{"id":"91a413eb-0ff7-4d25-885c-c7e618b8a3f4","arxiv_id":"2411.15867","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"PanoLlama uses token redirection on a fixed-size autoregressive image model to generate coherent, arbitrarily long panoramas without any extra training.","lead":"PanoLlama generates long panoramic images by having a pretrained next-token image model predict one overlapping crop at a time, a training-free trick the authors call token redirection. It reports higher coherence and faster generation than diffusion-based panorama methods, and it introduces a 1,000-prompt benchmark, making it a candidate new baseline for panorama generation research.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Horizontal token redirection asks the AR model to predict off-raster tokens: after a row suffix, the trained next token is the next row, not a rightward extension, so the central horizontal claim is unestablished.","rationale":"The reader's weakest assumption, that the AR model must generalize to suffix-conditioned off-grid continuation, is the right umbrella, and the conditional verdict is reasonable. My stress-test sharpens it: for horizontal expansion the off-grid point is not merely a resolution or position generalization issue; it is a mismatch between the token order the model was trained on and the spatial direction the method claims. The diagnostic above would settle whether the concern actually lands. If it lands, the horizontal SOTA claim and the main figures are invalid; if it does not, the method's main results survive and the remaining evaluation issues from the reader's verdict stand. Because the test has not been run, I keep the verdict at conditional/unchanged rather than moving to reject.","tokens_in":13469,"tokens_out":18066,"duration_ms":182843,"concrete_test":"Tokenize a set of natural wide images with the VQGAN encoder; for each row j of a 16×16 crop, form the context of the last √p−c tokens of row j, and compare fG's predicted next c tokens against (a) the true next c tokens of the same row (horizontal ground truth) and (b) the true first c tokens of row j+1 (vertical ground truth). If the model's predictions match (b) significantly more often than (a), Eq. 8 implements vertical continuation rather than horizontal expansion, and the central horizontal-coherence claim is refuted. If they match (a), the model has learned off-raster horizontal extrapolation and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's main quantitative results are for horizontal expansion (w′=5120, 10x wider). In LlamaGen's raster-scan token order, the token immediately after the last token of row j is the first token of row j+1, not a token to the right of row j. Equation 8 nevertheless takes the last √p−c tokens of row j as context and treats fG's output tokens as c new columns appended to row j. Those output tokens are drawn from the model's learned distribution for the start of the next row, a distribution trained to be vertically, not horizontally, adjacent. The paper supplies no argument or measurement that this off-grid prediction coincides with a horizontally coherent continuation. The same issue does not arise for vertical expansion, where redirection follows the raster order, but it undermines the headline 10x-horizontal experiments and the claimed SOTA horizontal coherence. This is the load-bearing step in the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"PanoLlama re-frames panoramic image generation as next-crop prediction using pre-trained autoregressive image models. The method is training-free: it first generates a fixed-size block with LlamaGen, then redirects generation by feeding the trailing tokens of the previous block as context to generate the next block vertically or horizontally, concatenates all token sequences, and decodes them into an arbitrarily larger panorama. The authors report state-of-the-art coherence, fidelity, and aesthetics on a newly constructed 1,000-prompt benchmark, and demonstrate additional capabilities such as multi-scale expansion, mask-free layout control, and multi-guidance synthesis. The paper is clearly written and the idea is simple and attractive, but the quantitative evidence does not yet support the central horizontal-expansion claim.","tokens_in":13631,"tokens_out":6617,"duration_ms":57474,"significance":"The autoregressive next-token-prediction paradigm for panoramic image generation is a potentially valuable departure from the joint-diffusion pipeline, and the training-free token-redirection idea is elegant if it works. The paper's strengths include a concrete algorithm (Algorithm 1), a code release, a new benchmark, and several applications. However, the central mechanism for horizontal expansion relies on the model predicting off-raster tokens, and the evaluation mixes absolute and relative comparisons across different base generators. If these issues are resolved, the contribution could be significant; in the current form, the reported state-of-the-art claims are not yet established.","major_comments":[{"comment":"In horizontal expansion, the last sqrt(p)-c tokens of row j are used as the context to generate c new tokens that are treated as a horizontal extension of that row. In the raster-scan token order used by LlamaGen and other VQGAN-based autoregressive models, the token that follows the end of a row in training is the first token of the next row, not a token to the right. The model therefore receives an out-of-distribution context, and its output is drawn from a distribution trained to predict vertically adjacent content. The paper supplies no quantitative