{"id":"982da6be-0fe2-4edd-9159-863352dda470","arxiv_id":"2509.01984","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"VARIN uses a Location-aware Argmax Inversion pseudo-inverse of Gumbel-max sampling to extract editable discrete noises, enabling training-free prompt-guided editing for visual autoregressive models.","lead":"A new training-free method, VARIN, lets next-scale visual autoregressive models like HART edit images by extracting and reusing the discrete 'inverse noise' that reproduces the source image. It claims to achieve prompt-guided edits in about one second per image while preserving background and structure, comparable to slower diffusion-based editing methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 compares VARIN (s=6) to Regeneration at s=8, not the stated default s=6; at matched s=6, Regeneration has higher CLIP Edited (23.11 vs 21.49), so the claim of superior prompt alignment is unsupported.","rationale":"The core LAI construction is mathematically sound: by truncating non-label Gumbel samples below the label's sampled value minus τ, the argmax condition guarantees exact token-map reconstruction from the source prefix, so the inversion part of the claim is credible. The load-bearing issue is not the inversion itself but the evidence for the editing benefit. Table 1 is the primary quantitative support for the headline claim that VARIN aligns with the target prompt while preserving background better than Regeneration. But the Regeneration baseline in that table is run at s=8, not at the paper's stated default s=6. This is an internal inconsistency: Appendix Table 3 shows the exact Table 1 numbers for Regeneration at s=8, while the text and Appendix C.1 recommend s=6. At s=6, Regeneration has substantially worse structure/background than VARIN (as expected), but its CLIP Edited score is higher (23.11 vs 21.49), so the claim that VARIN 'adher[es] to the target prompt more effectively' fails under the paper's own default. A matched-scale comparison may still show VARIN's background-preservation advantage, but the prompt-alignment claim would need to be revised. This does not invalidate the method's potential, but it does mean the current Table 1 cannot be taken as clean evidence for the central claim. The reader flagged the same inconsistency in the rationale, though their central weakest_assumption was about the interpolation mechanism; hence partial agreement. The verdict remains CONDITIONAL pending a corrected comparison at matched start scales.","tokens_in":15582,"tokens_out":11769,"duration_ms":126867,"concrete_test":"Re-run the Regeneration baseline on PIE-Bench at s=6 (the paper's stated default) and s=8, keeping HART and all other settings fixed, and compare against VARIN at s=6, τ=18. If Regeneration s=6 yields CLIP Edited > 21.49 and/or the reported Table 1 cell changes, the claim of superior prompt alignment is falsified; if the structure/PSNR advantage persists, the mechanism may still stand but the comparison must be reported at matched scales.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 states that for both Regeneration and VARIN the default start scale is s=6 and τ=18. However, the Regeneration row in Table 1 (Structure 25.56, PSNR 20.45, LPIPS 106.50, CLIP Edited 21.13) exactly matches the s=8 row in Appendix Table 3, not the s=6 row (Structure 42.85, PSNR 19.21, CLIP Edited 23.11). Appendix C.1 even calls s=6 the recommended scale. Thus the headline comparison pits VARIN at s=6 against Regeneration at s=8. At the stated default s=6, Regeneration achieves CLIP Edited 23.11 vs VARIN's 21.49, contradicting the paper's claim that VARIN 'adher[es] to the target prompt more effectively' than Regeneration. The central quantitative evidence for simultaneous alignment and preservation therefore rests on a mismatched baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VARIN, a training-free editing method for next-scale visual autoregressive models such as HART and Switti. The core idea is to invert the Gumbel-max sampling procedure used by discrete VAR models, via a new Location-aware Argmax Inversion (LAI) that produces inverse noises able to reconstruct the source token maps. Editing is performed by regenerating scales from a start scale s while interpolating fresh Gumbel noise with the inverse noise, with a hyperparameter τ controlling how much source information is retained. The method is evaluated on PIE-Bench against Regeneration, DICE, EditAR, and several diffusion-based editing baselines, reporting better background preservation and competitive prompt alignment, with a user study and qualitative examples. The reconstruction guarantee of LAI is mathematically sound by construction, and the method is simple and fast, but the quantitative evaluation contains an internal inconsistency in the Regeneration baseline and some claims are not supported by the reported numbers.","tokens_in":15728,"tokens_out":6880,"duration_ms":74581,"significance":"If the claims hold, this is a useful contribution: it is, to my knowledge, the first training-free noise-inversion editing method designed for next-scale autoregressive visual models, and it demonstrates that a simple pseudo-inverse construction can yield reasonable edits with real-time speed. The reconstruction property of LAI is correct by construction, and the