{"id":"381ee7c9-fde3-4bd1-aa9b-3c0520bf3287","arxiv_id":"2506.19839","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DFM applies flow matching with per-stage noise schedules to Laplacian pyramid levels in one shared DiT, improving FID and FDD on ImageNet and video benchmarks over prior progressive generation baselines.","lead":"This paper introduces Decomposable Flow Matching (DFM), which trains a single generative model to build images and videos coarse-to-fine by applying flow matching separately at each level of a Laplacian pyramid. The authors report consistent quality improvements over standard flow matching, cascaded models, and Pyramidal Flow on ImageNet and Kinetics-700, plus faster fine-tuning of FLUX.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline FDD gain may reflect DINOv2's structural bias rather than overall visual quality, since DFM is tuned to favor low-frequency structure and the paper admits high-frequency artifacts (Sec. 5, App. G).","rationale":"In good faith, the method is coherent: DFM is a clean extension of Flow Matching with per-scale timesteps, a shared transformer is fed aligned summed patch embeddings, and the compute-matching story is plausible because the token count matches the base DiT. The reader's weakest assumption (dependence on a scale-equivariant autoencoder) is mitigated by Appendix D, Table 4: on the original FLUX autoencoder DFM beats Flow Matching by a larger margin (FID 34.00 vs 54.50) than on the scale-equivariant autoencoder (32.89 vs 43.16), so the coarse-to-fine advantage does not require the spectral AE. The load-bearing soft spot is instead the headline metric itself. DINOv2 features emphasize global structure; DFM's tuned hyperparameters (p0_t=0.9, lognorm location 1.5, 30/10 stage steps) explicitly favor structure, and the paper's own Limitations section and Figure 7 document flattened high-frequency detail. The 35.2% FDD improvement is therefore not yet established as a general visual-quality gain. FID and IS do improve, which prevents this from being a fatal objection, but the magnitude of the headline claim is metric-dependent. This reinforces the reader's CONDITIONAL verdict rather than changing it; the proposed human and texture-sensitive evaluation would settle whether the concern lands.","tokens_in":21041,"tokens_out":20004,"duration_ms":211578,"concrete_test":"Conduct a two-alternative forced-choice human preference study on ImageNet-1K 512px samples generated at the reported optimal settings (cfg=1.5), with at least 1,000 pairwise comparisons per method pair and at least five independent raters, comparing DFM against Flow Matching and Pyramidal Flow; in parallel, compute Frechet distance on VGG-16 early-layer activations (relu1_2/relu2_2) as a texture-sensitive distribution metric. If DFM is not preferred by humans or does not win the texture-sensitive metric, the headline FDD gain overstates the visual-quality improvement; if it wins on both, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Abstract; Table 2) is a 35.2% FDD improvement on ImageNet-1K 512px. FDD is a Frechet distance on DINOv2 features, which are strongly biased toward global/structural content and relatively insensitive to high-frequency texture. DFM's selected training and sampling hyperparameters (p0_t=0.9, lognorm location 1.5, 30/10 per-stage steps) deliberately spend most capacity on the coarse stage, and the paper's own Limitations (Sec. 5) and Appendix G report a 'decreased presence of high-frequency details' plus artifacts in vegetation, fur, and thin structures (Fig. 7). The headline FDD gain may therefore partly measure a shift toward the structural prior captured by DINOv2, not an across-the-board visual-quality improvement. FID and IS also improve, but by smaller margins (e.g., FID 4.28 vs 4.57 over Pyramidal Flow at cfg=1.5), and no human evaluation is reported. If scored with a texture-sensitive metric or by human raters, the DFM advantage could shrink or reverse, weakening the 'improves visual quality' claim as stated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Decomposable Flow Matching (DFM) extends Flow Matching to progressive multi-stage generation. The input latent is decomposed with a Laplacian pyramid into S scales; each scale receives its own flow timestep, and a shared DiT with per-scale patchification, per-scale time embeddings, and masked losses is trained to predict per-scale velocities. At inference, scales are denoised sequentially from