Decomposable Flow Matching

1.1 Deduplication

First read ml-report-index.md. Recent coverage includes MeanFlow(one-step Flow Matching), MolJO/I2Mole(molecular generation/optimization), earlier long-video generation before LongLive, and theory/systems. One-step flows, molecular generation, and memory/privacy diffusion are excluded as core themes. Candidates may include video,3D, discrete flows, and alignment, but the drawn main question must differ. DFM focuses onmultiscale, single-model, progressive flow matching, distinct from MeanFlow’s one-step velocity identity and molecular work.

1.2 Fixed candidate pool(20 papers)

Order fixed before drawing. Entries1–15 prioritize accepted papers;16–20 are newer public ICLR2026 submissions/high-quality arXiv papers, with redraw only if selected material is insufficient.

1. SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer — ICLR 2026 Oral
2. GLASS Flows: Efficient Inference for Reward Alignment of Flow and Diffusion Models — ICLR 2026 Oral
3. Half-order Fine-Tuning for Diffusion Model — ICLR 2026 Oral
4. NoisePrints: Distortion-Free Watermarks for Authorship in Private Diffusion Models — ICLR 2026
5. Discrete Adjoint Matching — ICLR 2026
6. Disentangled Hierarchical VAE for 3D Human-Human Interaction Generation — ICLR 2026
7. NatADiff: Adversarial Boundary Guidance for Natural Adversarial Diffusion — ICLR 2026
8. ImagenWorld: Stress-Testing Image Generation Models — ICLR 2026 / arXiv
9. Adjoint Schrödinger Bridge Sampler — NeurIPS 2025 Oral
10. Edit Flows: Variable Length Discrete Flow Matching with Sequence-Level Edit Operations — NeurIPS 2025
11. Improving Progressive Generation with Decomposable Flow Matching — NeurIPS 2025
12. Self-Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion — NeurIPS 2025 Spotlight
13. Efficient Zero-shot Inpainting with Decoupled Diffusion Guidance — ICLR 2026
14. Subject-driven Video Generation Emerges from Experience Replays — public ICLR2026 submission
15. GenICF: Benchmarking Generative Methods — public ICLR2026 submission
16. E-CommerceVideo: A Benchmark and Approach for E-Commerce Video Generation — public ICLR2026 submission
17. Enhancing Physical Plausibility in Video Generation by Reasoning the Implausibility — arXiv
18. Generative Image Compression by Estimating Gradients of the Rate-variable Feature Distribution — arXiv / public ICLR2026 submission
19. MV-Diffus3R: Refining Multi-View Diffusions for Geometric Coherence 3D Reconstruction — public ICLR2026 submission
20. OneFlow: Concurrent Mixed-Modal and Interleaved Generation with Edit Flows — arXiv / public ICLR2026 submission

1.3 Actual draw

$ shuf -i 1-20 -n 1
11

Paper11 has official NeurIPS proceedings, a35-page full version, project page, appendix, code, and sufficient results/ablations, sono redraw was needed. Selection does not depend on popularity, citations, or affiliation.

3.1 Narrative

First, a Laplacian pyramid decomposes a latent into low-frequency approximation and high-frequency residuals. Second, each component gets an independent Flow Matching time t^s: the current stage receives random noise, completed stages favor low noise, and future stages stay fully noisy with masked losses. Third, one DiT shares its backbone, duplicating only scale-specific patchification, output projections, and time embeddings. Fourth, inference retains ordinary Euler integration, using threshold τ to start the next stage. ImageNet-1K, Kinetics-700, and internal-data FLUX fine-tuning validate training from scratch and adapting large models.

Three assumptions underlie this narrative:

1. Spectral decomposition carries semantics. Low-frequency latents mainly encode structure and high-frequency residuals detail. If the autoencoder mixes frequencies completely, stages lose interpretability.
2. The shared backbone has sufficient capacity. One DiT can fit multiscale velocities without expert models; stage-specific ports handle dimensional differences.
3. The joint training distribution approximates inference trajectories. Stage probabilities and shifted logit-normal sampling cover inference’s cross-scale noise combinations.