paper shows qualitative generalization to a second VAR architecture (Switti), which is a strength. The main weakness is that the headline quantitative claims are weakened by a mismatched baseline and by hyperparameters selected on the same evaluation set used for the final table. The underlying method is plausible, and the reconstruction guarantee is rigorous, but the evaluation needs to be redone or substantially qualified before the claims can be accepted.","major_comments":[{"comment":"The Regeneration row in Table 1 (Structure 25.56, PSNR 20.45, LPIPS 106.50, CLIP Edited 21.13) is exactly the s=8 row of Table 3, not the stated default s=6. At s=6, Regeneration has CLIP Edited 23.11 and Whole 26.83, while VARIN reports 21.49 and 25.05. Therefore the text's claim that VARIN 'adher[es] to the target prompt more effectively' than Regeneration is not supported at the stated default setting. The background-preservation comparison at matched s=6 still favors VARIN (PSNR 26.54 vs 19.21), but the headline claim of simultaneous alignment and preservation must be re-evaluated with matched baselines.","section":"Section 5.2, Table 1, Appendix Table 3"},{"comment":"The same claim of 'adhering to the target prompt more effectively' is also contradicted by the DDPM-Inversion row: DDPM-Inversion has CLIP Edited 23.02 and CLIP Whole 26.22, both higher than VARIN's 21.49 and 25.05. The sentence in Section 5.2 asserting that VARIN 'adher[es] to the target prompt more effectively' than DDPM-Inversion is not supported by the table. Please qualify the claim or present the comparison on the quantities that are actually favorable.","section":"Section 5.2, Table 1"},{"comment":"The editing hyperparameters s=6 and τ=18 are selected based on ablations on PIE-Bench, and the final quantitative results are reported on the same PIE-Bench set, with no validation split. In particular, Appendix C.2 explicitly chooses τ based on 'best visual editing results' in Table 4, and Section 5.2 sets s=6 based on qualitative observation on the same benchmark. This test-set tuning risks optimistic bias in the reported margins. I ask the authors to provide a held-out validation protocol or to discuss the sensitivity of the claimed improvements to the chosen hyperparameters.","section":"Section 5.2, Appendix C.2, Table 4"}],"minor_comments":[{"comment":"The pseudocode for VARIN inversion does not list csrc as input and does not define p_t. In the main text and Figure 3 it is clear that p_t is computed from the source prompt, but the algorithm should explicitly include 'Input: csrc' and a line such as 'pt ← pθ(·|r<t, csrc)' for reproducibility.","section":"Algorithm 2"},{"comment":"Algorithm 6 uses 'Itgt' in line 1 where the source image must be intended (the input list has Isrc), and the label 'Regeneration:' above the algorithm is misleading. Please correct.","section":"Algorithm 6"},{"comment":"Typos: 'staring scale' in Algorithm 3, 'editings = 6' in Section 5.2, 'unormalized' and 'algins' in Section 3.2/4.2. These should be fixed.","section":"Various"},{"comment":"The user study reports percentages from 25 people on 10 images without confidence intervals or significance testing. Please add uncertainty estimates or at least state the small-sample limitation.","section":"Table 6"},{"comment":"The phrase 'Parallel Autoregressive Inversion' is confusing because the loop in Algorithm 2 is sequential over scales. Please clarify what 'parallel' refers to.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core reconstruction construction is sound and the method is a plausible contribution, but the mismatched Regeneration baseline in Table 1 is a serious evaluation issue that must be corrected before publication. The authors should rerun Regeneration at s=6, report matched comparisons, and either add a validation split for hyperparameter selection or explicitly discuss the risk of test-set tuning. If the evaluation is corrected and the claims are appropriately qualified, the paper could be publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: VARIN is worth reading for the Location-aware Argmax Inversion (LAI) construction, which is genuinely new and mathematically sound. But the quantitative headline is compromised: the Regeneration row in Table 1 is run at s=8 while VARIN is at s=6, even though the text sets s=6 as default. At matched s=6, Regeneration actually has higher CLIP Edited (23.11 vs 21.49). So the claim that VARIN adheres to the target prompt more effectively doesn't hold for the presented comparison.\n\nThe core idea is solid. LAI truncates non-label Gumbel samples so argmax(q)=r by construction, giving exact token-level reconstruction, and the location-aware bias toward the predicted logits makes the inverse noise closer to standard Gumbel than one-hot inversion. That's a real contribution, and applying noise inversion to next-scale AR models (HART, Switti) is new. The editing pipeline—interpolating inverse noise with fresh Gumbel noise via a decreasing lambda—is heuristic but plausible, and the qualitative results look reasonable.\n\nSoft spots besides the baseline mismatch:\n\n- tau and start scale s are selected via ablations on the same PIE-Bench set used for final numbers, with no error bars. This inflates confidence in the reported numbers.