coarse to fine according to a threshold schedule with per-stage step counts. Experiments on ImageNet-1K 512/1024 and Kinetics-700 compare DFM with Flow Matching, a cascaded model, and Pyramidal Flow under matched training compute; the paper reports consistent gains, e.g., FDD 200.6 vs 282.9 for Flow Matching on ImageNet 512 at cfg=1.0, and a 35.2% relative FDD improvement over the base architecture is quoted in the abstract. An additional experiment fine-tunes FLUX-dev on an internal dataset and reports faster convergence than standard fine-tuning. The appendix contains extensive ablations over training and sampling hyperparameters, decomposition choices, failed experiments, and limitations.","tokens_in":21236,"tokens_out":11330,"duration_ms":114554,"significance":"The framework is simple, decomposition-agnostic, single-model, and clearly described; these are genuine strengths. The ablation study is unusually thorough, the baseline comparisons are matched in training compute, and the convergence curves show the gain appears throughout training rather than only at the final checkpoint. The candor about failed experiments and limitations is helpful. However, the paper's quantitative claims would be more convincing with repeated seeds, a held-out validation protocol for hyperparameters, and a metric or human study that is sensitive to the high-frequency artifacts the authors acknowledge; the method's main reported gain is on FDD, a DINOv2-based structural metric. The lack of code and the private FLUX dataset also limit independent verification.","major_comments":[{"comment":"The main ImageNet-1K 512px configuration is tuned on the same benchmark used for the headline numbers: p0_t=0.9 and logit-normal location 1.5 are selected from Table 1(a)-(b), and threshold 0.7 with 30/10 per-stage steps is selected from Figure 3 and §4.3. With no validation split and no repeated-seed variance, part of the 35.2% FDD margin in Table 2 may be a selection effect rather than an intrinsic property of DFM. Please evaluate the ablated configurations on a held-out split, or report the sensitivity of the Table 2 comparisons to the selected hyperparameters.","section":"§4.2, §4.3, Table 1, Figure 3"},{"comment":"All main results are single runs without error bars. Several margins are small enough to be within typical run-to-run variability, notably Kinetics-700 at cfg=1.0 (FDD 336.7 vs 353.8 for Pyramidal Flow, FVD 260.2 vs 265.6) and ImageNet 512 at cfg=1.5 (FID 4.28 vs 4.57 for Pyramidal Flow). The claim of uniform superiority over all baselines on all metrics therefore needs at least a few seeds for the main comparisons, or bootstrap confidence intervals on the evaluation metrics, to establish that the differences are not noise.","section":"§4.4, Table 2"},{"comment":"The abstract's headline '35.2% improvement in FDD' is a Frechet distance on DINOv2 features, which are biased toward global structure, while the paper itself reports 'a decreased presence of high-frequency details' and artifacts in vegetation, fur, and thin structures (§5, App. G, Fig. 7). The FDD gain may therefore partly reflect a shift toward the structural prior encoded by DINOv2 rather than an across-the-board improvement in visual quality. Because FID and IS also improve, I do not see this as invalidating the method, but the 'improves visual quality' claim should be qualified, or supported by a human evaluation or a texture-sensitive metric.","section":"§5, Appendix G, Table 2"}],"minor_comments":[{"comment":"The mask term M_s in Eq. (4) is central to the training loss but is never defined in the main text; please specify which stages are masked and how M_s relates to the architectural input masking described in §3.4.","section":"Eq. (4)"},{"comment":"The sentence 'with the exception of Kinetics-700 FDD' is imprecise: at cfg=1.0 and 1.25 DFM has the best FDD on Kinetics-700; only at cfg=1.5 is DFM slightly behind Cascaded (236.5 vs 235.0).","section":"§4.4, Table 2"},{"comment":"The main text says that removing standardization improves performance and that the main experiments exclude standardization, but Appendix A.1 says the model normalizes data across scales with scale-wise pre/post-conditioning; please clarify how the 'standardization' ablated in Table 1(c) differs from the preconditioning described in Appendix A.1.","section":"Appendix A.1 / Table 1(c)"},{"comment":"FDD is cited as 'Frechet Distance computed against DINOv2 features' but no details are given on the DINOv2 variant, layer, pooling, or number of samples used for the covariance estimate; please add these details.","section":"§4.1, evaluation metrics"},{"comment":"The paper does not release code, and the FLUX experiment in §4.5 is on a private internal dataset; this limits independent verification of the empirical claims, especially given the absence of error bars.","section":"Overall"},{"comment":"There are a few typographical errors, including 'add-hoc sampler' in the abstract and 'Addtionally' in Appendix C; please proofread.