3.2 Key tradeoffs

4.1 Standard Flow Matching

For data sample X1 ~ p_d and noise X0 ~ N(0,I), the paper uses a linear probability path:

[ X_t=tX_1+(1-t)X_0,v_t==X_1-X_0. ]

Generator G is trained by velocity regression:

[ L_{FM}=E_{t,X_1,X_0}|G(X_t,t)-v_t|_2^2. ]

Core Eq.(1)(PDF p.4; HTML lines 103–107). Inference integrates an ODE; training need not simulate full trajectories.

4.2 Multiscale representation

Given S resolutions, Laplacian components are:

[ X^1=(X,1), X^s=(X,s)-((X,s-1),s), s>1. ]

Summing through level s produces intermediate preview \bar X^s=\sum_{k=1}^{s}X^k. This mathematically enables viewing generation partway through(Eq.(2), PDF p.4; HTML lines 110–116).

4.3 Decomposable Flow Matching

DFM assigns each scale an independent time:

[ P_{t^1,,tS}={X^s_{ts}}_{s=1}^S, X^s_{ts}=t^sXs_1+(1-t^s)Xs_0. ]

The shared network predicts all scale velocities, while mask M^s excludes inactive stages:

[ L_{DFM}=E{s=1}^{S}Ms |G(P{t^1,,tS},t^1,,tS)-v^s|_2^2. ]

These are Eqs.(3)–(4)(PDF p.4; HTML lines 119–129). The loss itself remains similar; joint time distribution p_t becomes a curriculum. For current stage k, future scales are fully noisy and excluded from loss; past scales use low-noise-shifted logit-normal samples; the current scale uses ordinary logit-normal sampling.

4.4 Sampling pseudocode

Inputs: noise pyramid{X0^s}, shared DiT G, thresholdτ,40 total steps
Training:
  1  X -> Laplacian pyramid{X1^s}
  2  Sample main stage k ~ Categorical(p_stage)
  3  For s<k: t^s ~ LogitNormal(location=1.5, scale=1)
  4  For s=k: t^s ~ LogitNormal(location=0, scale=1)
  5  For s>k: t^s=0, M^s=0; otherwise M^s=1
  6  Predict active-scale velocities with shared G; minimize Eq.(4)

Inference:
  1  Integrate the coarsest stage; activate the next atτ
  2  Active stages continue in parallel on individual linear schedules
  3  Default coarse/fine allocation:30/10 Euler steps
  4  Sum{X^s}; decode images/video with the autoencoder

4.5 Architectural interface

For256→512 stages, low-resolution1×1 patches and high-resolution2×2 patches yield equal token counts. Scale patch embeddings are summed into one Transformer, with separate velocity-output heads. Per-scale time-MLP outputs are also summed into unified conditioning(Section 3.4, PDF p.5; HTML lines 130–135). More precisely than claiming no DiT changes:backbone blocks stay unchanged, but input/output ports and modulation entry points change.

6.1 ImageNet-1K 512/1024

Image experiments report FID, FDD, and IS. FID uses Inception features; FDD uses DINOv2, considered more semantic/structural by the authors; IS combines class confidence and diversity.512px uses 50k samples,1024px10k, so sampling errors cannot be assumed equal. ImageNet carries research-only, noncommercial terms(Appendix A.3, PDF p.15).

6.2 Kinetics-700 512

Video metrics are frame-FID, FDD, and FVD. Frame-FID/FDD assess individual-frame visual quality; FVD incorporates temporal features. All are lower-is-better. Experiments use 10k samples,200k training steps, CogVideo AE, and unit-variance normalization of decomposed stages. Kinetics-700 is used under CC BY4.0(Appendix A.3, PDF p.15).