or qualitative ablation that separates the effect of this off-raster redirection from the base model's own continuity. Since the headline results (w'=5120, 10x horizontal expansion) are based on this mechanism, the central claim is not yet established. Please provide evidence that the predicted tokens are horizontally coherent, for example by measuring agreement with ground-truth right-neighbor tokens, comparing against a raster-order baseline, or using human evaluation.","section":"Sec. 3.2, Eq. (8)"},{"comment":"The coherence scores (LPIPS, DISTS, TV, SSIM) are absolute values computed on outputs from different base generators: PanoLlama uses LlamaGen while all diffusion baselines use SD or SDXL. As the paper itself states, PIG performance is closely tied to the reference model, so a direct comparison of absolute coherence scores can confound the PIG method with the base model. To support the claim of state-of-the-art coherence, the comparison should be controlled: use the same base generator where possible, report the coherence degradation relative to each base model, or at least include coherence scores for the unexpanded base models. Without this, the reported 47.50% TV improvement and the 21.15% LPIPS improvement are not interpretable.","section":"Sec. 4.1, Table 1"},{"comment":"The dataset description is internally inconsistent. The main text says the dataset contains 1,000 prompts and that \"using all prompts ... and 25 sets of random seeds\" yields 2,000 panoramas; Sec. 4.2 says 80 prompts per theme (2,000 prompts for 25 themes); Appendix D says \"400 prompts per theme\" (10,000 prompts). These numbers cannot all be correct, and they prevent the experiments from being replicated. Please report the exact number of prompts, seeds, and generated images, and correct the statements.","section":"Sec. 4.1 / Sec. 4.2 / Appendix D"},{"comment":"The method assumes the pre-trained tokenizer decoder fT_d can decode concatenated token sequences longer than the training grid into an arbitrary-width panorama. The paper does not verify or discuss whether LlamaGen's VQGAN decoder preserves quality outside its training resolution, nor whether the concatenation order in Eqs. (7)-(9) matches the decoder's expected spatial arrangement. This assumption is part of the method and should be tested, for example by decoding a horizontally concatenated sequence of two unmodified generated blocks and measuring seam artifacts.","section":"Sec. 3.2, Eq. (10)"}],"minor_comments":[{"comment":"The loss function and the optimization for theta* are not used by the training-free method; they are only motivational. Please state this explicitly to avoid the appearance that the method optimizes Eq. (2).","section":"Sec. 3.1, Eqs. (1)-(3)"},{"comment":"The word \"construsting\" should be \"constructing\".","section":"Sec. 4.2"},{"comment":"The notation v1,r*sqrt(p) and similar expressions lack clear subscripts; please typeset as v_{1,r*sqrt(p)} and v_{1,p} for readability.","section":"Eq. (6)"},{"comment":"The description of prompt creation (\"400 prompts per theme\") should be aligned with the 1,000-prompt total mentioned in the abstract and Sec. 4.1.","section":"Appendix D"},{"comment":"The axes of the quality-efficiency plot are not labeled; please add labels and units.","section":"Fig. 4"},{"comment":"The abbreviations \"V AR\" (visual AR) and \"VAR\" (visual autoregressive) are used inconsistently; please unify them.","section":"Secs. 1-2"}],"recommendation":"major_revision","confidential_remarks":"The horizontal redirection issue is the most serious concern: the central quantitative claim depends on an off-raster prediction that is not validated. The dataset inconsistency also needs a clear correction. If these are addressed with controlled experiments, the paper could become a useful contribution; in the current form I would not recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a genuinely interesting twist on using fixed-size autoregressive image models for panorama generation, and the qualitative results are impressive. But the main quantitative claims are built on horizontal expansion, and the paper never validates that the model can actually predict tokens in a direction it was never trained on. That's the thing to check first.\n\nWhat's new: token redirection as an inference-time trick. You generate one crop, then feed the trailing tokens (in a specific arrangement) back as context to produce the next crop. This is training-free and works with any fixed-size AR model. The paper also shows nice applications like mask-free layout control and multi-guidance. The benchmark, while self-created, is a useful step toward standardizing PIG evaluation.\n\nWhat it does well: the writing is clear, the method is simple to grasp, and the qualitative comparisons are extensive. The vertical expansion is conceptually sound because it follows the raster-scan order. The idea of framing PIG as next-crop prediction is a natural re-description, but it's fine.\n\nThe soft spots are real. First, and most important: horizontal expansion. In LlamaGen's raster order, the token after the last token of a row is the first token of the next row, not a token to the right. Equation 8 takes a row suffix and treats the model's output as new columns for that same row. That asks the model to produce tokens for positions it never saw during training. The paper gives no argument, measurement, or qualitative close-up that this off-raster continuation is actually horizontally coherent. The headline 512×5120 results are all horizontal, so this cuts the central claim. Vertical expansion is unaffected, but that's not what the SOTA numbers are about.