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid candidate if the empirical concerns are addressed. I am not recommending rejection: the method is sensible, the ablations are extensive, and the improvements appear on multiple metrics. The main risks are selection of hyperparameters on the evaluation benchmark, lack of variance estimates, and overreliance on FDD as the headline metric. I would like the authors to address these with additional experiments or careful qualification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline FDD number is the least convincing evidence in the paper; the method is more credible than that one metric suggests.\n\nWhat's actually new: applying flow matching independently to each level of a Laplacian pyramid with per-stage timesteps and input masking in a single shared DiT. That's a simple, architecture-agnostic recipe, and the paper documents the design space carefully. The ablations on p0_t, logit normal shifts, threshold, per-stage steps, and parameter specialization are useful; the honest reporting of failed experiments (DCT-space DFM, Tokenformer) is a plus. The matched-compute baselines (cascaded, Pyramidal Flow) are fair and consistently beaten on FID, FDD, and IS across ImageNet-512, ImageNet-1024, and Kinetics-700.\n\nThe soft spots are real but proportional. The 35.2% FDD improvement over the base architecture is on DINOv2 features, which are biased toward global/structural content. The paper itself admits decreased high-frequency detail and artifacts in vegetation, fur, and thin structures (Sec. 5, App. G). The FID and IS gains over Pyramidal Flow are much smaller — e.g., FID 4.28 vs 4.57 and IS 196.8 vs 191.0 at cfg=1.5 on ImageNet-512 — so the headline overstates the across-the-board quality gain. No human evaluation is reported. That said, the improvements are consistent across three metrics and three benchmarks, so the method is not a one-metric artifact.\n\nThe bigger issues are methodological hygiene: hyperparameters are selected on the same ImageNet-512 benchmark used for the headline numbers, there are no error bars or multiple seeds, the FLUX finetuning study uses a private internal dataset, and no code is released. These are common in this area, but they push the verdict to \"conditional\" rather than \"settled.\" The method is described precisely enough to reproduce from the paper, and the ablation tables give a fair picture of what matters.\n\nWho should read it: anyone working on progressive or coarse-to-fine generation, and anyone relying on FDD-style metrics. The paper makes a good case that explicit spectral decomposition helps flow matching, and it gives a clear warning about the sweet spot between structure and texture.\n\nRecommendation: send it to peer review. Ask the authors for error bars, a texture-sensitive or human evaluation, and code. The central idea is simple and likely useful; it deserves referee time despite the evaluation caveats.","headline":"DFM is a simple, well-ablated progressive generation method whose FDD headline overstates the gain; the FID/IS evidence is thinner but consistent, and the paper deserves a serious referee.","tokens_in":21804,"tokens_out":3168,"would_cite":true,"duration_ms":32037,"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":"This paper establishes that progressive coarse-to-fine generation can be implemented inside a single Flow Matching model by giving each level of a multiscale representation its own flow timestep, and that this improves image and video…","keywords":["decomposable flow matching","progressive generation","flow matching","laplacian pyramid","multi-scale generation","spectral autoregression","diffusion transformer","image and video generation"],"falsifier":"Run the ImageNet-1K 512px comparison at fixed training compute with an autoencoder whose latent frequency bands are permuted or otherwise misaligned with RGB frequencies; if DFM's FDD advantage over vanilla Flow