6.3 Internal FLUX fine-tuning set

FLUX-dev first adapts to the scale-equivariant AE for 24k steps. From that checkpoint, FLUX-FT and FLUX-DFM each train32k steps at batch 192,1024px, multiple aspect ratios, evaluated on 10k internal prompts(Table 3/Appendix E.1, PDF pp.9,18). Unreleased data, prompts, and AE checkpoint limit evidence to relative internal-domain convergence, not public benchmark rankings.

6.4 No fabricated five-benchmark×six-method matrix

The complete equal-compute matrix contains only 3 dataset/resolution settings ×4 methods ×3 CFG values; FLUX has only 3 models. There is no fifth public benchmark or six strictly comparable methods. Importing other papers’ numbers would break backbone/budget/protocol consistency. This report therefore transcribes Table 2 completely rather than inventing5×6 coverage.

7.1 Table 2:3 benchmark groups ×4 methods ×3 CFG values at equal training compute

All values come from core Table 2(PDF p.8; arXiv HTML lines 220–240). Bold marks the best within each CFG/column;↓ lower is better,↑ higher is better.

7.2 Against the strongest non-DFM baseline at the same CFG

* The abstract claims26.4% improvement over the best progressive baseline; Table 2’s116.2→85.8 recalculates to 26.16%. Against the actual best non-DFM column value114.0, it is 24.74%. This does not reverse conclusions but requires explicit comparator choice.

7.3 Table 3: FLUX fine-tuning

Source: core Table 3(PDF p.9; HTML lines 244–253).

Against equal-training-compute FLUX-FT, DFM reduces FID by 3.73 / 29.70%and FDD by 37.4 / 24.35%, while CLIPSIM increases by 0.0120 / 3.68%. Main-text line253 gives FID improvement as 28.7%, but table recalculation gives 29.70%; introduction line73 also says 29.7%, so 28.7% appears to be an internal typo.

8.1 Training-stage probability: structure-first gains are not guaranteed linear

Table 1(a)(PDF p.6) shows on DiT-B that as p0 moves.7→.8→.9→.95, FID is 36.02→32.09→29.56→27.50, FDD676→622→571→540, and IS32.2→36.1→41.0→43.7..7→.95 reduces FID relatively by 23.65%and FDD by 20.12%. Yet main experiments choose .9, reasoning that larger models may have capacity for nonstructural detail and need closer inference-trajectory matching(Sec.4.2, PDF p.6). This isspeculative hyperparameter transferrather than an XL-sweep-supported conclusion.

8.2 Decoupled times are a discrete turning point

Changing previous-stage logit-normal location across 0,.5,1,1.5,2 yields FID 28.89–29.96. Tying all stage times instead jumps FID from 29.56 to 55.49（+87.72%）and FDD from 571 to 885（+54.99%）, while IS falls from 41.0 to 21.4（-47.80%）(Table 1(b), PDF p.6). More important than the optimal location is thatmultiple timelines are the mechanism, not decoration.

8.3 Coupled masking, normalization, and loss weighting

Removing masks slightly worsens FID 29.56→29.90 and FDD571→584. Removing stage standardization instead improves FID 29.56→27.13 and FDD571→505(by 8.22%、11.56%respectively). The authors attribute this to smaller second-stage amplitudes, effectively allocating more capacity to structure. Normalization implicitly reweights loss rather than merely stabilizing numerics. It is disabled for ImageNet but enabled for Kinetics, demonstrating autoencoder dependence(Table 1(c), Appendix A.5, PDF pp.6,15).

8.4 Parameter specialization: fairness excludes the best result

Among None/Mod/Proj/Cond/Attn/MLP/Full specialization, Full FID/FDD/IS is 26.63/526/45.5, improving over 29.56/571/41.0 by 9.91% / 7.88% / 10.98%. Main experiments omit Full to match baseline parameter counts(Table 1(d), PDF p.6). This separates representation/time decomposition from simply adding scale-specific parameters.

