper reports tokenizer ablations over the conditioning group count K and the timeshift parameter beta, compares against several strong baselines, and is transparent about the tokenizer parameter count in a table note. The reported 2.09 FID with a 775M AR backbone is competitive among vanilla autoregressive methods. The main value is the sequential diffusion tokenizer idea, which is novel relative to prior tokenizers with diffusion decoders. The primary weakness is that two of the advertised properties, consistent streaming previews and zero-shot layout control, are supported only by qualitative examples and distribution-level FID numbers, not by a direct consistency evaluation.","major_comments":[{"comment":"The headline parameter count of 775M for D-AR-XL counts only the autoregressive backbone and excludes the 300M sequential diffusion tokenizer. While the Table 3 note discloses this, the abstract, introduction, and conclusion repeat the 775M figure without the caveat, and the surrounding text compares D-AR-XL to IBQ-XXL 2.1B as if the full systems were comparable. Please report total system parameters (1.075B for D-AR-XL) in the headline statements and revisit the parameter-region comparisons, or clearly state that all compared methods exclude tokenizer parameters and provide the tokenizer budget for each.","section":"§5.1, Table 3"},{"comment":"The claimed property of 'consistent previews' rests on the sufficiency of the fixed condition schedule c(t), namely that each disjoint group of tokens consumed in order provides the conditioning needed at its diffusion step. The quantitative evidence in Table 5, however, consists of FID, IS, precision, and recall of the partial-token jump estimates, which are distribution-level metrics and do not establish whether a preview at, say, 64 tokens is the same image as the final sample at 256 tokens. Please add a paired consistency metric (for instance, LPIPS or segmentation/label agreement between each preview and its own final image, or a same-prefix-multiple-completions study) and report the numbers. Without this, the streaming-preview property is supported only by the qualitative Figures 5 and 11, which is insufficient for a headline claim.","section":"§3.1, Eq. (5); §3.3, Eq. (8); §5.1, Table 5"}],"minor_comments":[{"comment":"The text contains a typo: 'Dffusion via Autoregressive models' should read 'Diffusion via Autoregressive models.'","section":"Section 1"},{"comment":"The phrase 'coodebook utilization' should be corrected to 'codebook utilization.'","section":"Appendix III"},{"comment":"The statement that Eq. (8) is the 'reversed form' of Eq. (5) is ambiguous: Eq. (8) is not the algebraic inverse of Eq. (5). Please clarify the exact relationship between these two schedules.","section":"Eq. (5) and Eq. (8)"},{"comment":"The zero-shot layout-controlled synthesis property is currently demonstrated only with qualitative examples (Figure 6 and Figure 12). If this is retained as a headline property, a quantitative layout-fidelity metric or a comparison with a simple conditioning baseline would strengthen the claim.","section":"§5.1, Zero-shot layout-controlled synthesis"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The central idea is sound and the experimental core is competently executed, but the missing quantitative consistency check for streaming previews is a real gap for one of the paper's three advertised properties, and the parameter-count presentation needs adjustment. I do not see a circularity problem; the concern is about evidence strength, not internal consistency. The relation to DDT-LLaMa is discussed, but during revision the authors should make the technical distinction crisper, since both works use diffusion decoders with sequential discrete conditioning. The paper fits the scope of the journal if the requested revisions are made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nHere is my take on arXiv:2505.23660. The core idea is genuine and worth taking seriously: a sequential diffusion tokenizer that groups discrete tokens by diffusion timestep, trained with flow matching and quantized conditioning, lets a stock Llama-style next-token predictor reach 2.09 FID on ImageNet 256 with 256 tokens. That is a real result. The distinction from DDT-LLaMa is also real: reversing the token order and not supporting partial-token decoding are structural differences, not cosmetic ones.\n\nWhat the paper does well: it builds a clean pipeline with standard cross-entropy loss, reports honest ablations on beta and K, and includes a limitation section that acknowledges the moderate scale and missing text-to-image generation. The rFID of 1.52 with 256 tokens is competitive, and the qualitative trajectory/preview figures are suggestive. The related-work discussion of diffusion-decoder tokenizers and DART/MAR is fair, and the novelty claim about the timestep-ordered condition schedule holds up against the cited prior work.\n\nThe soft spots, in decreasing order of importance. First, 'consistent previews' is advertised as a headline property, but the evidence is qualitative plus partial-token FIDs in Table 5. FID does not tell you whether the early jump-estimate shows the same object as the final sample. The stress-test note is right on this: without a quantitative consistency metric, say CLIP or DINO feature similarity between preview and final, the most distinctive claimed property is under-supported. I do not think this is fatal; the loss couples each token group to its timestep, so the schedule-sufficiency assumption is plausible. But it is an assumption, and it deserves a direct test. Second, Table 3 counts only AR parameters. D-AR's tokenizer is 300M versus roughly 72M for LlamaGen, and the 185M diffusion decoder runs at every sampling step. The footnote is transparent, but the headline comparison still overstates efficiency. Report total parameters and approximate inference FLOPs. Third, no code or models at submission. That is a reproducibility issue, especially for an architectural method others will want to build on. Fourth, the layout-control evidence is qualitative only, which is minor given the mechanism.