Matching collapses or reverses, the mechanism carrying the result is the latent spectral structure rather than the per-stage timestep schedule itself.","tokens_in":20808,"feed_emoji":"🖼️","tokens_out":11064,"duration_ms":98582,"temperature":0.7,"pith_summary":"Generating high-dimensional images and videos is expensive, and the paper's claim is that the task becomes both cheaper and higher-quality if the signal is generated coarse-to-fine inside an ordinary Flow Matching pipeline. Decomposable Flow Matching (DFM) decomposes the input into a multiscale representation, typically a Laplacian pyramid, assigns each level its own flow timestep $t_s$, and trains one shared model to predict per-scale velocities. At inference the model generates scales sequentially, starting with coarse structure and adding fine detail only once earlier scales reach a threshold, using a standard sampler. On ImageNet-1K at 512px the paper reports 35.2% FDD improvement over base Flow Matching and 26.4% over the best prior multi-stage baseline under equal training compute, with similar gains at 1024px, on Kinetics-700 video, and in finetuning a large text-to-image model. The wider point is that progressive generation can live in the scheduling of per-stage timesteps rather than in a new generative model family.","feed_headline":"One model, coarse-to-fine, lifts image quality 35 percent","feed_subtitle":"Giving each image scale its own flow timestep lets one network beat multi-stage baselines at equal compute.","key_machinery":"The load-bearing object is the per-stage flow timestep vector $(t_1,\\dots,t_S)$, which breaks the usual single global timestep into one per level of a multiscale decomposition and thereby turns a Flow Matching model into a coarse-to-fine progressive generator. The decomposition itself is a Laplacian pyramid, in which each level holds the difference between successive downsampled approximations, so levels contain progressively finer detail. Around these two objects the method adds per-scale patchification and timestep-embedding layers inside a DiT-style transformer so one network handles all scales, a masking term that suppresses loss from future fully-noisy scales, and a scale-equivariant autoencoder that maps RGB frequency bands to matching latent frequency bands. These pieces together replace stage-transition machinery with a simple schedule of per-scale timesteps.","core_discovery":"The central claim is that Flow Matching and progressive generation can be fused without special diffusion processes, separate expert models, or ad-hoc samplers. The input is decomposed into $S$ scales $X^s$; each scale receives an independent flow timestep $t_s$, so the forward process is $X^s_{t_s} = t_s X^s_1 + (1-t_s) X^s_0$ per scale, and the model predicts per-scale velocities $v^s$ from the noisy versions of all scales. Training simulates the progressive schedule by sampling a current stage, drawing its timestep from a logit-normal distribution, drawing less-noisy timesteps for previous stages, setting future stages to full noise, and masking those future scales out of the loss. Inference advances the stages one by one with a standard ODE sampler, moving to the next scale once the current one reaches a threshold $\\tau$. On ImageNet-1K, Kinetics-700, and a large-model finetuning task, the paper reports that this single-model scheme outperforms vanilla Flow Matching, cascaded diffusion, and Pyramidal Flow at matched training compute.","pith_inferences":["If the per-stage timestep schedule is the active ingredient, the same factorization could be applied to non-diffusion generative models, e.g., next-scale autoregressive models, to test whether progressive structure itself or the flow formulation drives the gain.","The hand-set switch threshold and per-stage step counts suggest a natural extension: learn the sampling schedule along with the model, which might reach the same quality with fewer total steps.","For video, the same decomposition idea can be applied along time as well as space, which would test whether coarse-to-fine generation improves temporal consistency the way it appears to improve spatial structure.","A decisive comparison would use a spectrally disentangled tokenizer (e.g., wavelet-based) instead of a scale-equivariant fine-tuned autoencoder, isolating how much of the benefit comes from the latent structure versus the progressive schedule."],"forward_implications":["On ImageNet-1K 512px, DFM reports a 35.2% FDD improvement over