8.5 Compute allocation and decomposition scales

Spending saved compute on larger batches gives FID 31.95, versus 29.56 with smaller patches, motivating equal-token patching. At 1024px,512→1024,256→512→1024, and 256→1024 yield FID 58.03,50.61,41.06. More stages are not always better; the coarse starting point must be low enough without excessive stages.512→1024 trails256→1024 by 41.32%, suggesting the first stage must genuinely isolate structure(Table 1(e,f), PDF p.6).

8.6 Inference thresholds and steps

Main512px experiments use τ=.7and 1024px use .95; fixed 40-step sampling allocates30/10. Figures 4/8 provide curves without complete threshold×step tables, so exact decimals cannot safely be extracted. Supported trends: at high CFG, τ is less sensitive; more coarse steps improve structure, while fine steps visibly repair vegetation, grass, hair, and complex textures with diminishing returns(Sec.4.3, Appendix C/G, PDF pp.7,17,19). Universal optimality of 30/10 is unsupported.

8.7 Statistical significance

not reported by the papermulti-seed means, standard deviations, confidence intervals, significance tests, or FID/FDD bootstrap estimates for main tables/ablations. Evaluating10k/50k samples does not replace training-seed variance. Large differences such as FDD132.5→85.8 have engineering significance, but whether Kinetics FID 6.98→6.85 or FDD235.0→236.5 exceeds noise remains unknown.

9.1 Flow Matching（Lipman et al., ICLR 2023）

Problem addressed: Conditional-path vector-field regression avoids training-time CNF integration, providing scalable simulation-free continuous generation.Remaining gap: One full-resolution time t does not explicitly separate structure and texture.How DFM responds: Preserves linear paths and MSE, extending t to {t^s} and changing their joint training distribution.Transferred design: Standard ODE solvers, velocity prediction, and classifier-free guidance remain, making DFM a structured FM wrapper rather than a replacement.

9.2 Cascaded Diffusion Models（Ho et al., JMLR 2022）

Problem addressed: Low-resolution generators plus one or more super-resolution diffusion models split high-dimensional generation into stages.Remaining gap: Separate models per stage add training/maintenance costs and exposure gaps from earlier-stage errors.How DFM responds: One backbone jointly trains all scales, masking future stages with noise and conditioning on low-noise past stages.Transferred design: Retains the low-resolution-first curriculum; DFM’s p0=.9 and 30/10 allocation inherit cascade-style first-stage budget emphasis.

9.3 f-DM（Gu et al., ICLR 2023）

Problem addressed: Progressive signal transformations such as downsampling enter diffusion dynamics, changing resolution over time.Remaining gap: Generation/sampling becomes transformation-dependent, with specialized stage transitions.How DFM responds: Each component has a standard linear flow; transitions change which timelines advance rather than changing state spaces.Transferred design: f-DM demonstrates useful progressive signals; DFM moves transforms into representations and returns dynamics to ordinary FM.

9.4 Matryoshka Diffusion Models（Gu et al., ICLR 2024）

Problem addressed: NestedUNet jointly denoises multiple resolutions within one model and supports low-to-high progressive training.Remaining gap: Nested architectures and pixel-space multiscale designs are heavy and bind decomposition to architecture.How DFM responds: Shares an ordinary DiT backbone, duplicating only ports/time embeddings in latent space.Transferred design: Retains joint multiscale output training and low-to-high feature support, replacing nested sharing with equal token counts and summed embeddings.

9.5 Pyramidal Flow Matching（Jin et al., ICLR 2025）

Problem addressed: Spatial/temporal latent pyramids generate video with full resolution only at the final stage, reducing long-video cost.Remaining gap: Interstage upsampling, re-noising, and continuity require a specific sampler.How DFM responds: Each Laplacian component flows independently; thresholds only activate stages, without manually converting distributions.Transferred design: Coarse-to-fine latent flows, equal-compute stage comparisons, and video evaluation transfer directly. Table 2 uses Pyramidal Flow as the strongest neighboring baseline.