\n\nSecond, the quantitative evaluation is not convincing as a comparison. Baselines use SD/SDXL backbones, PanoLlama uses LlamaGen. Coherence metrics are absolute, so part of the gap may be the base model, not the method. There are no error bars. And the benchmark is internally inconsistent: the main text says 1,000 prompts, Appendix D says 400 prompts per theme (which would be 10,000 for 25 themes), and the experimental section says 2,000 panoramas from 1,000 prompts and 25 seeds. That doesn't add up.\n\nThe paper also doesn't discuss prior autoregressive infinite-synthesis work, which would help position the contribution.\n\nWho should read this: anyone working on extending generative models to arbitrary sizes, and the PIG community specifically. It deserves a serious referee because the idea is good and the problems are addressable. But the referee should ask for a matched-backbone comparison, a public benchmark, and direct validation of the horizontal token redirection assumption.","headline":"Token redirection is a neat training-free idea, but the horizontal expansion relies on an unvalidated off-raster assumption that undermines the headline numbers.","tokens_in":14171,"tokens_out":4001,"would_cite":true,"duration_ms":31400,"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":"By reusing the trailing tokens of each generated crop as the starting condition for the next, PanoLlama turns a fixed-size visual autoregressive model into a training-free generator of coherent, arbitrarily long panoramas.","keywords":["panoramic image generation","next-token prediction","autoregressive image models","training-free extrapolation","token redirection","text-to-image generation","image generation benchmark","multi-scale synthesis"],"falsifier":"A decisive experiment is to tokenize the left half of a real wide photograph, apply PanoLlama to extend it rightward, and measure how closely the synthesized right half matches the real right half with a perceptual metric such as LPIPS. If the match is good only for uniform textures but fails for scenes with long-range geometry—crowds, facades, repeating patterns—then the redirection strategy would be shown to produce local continuity without the multilevel coherence the paper claims.","tokens_in":13223,"feed_emoji":"🌄","tokens_out":8343,"duration_ms":70037,"temperature":0.7,"pith_summary":"PanoLlama argues that panoramic image generation is best modeled as a next-token prediction problem: a panorama is a sequence of crops, and each crop should be generated in the context of everything before it. Rather than train a new model, it takes a pre-trained visual autoregressive image generator (which normally produces one fixed square image) and, after each crop, feeds the last tokens of the current crop back as the prefix for the next crop. This 'token redirection' gives spatial continuity in both horizontal and vertical directions without any fine-tuning. The paper reports that on its own 1,000-prompt benchmark of 512 by 5120 panoramas, the approach beats diffusion-based panorama methods on crop-connection coherence, fidelity, diversity, and aesthetic score, while also enabling mask-free layout control and multi-directional expansion.","feed_headline":"Redirection makes fixed-size image models generate endless panoramas","feed_subtitle":"Tail tokens become the next crop's prefix, creating endless panoramas with no retraining.","key_machinery":"The central object is token redirection, a manipulation of the raster-scan token order produced by the image tokenizer. Because an autoregressive image generator arranges its $p$ image tokens in a fixed grid of $\\sqrt{p}$ by $\\sqrt{p}$, the trailing boundary tokens of one crop can be fed back as the conditional prefix for the next block, so the model's learned next-token distribution supplies the spatial continuation. The two redirection rules are Eq. 6 for vertical expansion (feed the last $p - r\\sqrt{p}$ tokens back, producing $r$ new rows) and Eq. 8 for horizontal expansion (for each row $j$, feed the last $\\sqrt{p} - c$ tokens of that row back, producing $c$ new columns per row). Together with the tokenizer decoder from Eq. 10, these rules turn a fixed-square generator into an endless panorama generator; the overlap parameters $r$ and $c$ set the expansion stride and therefore the trade-off between coherence and generation time.","core_discovery":"The central claim is that next-crop prediction can be obtained for free from a fixed-size autoregressive image model by exploiting the raster-scan order of its discrete tokens. For vertical expansion, the next crop starts from the last $p - r\\sqrt{p}$ tokens of the previous crop (Eq. 6); for horizontal expansion, each row of the next crop starts from the last $\\sqrt{p} - c$ tokens of the corresponding row of the previous crop (Eq. 8). Iterating this process and decoding the concatenated token stream yields a panorama of arbitrary length. The paper's experiments indicate that this simple reuse of the model's own conditional distribution produces smoother crop junctions than heuristic diffusion-based fusion, and the accompanying benchmark of 1,000 prompts across 25 themes is offered as a standardized test bed. The paper states openly that the method approximates global dependencies with a partial token context rather than fully conditioning on the entire panorama.","pith_inferences":["Not tested in the paper, but a natural next stress test is very long extrapolation (well beyond 10x): with each step reusing only a bounded context, any slow drift in style or content