the base Flow Matching model and 26.4% over the best-performing prior baseline under equal training compute.","At 1024px and on Kinetics-700 video, DFM also beats Flow Matching, cascaded models, and Pyramidal Flow on FID, FDD, and related metrics at matched compute.","Finetuning a large text-to-image model with DFM reaches 28.7% lower FID and 24.3% lower FDD than standard full finetuning after the same number of steps, with higher CLIP similarity.","Because each scale is generated with its own timestep, intermediate stages can be decoded during inference, giving low-resolution previews without any extra model.","The decomposition is user-defined, so DWT, DCT, Fourier, or multiscale-autoencoder decompositions are drop-in alternatives to the Laplacian pyramid used in the main results."],"supporting_citations":[{"why":"Defines the rectified-flow objective that DFM extends by introducing per-scale timesteps.","marker":"[27]"},{"why":"Provides the Flow Matching framework whose velocity-matching loss is factorized per stage.","marker":"[26]"},{"why":"Supplies the DiT transformer backbone that DFM adapts with per-scale patchification and timestep embedders.","marker":"[33]"},{"why":"Pyramidal Flow Matching is the main progressive-generation baseline DFM is compared against.","marker":"[18]"},{"why":"Cascaded diffusion models represent the multi-model baseline paradigm in the comparison.","marker":"[13]"},{"why":"Scale-equivariant autoencoder finetuning gives the latent space the spectral structure the main DFM results assume.","marker":"[38]"},{"why":"Supplies the logit-normal timestep sampling distribution used for DFM training.","marker":"[7]"},{"why":"FLUX is the large text-to-image model used in the finetuning experiments.","marker":"[23]"},{"why":"ImageNet-1K is the main image benchmark for the reported comparisons.","marker":"[5]"},{"why":"Kinetics-700 is the video benchmark for the reported video experiments.","marker":"[3]"}],"fun_headline_variants":["One flow model, per-scale timesteps: beats multi-stage baselines","Decomposable Flow Matching: 35% better images, no cascade needed","Give each image scale its own timestep, cut complexity, lift quality","Single-network progressive generation: per-scale flow, top results","Coarse-to-fine flow matching: one model, no ad-hoc samplers, 35% gain"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The large quality gains depend on a compression model whose internal frequency bands line up with the visual frequency bands in RGB; without that spectral alignment the paper's own appendix shows DFM still beats Flow Matching, but the margin shrinks.","fun_headline_variants_meta":{"raw":{"variants":["One flow model, per-scale timesteps: beats multi-stage baselines","Decomposable Flow Matching: 35% better images, no cascade needed","Give each image scale its own timestep, cut complexity, lift quality","Single-network progressive generation: per-scale flow, top results","Coarse-to-fine flow matching: one model, no ad-hoc samplers, 35% gain"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1650,"prompt_tokens":1023,"completion_tokens":627,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":523}},"tokens_in":639,"tokens_out":627,"duration_ms":6413,"temperature":1.0,"reasoning_tokens":523,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:24:03.323136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the ImageNet-1K 512px comparison at fixed training compute with an autoencoder whose latent frequency bands are permuted or otherwise misaligned with RGB frequencies; if DFM's FDD advantage over vanilla Flow Matching collapses or reverses, the mechanism carrying the result is the latent spectral structure rather than the per-stage timestep schedule itself.","supporting_citations":[{"cited_title":"Scalable diffusion models with transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the DiT transformer backbone that DFM adapts with per-scale patchification and timestep embedders."},{"cited_title":"Improving the diffusability of autoencoders","cited_arxiv_id":null,"evidence_quote":"Scale-equivariant autoencoder finetuning gives the latent space the spectral structure the main DFM results assume."},{"cited_title":"Flux, 2024","cited_arxiv_id":null,"evidence_quote":"FLUX is the large text-to-image model used in the finetuning experiments."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"ImageNet-1K is the main image benchmark for the reported comparisons."}],"review_version":2}