9.6 Nonlinear technical evolution

                                                  ┌─ Cascaded Diffusion ── Multiple models, super-resolution conditioning
Diffusion / Score ───────┼─ f-DM / Relay ─────── Modified dynamics/stage transitions
                         └─ Matryoshka ───────── One nested multiresolution model

Flow Matching ───────────┬─ Pyramidal Flow ───── Latent pyramids + upsampling/re-noising
                         └─ DFM ──────────────── Multiple timelines + arbitrary decomposition + standard solver
                                                │
Scale-equivariant AE ────────────────────────────┘ Makes latent frequency hierarchy usable

Three shifts drive evolution: full-resolution generation to coarse-to-fine; one model per stage to shared models; and stage-specific dynamics to representation/time scheduling. DFM mainly contributes to the third.

9.7 Cross-paper comparison

10.1 Actual contributions

1. Decouples progression from diffusion equations. Progression comes from representation {X^s} and time vector {t^s}without modifying single-scale noise dynamics.
2. Specifies a joint training distribution. Stage index, current/past/future time sampling, and masks simulate inference states; tied-time ablation establishes their central role.
3. Fair neighboring-method controls. Flow Matching, cascades, and Pyramidal Flow share DiT-XL and equal training compute; tables span images, video, and three CFG values.
4. Connects the method to FLUX fine-tuning. Shows applicability beyond small from-scratch prototypes, though evidence is internal-domain.

10.2 Reusable designs

• stage-specific ports + shared trunk Transferable to wavelets, DCT(after inverse spatial transformation), or multiscale learned tokenizers.
• vector timestep conditioning Suitable for block-decomposable states such as spacetime blocks, semantics/texture, or geometry/material.
• Loss masks simulate inactive future components as a general curriculum tool, preventing pure-noise branches from contaminating shared representations.
• Stage budgets under fixed total NFE can become sample-adaptive controllers: more coarse budget for complex structure, more fine budget for complex textures.

10.3 Transfer potential

Video could combine spatial Laplacian and temporal pyramids into two-dimensional stages;3D could separate occupancy/coarse geometry from textures/materials; audio could separate low-frequency envelopes and high-frequency transients; discrete generation could separate lexical skeletons and surface forms. Each layer must support linear interpolation and backbone sharing; image success alone does not establish transfer.

11.1 Limitations stated by the paper

1. Additional hyperparameters. Stage probabilities, past-stage noise shifts, inference thresholds, and step allocations jointly control frequency balance(Discussion, PDF p.9; HTML lines 255–259).
2. High-frequency artifacts. Vegetation, hair, fine structures, and clutter may flatten; more second-stage steps alleviate but do not eliminate this(Appendix G, PDF p.19).
3. Limited decomposition/scale coverage. Main results center on Laplacian; direct DCT input fails, with wavelets, multiscale AEs, and larger scales deferred(Discussion/Appendix F, PDF pp.9,18–19).