\n- No quantitative comparison against the one-hot argmax inversion (OAI) they criticize, even though it's the natural baseline for the pseudo-inverse claim.\n- No ablation of the lambda schedule; they just assert linear decay.\n- \"First editing technique for VAR\" overstates prior art—EditAR and Regeneration are editing techniques, even if not inversion-based. The precise claim should be \"first noise inversion-based editing for VAR.\"\n- The failure cases they show are honest, and the limitation section narrows scope to moderate edits, which is good.\n\nNone of these invalidate the mechanism. The reconstruction guarantee is exact by construction, and a corrected comparison at matched settings would likely still show VARIN trading some prompt alignment for much better background preservation (e.g., PSNR 26.54 vs 19.21 at s=6). The problem is the paper doesn't present it that way; it presents a favorable mismatch.\n\nThis paper deserves a serious referee. The math is checkable, the method is training-free and fast, and the idea will be useful to people working on AR-based editing. I'd send it to review with a request to fix the baseline comparison, add error bars, and include the OAI baseline and lambda ablation. I would bring it to a reading group, mostly to talk about the evaluation mismatch as a cautionary tale.","headline":"LAI is a genuinely new pseudo-inverse for argmax with an exact reconstruction guarantee, but the headline comparison uses a mismatched Regeneration baseline; the central idea survives, the evidence as presented does not.","tokens_in":16343,"tokens_out":4121,"would_cite":true,"duration_ms":41323,"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":"VARIN proposes a pseudo-inverse of the Gumbel argmax that recovers source image noise, enabling training-free text-based editing for visual autoregressive models while preserving background details.","keywords":["visual autoregressive models","text-based image editing","noise inversion","Gumbel-max trick","argmax pseudo-inverse","training-free editing","next-scale prediction","HART"],"falsifier":"Take a held-out set of source images and target prompts with large structural changes (e.g., pose or object interaction), run VARIN with the paper's fixed settings (s=6, tau=18), and measure whether CLIP edit similarity and background PSNR remain near PIE-Bench levels; the paper's own failure cases show such prompts degrade, so a systematic curve would bound the claim.","tokens_in":15390,"feed_emoji":"🖼️","tokens_out":4950,"duration_ms":50146,"temperature":0.7,"pith_summary":"The paper introduces VARIN, a training-free way to edit images with visual autoregressive (VAR) models by inverting the discrete sampling step. Because VAR models generate tokens by argmax over Gumbel-perturbed logits, the argmax cannot be inverted uniquely; the paper proposes Location-aware Argmax Inversion (LAI), a pseudo-inverse that recovers Gumbel noises which reproduce the source token maps exactly while remaining close to standard Gumbel noise. During editing, VARIN starts from the source token maps through scale 6, then mixes recovered inverse noises with fresh Gumbel noise under the target prompt, with the inverse-noise weight decaying to zero by the final scale. On PIE-Bench with HART, this beats the regeneration baseline on structure distance (11.46 vs 25.56) and background PSNR (26.54 vs 20.45) while matching prompt alignment, and runs in about one second per image. The paper argues this makes discrete autoregressive image models practical for prompt-guided editing without fine-tuning or optimization.","feed_headline":"Inverse Gumbel noise gives autoregressive image models an edit mode","feed_subtitle":"A pseudo-inverse of argmax recovers the source image's sampling noise, so edits keep the background intact.","key_machinery":"Location-aware Argmax Inversion (LAI): a pseudo-inverse of the argmax sampling step. For each token it draws the winning value from Gumbel(p_label, 1) and every losing value from a Gumbel truncated at winning_value - tau. The truncation enforces the source label as argmax, the shared location keeps the noise Gumbel-like, and tau is the dial that trades edit freedom against background preservation.","core_discovery":"VARIN's central claim is that a discrete next-scale autoregressive model can be edited by noise inversion if one replaces the non-invertible argmax of the Gumbel-max trick with a location-aware pseudo-inverse. LAI samples the perturbed logit at the ground-truth token from a Gumbel centered at the model's predicted logit, and samples all other positions from a Gumbel truncated so they cannot exceed that value minus tau. This guarantees argmax equals the source token, so the extracted noises reconstruct the source image token maps perfectly; because the sampling locations come from the model's own logits, the noises behave like standard Gumbel noise, and tau controls how much source bias they","pith_inferences":["If the mechanism transfers as claimed, the same LAI pseudo-inverse should also work for next-token (not just next-scale) discrete autoregressive image models, since the Gumbel-max inversion is scale-agnostic.","The linear lambda schedule and the fixed start scale 6 are chosen empirically from PIE-Bench; prompts or images with very different structure (e.g., large pose changes) may need per-input tau or a later