would accumulate differently from what a single 10x benchmark shows.","The benchmark's coherence metrics compare adjacent non-overlapping crops; an inference from that design is that local junction quality is measured directly, while global long-range consistency is only indirectly reflected, so future work could add explicit long-range layout agreement tests.","The redirection idea is stated for one raster-scan generator family; whether it transfers to hierarchical or masked next-scale image generators is open, because those models do not expose a single per-row token ordering."],"forward_implications":["Any future fixed-size autoregressive image generator becomes a panoramic generator with no retraining, so the redirection trick is a drop-in capability for the whole class of models.","Because the redirection is row-wise and column-wise, the same loop supports horizontal, vertical, and two-dimensional multi-scale panoramas from a single pipeline.","The overlap parameters $r$ and $c$ give a direct knob for the quality/efficiency tradeoff: smaller strides mean more overlap and higher local coherence but longer generation.","The main results are measured on a new 1,000-prompt benchmark spanning 25 themes, so future panorama methods can be compared on a standardized test bed rather than small hand-picked prompt sets.","The method's token-prefix conditioning also enables mask-free layout control and multi-guidance generation, letting different regions be steered by different prompts or images within one panorama."],"supporting_citations":[{"why":"Supplies the fixed-size visual autoregressive generator that PanoLlama extends; the whole method builds on this pre-trained model's next-token predictions.","marker":"[33]"},{"why":"Provides the discrete image tokenizer whose raster-scan token order the redirection strategy exploits, and decodes the final token sequence into pixels.","marker":"[37]"},{"why":"Defines the joint-diffusion paradigm for panorama generation and is the primary baseline that PanoLlama claims to surpass in coherence.","marker":"[3]"},{"why":"A synchronized joint-diffusion baseline used in the performance comparison, against which the paper reports improved coherence and speed.","marker":"[18]"},{"why":"A merge-attend-diffuse baseline that the paper identifies as the strongest coherence competitor, and whose scores mark the improvements in LPIPS, DISTS, TV, and SSIM.","marker":"[24]"},{"why":"Supplies the diffusion reference model used by the diffusion baselines, so the paper's relative fidelity and compatibility scores are computed against this generator.","marker":"[26]"}],"fun_headline_variants":["Fixed-size AR models get endless panoramas via token redirection","Token redirection makes fixed-size autoregressive models generate endless panoramas","Endless panoramas from fixed-size AR models by reusing tokens as context","Next-crop prediction: training-free redirection for endless panoramas"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that when the token sequence of a generated image is cut short and its tail is fed back as a prompt, the model genuinely continues drawing the adjacent part of the scene rather than merely producing plausible-looking but structurally disconnected pixels.","fun_headline_variants_meta":{"raw":{"variants":["Fixed-size AR models get endless panoramas via token redirection","Token redirection makes fixed-size autoregressive models generate endless panoramas","Endless panoramas from fixed-size AR models by reusing tokens as context","Next-crop prediction: training-free redirection for endless panoramas"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000673,"raw_usage":{"total_tokens":3084,"prompt_tokens":983,"completion_tokens":2101,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":2025}},"tokens_in":599,"tokens_out":2101,"duration_ms":13893,"temperature":1.0,"reasoning_tokens":2025,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:47:58.941631+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive experiment is to tokenize the left half of a real wide photograph, apply PanoLlama to extend it rightward, and measure how closely the synthesized right half matches the real right half with a perceptual metric such as LPIPS. If the match is good only for uniform textures but fails for scenes with long-range geometry—crowds, facades, repeating patterns—then the redirection strategy would be shown to produce local continuity without the multilevel coherence the paper claims.","supporting_citations":[{"cited_title":"Neural discrete representation learning","cited_arxiv_id":null,"evidence_quote":"Provides the discrete image tokenizer whose raster-scan token order the redirection strategy exploits, and decodes the final token sequence into pixels."},{"cited_title":"Multidiffusion: Fusing diffusion paths for controlled image generation","cited_arxiv_id":null,"evidence_quote":"Defines the joint-diffusion paradigm for panorama generation and is the primary baseline that PanoLlama claims to surpass in coherence."},{"cited_title":"Syncdiffusion: Coherent montage via synchronized joint diffusions","cited_arxiv_id":null,"evidence_quote":"A synchronized joint-diffusion baseline used in the performance comparison, against which the paper reports improved coherence and speed."},{"cited_title":"Merging and splitting diffusion paths for se- mantically coherent panoramas","cited_arxiv_id":null,"evidence_quote":"A merge-attend-diffuse baseline that the paper identifies as the strongest coherence competitor, and whose scores mark the improvements in LPIPS, DISTS, TV, and SSIM."}],"review_version":1}