11.2 Independent critique(evidence→inference)

1. [Statistics] Because Tables 1–3 lack seed counts, standard deviations, CIs, or bootstrap estimates, Kinetics FID’s 0.13 advantage and FDD’s 1.5 disadvantage cannot be treated as stable. Engineering judgments are safer for large FDD gaps such as 46.7.
2. [Reproduction] Because the official repository explicitly reimplements on SiT with a different architecture and forthcoming checkpoints, open code does not make the main tables reproducible. Third parties can currently reproduce the concept, not necessarily Table 2 values.
3. [Data extrapolation] Because FLUX’s24k+32k fine-tuning,10k prompts, and scale-equivariant AE use internal data, FID 12.56→8.83 cannot generalize to public prompt/aesthetic distributions or original FLUX-dev quality. It supports faster internal-domain adaptation only.
4. [Attribution] Because the pipeline adds 10k AE fine-tuning steps and Table 4’s scale-equivariant AE alone improves ordinary FM FID 54.50→43.16, gains are not solely from extra timelines. DFM also improves54.50→34.00 on the original AE, establishing independent value, but DFM–AE regularization interactions remain incompletely isolated.
5. [Hyperparameter selection] Because in DiT-B ablations p0=.95 has FID/FDD/IS 27.5/540/43.7 all superior to .9 interval's 29.56/571/41.0, while main experiments choose .9 based only on speculation that larger models need more detail capacity, main settings are not directly supported ablation optima. Equal-cost XL sweeps are needed.
6. [Video evidence] Because at cfg1.5 DFM Kinetics FDD 236.5 is worse than Cascaded’s 235.0and FVD beats Pyramidal Flow by only 8.5 / 3.03%, comprehensive image/video superiority should be narrowed to most metrics except Kinetics FDD. Video gains are notably weaker.
7. [Evaluation validity] Because public benchmarks report only FID/FDD/IS/FVD, without human structural preferences, local texture errors, prompt compositionality, or safety, claims of better structural detail rely mainly on qualitative images and DINO distances. Lower FDD does not prove every semantic relation is more correct.
8. [Numerical audit] Because Table 3’s12.56→8.83 recalculates to 29.70%but text says 28.7%, and Pyramidal116.2→85.8 recalculates to 26.16% rather than 26.4%, at least two percentage reports are imprecise. They do not overturn conclusions but reduce confidence in exact reuse.

Theory

1. Sufficient conditions for decomposition equivariance. Fix DiT, parameters, and training FLOPs; compare Laplacian, Haar wavelets, inverse-transformed DCT, and learned spectral tokenizers. Measure cross-scale mutual information, velocity-field Lipschitz estimates, and FID/FDD. If frequency decoupling helps, predictable correlations should emerge beyond empirical rankings.
2. Training-distribution/inference-trajectory coverage. Compare hand-designed logit-normal sampling, joint p(t^1,…,t^S)fitted to actual sampler trajectories, and adversarial worst-case timestep sampling. At 500k steps, evaluate train/inference hidden-state MMD and generation metrics for exposure gaps.

Engineering and applications

3. Equal GPU-hours rather than iterations. Compare FM, DFM, Cascaded, and Pyramidal Flow at equal H100-hours, parameters, and NFE. Report training throughput, peak memory, communication, sampling latency, and energy to test whether half the iterations really means half the cost.
4. Public FLUX adaptation. Use a public LAION-Aesthetics subset and fixed 10k prompts, releasing AE,24k-start, and both32k-branch checkpoints. Compare FLUX-FT/DFM/LoRA-DFM with blinded human evaluation, GenEval, DPG-Bench, and FID/FDD, separating data from method gains.
5. Texture-adaptive budgets. A lightweight controller allocates remaining NFE using coarse-preview edge spectra/semantic complexity. Compare against30/10 at mean 40 NFE on hair, vegetation, text, and regular-geometry failure sets, measuring LPIPS patch error, human ratings, and P95 latency.

Emerging directions

6. Two-dimensional spatial×temporal decomposition. Compare spatial-only DFM, temporal-only pyramids, and Cartesian-product stages at equal token-FLOPs on Kinetics, long-video consistency, and physical plausibility, testing whether more stages dilute optimization.
7. 3D geometry/material separation. Within one3D DiT, assign independent flows to coarse occupancy, surface detail, and texture. Compare single flows, model cascades, and triplane pyramids using Chamfer, normal consistency, texture FID, and editability, checking geometry-error propagation into materials.
8. Safety and provenance. Use stage ports for low-frequency semantic/high-frequency invisible watermarks; compare NoisePrints and ordinary latent watermarking. Test detection and quality after cropping, compression, and re-diffusion to assess new multiscale provenance controls.