\n\nOverall, the central architectural claim — that diffusion-ordered discrete tokens enable vanilla AR generation with competitive quality — holds up well enough to deserve serious refereeing. For anyone building unified multimodal AR/LLM stacks, this is a concrete route worth engaging. I would send it to reviewers rather than desk reject, and I would ask for a quantitative preview-consistency metric and a fairer efficiency comparison before acceptance.","headline":"Diffusion-ordered discrete tokens let a vanilla Llama produce competitive ImageNet FID; the architecture is real, but preview consistency needs a quantitative metric before I'd call the headline property proven.","tokens_in":16347,"tokens_out":6890,"would_cite":true,"duration_ms":58852,"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":"D-AR recasts image diffusion as a vanilla autoregressive next-token prediction problem, and shows that a 775M Llama-style transformer generates ImageNet 256×256 images at 2.09 FID with 256 discrete tokens while natively supporting…","keywords":["diffusion models","autoregressive image generation","sequential diffusion tokenizer","next-token prediction","flow matching","visual tokenization","ImageNet class-conditional generation","KV cache"],"falsifier":"Swap two token groups at decode time (for example, feed group 2's tokens as the condition for the first denoising step) and measure FID and layout similarity against the original schedule; if quality and coarse structure do not collapse, the claimed coarse-to-fine ordering is not actually determining the output, and the central equivalence claim fails.","tokens_in":15290,"feed_emoji":"🖼","tokens_out":6531,"duration_ms":57329,"temperature":0.7,"pith_summary":"Diffusion via Autoregressive models (D-AR) claims that the image diffusion process can be recast as a standard autoregressive next-token prediction problem, with no changes to the causal mask, loss, or inference loop of a vanilla LLM-style transformer. The paper builds a sequential diffusion tokenizer that maps an image to 256 discrete tokens, with later tokens feeding later denoising steps of a pixel-space flow-matching decoder, so autoregressive token generation directly mirrors diffusion denoising. On ImageNet 256×256, D-AR-XL (775M parameters) achieves 2.09 FID, outperforming vanilla autoregressive baselines and matching much larger models, while gaining native KV-cache inference, consistent partial-token previews, and zero-shot layout condition from prefix tokens. The significance is architectural: it suggests LLM stacks need no vision-specific modifications to do diffusion-quality generation.","feed_headline":"Image diffusion becomes plain next-token prediction in D-AR","feed_subtitle":"A 775M Llama-style model reaches 2.09 FID on ImageNet with 256 tokens, gaining KV-cache speed and streaming previews.","key_machinery":"The sequential diffusion tokenizer: a transformer encoder with learnable query tokens and vector quantization produces 256 ordered discrete tokens, and a 185M-parameter diffusion transformer decoder takes, at each flow-matching timestep $t$, only the token group $g_{\\lceil t' K \\rceil}$ selected by the shifted condition schedule $c(t)=g_{\\lceil t' K \\rceil}$ with $t'=t/(t+(1/\\beta)(1-t))$. This makes diffusion look like a Markov chain of conditioning groups, so an LLM generating tokens one at a time is literally supplying the next denoising condition. The reverse schedule $t_i = (i/K)/(i/K+\\beta(1-i/K))$ fixes 8 denoising steps, each using one group of 32 tokens.","core_discovery":"The central discovery is that a discrete token sequence can serve as a compressed control program for a diffusion process: instead of sampling noise and iterating a denoiser, one samples tokens left to right, and each group of tokens conditions the next denoising step in pixel space. Because the condition schedule $c(t)$ reads token groups in order as the flow-matching time $t$ goes from 0 to 1, early tokens encode coarse layout and global structure, while later tokens add fine detail, producing a natural linear ordering for autoregressive generation. Training uses standard cross-entropy on discrete tokens plus a flow-matching tokenizer loss with perceptual and representation-alignment terms; decoding uses a fixed 8-step reverse schedule with each of the $K=8$ groups consumed exactly once. The paper shows that this yields state-of-the-art results among vanilla (unmodified) autoregressive image models on ImageNet, and that partial token sequences can be decoded immediately to produce consistent coarse-to-fine previews.","pith_inferences":["If the schedule sufficiency holds beyond the tested regime, the same conditioning-group trick could linearize other coarse-to-fine generative processes (video, audio, 3D) into vanilla autoregression, but the paper explicitly leaves native text-to-image and parameter scales above 1B untested, so that breadth is extrapolation.","The fixed $K=8$ reverse schedule is a design choice rather than a necessity; an adaptive schedule where later groups re-condition earlier timesteps could test whether the Markovian claim