start scale to avoid failures.","A testable extension is to make tau image-adaptive (e.g., set by estimated edit-region size) rather than a global constant, which could improve preservation on sparse edits.","The method's comparison to optimization-based diffusion editing suggests that discrete inversion alone can reach similar fidelity, but head-to-head user studies on harder structural edits would clarify where the ceiling is."],"forward_implications":["VARIN gives any next-scale VAR model (demonstrated with HART and Switti) a training-free text-editing capability, so editing costs no extra optimization or fine-tuning.","Because inverse noises reconstruct source token maps exactly, edits can be localized to scales 6 and above while early-scale layout tokens remain untouched.","The tau parameter directly controls the preservation-editability trade-off, with tau around 14-20 giving the best balance on PIE-Bench.","Editing runs in roughly one second per image, about 10x faster than DDPM-Inversion, making real-time interactive editing plausible.","The approach is orthogonal to attention-control and test-time optimization methods and can be combined with them for stronger edits."],"supporting_citations":[{"why":"Defines the next-scale visual autoregressive model family that VARIN edits.","marker":"[50]"},{"why":"Provides the HART base model used for the main editing experiments and ablations.","marker":"[48]"},{"why":"SDEdit motivates the Regeneration baseline by starting generation from a middle noise scale.","marker":"[31]"},{"why":"DICE is the discrete-diffusion inversion baseline that VARIN is compared against.","marker":"[17]"},{"why":"EditAR is the autoregressive editing baseline that VARIN outperforms on structure and background metrics.","marker":"[34]"},{"why":"Null-text Inversion is the optimization-based diffusion method whose editing quality VARIN matches without optimization.","marker":"[33]"},{"why":"Supplies the PIE-Bench dataset and evaluation protocol used for all editing experiments and ablations.","marker":"[23]"},{"why":"Provides the Gumbel truncation sampling routine used inside LAI to enforce the argmax condition.","marker":"[25]"},{"why":"Switti is the second base model demonstrating VARIN's transfer across next-scale autoregressive architectures.","marker":"[56]"},{"why":"DDPM-Inversion is the training-free diffusion inversion baseline used for speed and quality comparison.","marker":"[22]"}],"fun_headline_variants":["Noise inversion enables training-free edits for autoregressive models","Inverse Gumbel noises make autoregressive models editable","Discrete noise inversion edits autoregressive images without retraining","Location-aware argmax inversion enables text-guided image editing","Pseudo-inverse of argmax lets autoregressive models edit images"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that interpolating the reconstructed inverse Gumbel noises with fresh Gumbel noise, under a fixed start scale and decay schedule, produces semantically meaningful target-aligned edits, and that this behavior holds across VAR architectures and prompts; the paper validates this only empirically on PIE-Bench.","fun_headline_variants_meta":{"raw":{"variants":["Noise inversion enables training-free edits for autoregressive models","Inverse Gumbel noises make autoregressive models editable","Discrete noise inversion edits autoregressive images without retraining","Location-aware argmax inversion enables text-guided image editing","Pseudo-inverse of argmax lets autoregressive models edit images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001483,"raw_usage":{"total_tokens":5768,"prompt_tokens":695,"completion_tokens":5073,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":439,"completion_tokens_details":{"reasoning_tokens":4991}},"tokens_in":439,"tokens_out":5073,"duration_ms":37595,"temperature":1.0,"reasoning_tokens":4991,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:00:37.388263+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of source images and target prompts with large structural changes (e.g., pose or object interaction), run VARIN with the paper's fixed settings (s=6, tau=18), and measure whether CLIP edit similarity and background PSNR remain near PIE-Bench levels; the paper's own failure cases show such prompts degrade, so a systematic curve would bound the claim.","supporting_citations":[{"cited_title":"EditAR: Unified Conditional Generation with Autoregressive Models","cited_arxiv_id":"2501.04699","evidence_quote":"EditAR is the autoregressive editing baseline that VARIN outperforms on structure and background metrics."},{"cited_title":"Mokady, A","cited_arxiv_id":null,"evidence_quote":"Null-text Inversion is the optimization-based diffusion method whose editing quality VARIN matches without optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Gumbel truncation sampling routine used inside LAI to enforce the argmax condition."},{"cited_title":"V oronov, D","cited_arxiv_id":null,"evidence_quote":"Switti is the second base model demonstrating VARIN's transfer across next-scale autoregressive architectures."},{"cited_title":"Huberman-Spiegelglas, V","cited_arxiv_id":null,"evidence_quote":"DDPM-Inversion is the training-free diffusion inversion baseline used for speed and quality comparison."}],"review_version":1}