survives without the fixed-step implementation.","If the coarse-to-fine ordering is genuinely load-bearing, token count may scale with the number of diffusion steps rather than spatial resolution, so high-resolution generation might stay token-cheap and fit within LLM context windows.","Because D-AR outputs discrete tokens in a standard causal format, it should integrate directly into multimodal LLM pipelines, but the paper does not benchmark wall-clock speedups, so the practical speed advantage remains an invitation rather than a measured result."],"forward_implications":["LLM serving infrastructure such as KV-cache acceleration applies unchanged to high-quality image generation via D-AR, since the image model is a vanilla decoder-only transformer.","Streaming generation gives consistent previews at any partial token count with near-zero extra cost, because the pixel-space diffusion decoder can jump-estimate the target image from whatever condition tokens are available.","Layout-controlled synthesis is achievable zero-shot by fixing a few prefix tokens as conditions, with no finetuning, and the control strengthens as more prefix tokens are provided.","D-AR-XL reaches 2.09 FID with 775M parameters, outperforming LlamaGen-XXL at 1.4B and matching IBQ-XXL at 2.1B within the vanilla autoregressive category.","The sequential diffusion tokenizer reconstructs ImageNet validation images at rFID 1.58 with 256 tokens and a 16384-entry codebook, beating LlamaGen's tokenizer at the same budget (2.19 rFID), and degrades less when the codebook is shrunk."],"supporting_citations":[{"why":"Supplies the 1D query-token encoder design that D-AR adapts for its sequential diffusion tokenizer.","marker":"[51]"},{"why":"Defines the vanilla autoregressive baseline and the Llama-style backbone and VQ tokenizer recipe that D-AR builds on and must beat.","marker":"[8]"},{"why":"Establishes the diffusion-decoder tokenizer training recipe with LPIPS and representation-alignment losses that D-AR's tokenizer training relies on.","marker":"[42]"},{"why":"Shows diffusion decoders as scalable image tokenizers; D-AR's approach builds on and differentiates from this line by sequentializing the conditions.","marker":"[45]"},{"why":"Another diffusion-decoder tokenizer; D-AR contrasts its sequential group schedule against this full-sequence conditioning baseline.","marker":"[44]"},{"why":"Conceptually related work using a diffusion decoder with a reversed token order; D-AR claims its key distinction lies in representing diffusion steps as a sequential AR process decodable from partial tokens.","marker":"[50]"},{"why":"Provides the diffusion transformer architecture used in D-AR's decoder and serves as a strong diffusion baseline in comparisons.","marker":"[53]"},{"why":"Supplies the training recipe and linear classifier-free guidance schedule used for D-AR's autoregressive models.","marker":"[15]"},{"why":"Represents the tailored autoregressive approach that injects diffusion into each AR step; D-AR distinguishes itself by staying fully discrete and vanilla.","marker":"[27]"}],"fun_headline_variants":["Diffusion recast as autoregressive token prediction","D-AR: tokens encode diffusion steps, one by one","Autoregressive model streams diffusion-style image previews","Image generation: next token maps to next denoising step","From noise to tokens: D-AR flips diffusion into AR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each diffusion denoising step needs only its own group of $N/K$ tokens as conditioning, consumed exactly once in a fixed order through the schedule $c(t)$; if later tokens carry information required at earlier denoising steps, the equivalence between autoregressive token generation and the diffusion procedure breaks.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion recast as autoregressive token prediction","D-AR: tokens encode diffusion steps, one by one","Autoregressive model streams diffusion-style image previews","Image generation: next token maps to next denoising step","From noise to tokens: D-AR flips diffusion into AR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1760,"prompt_tokens":1003,"completion_tokens":757,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":675}},"tokens_in":619,"tokens_out":757,"duration_ms":7366,"temperature":1.0,"reasoning_tokens":675,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:40:01.116788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Swap two token groups at decode time (for example, feed group 2's tokens as the condition for the first denoising step) and measure FID and layout similarity against the original schedule; if quality and coarse structure do not collapse, the claimed coarse-to-fine ordering is not actually determining the output, and the central equivalence claim fails.","supporting_citations":[{"cited_title":"An image is worth 32 tokens for reconstruction and generation.Advances in Neural Information Processing Systems, 37:128940–128966, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the 1D query-token encoder design that D-AR adapts for its sequential diffusion tokenizer."},{"cited_title":"Scalable diffusion models with transformers","cited_arxiv_id":null,"evidence_quote":"Provides the diffusion transformer architecture used in D-AR's decoder and serves as a strong diffusion baseline in comparisons."},{"cited_title":"Autoregressive image generation without vector quantization","cited_arxiv_id":null,"evidence_quote":"Represents the tailored autoregressive approach that injects diffusion into each AR step; D-AR distinguishes itself by staying fully discrete and vanilla."}],"review_version":1}