DiT-Flow: Diffusion Transformers & Flow Matching

• DiT-Flow is a paradigm combining diffusion transformers with flow matching objectives, where models learn continuous-time velocity fields to transport samples from noise to data.
• It employs an ODE-driven formulation that replaces traditional denoising with velocity regression, facilitating efficient and versatile generation across several modalities.
• Various systems illustrate its impact by demonstrating improved generation metrics and practical enhancements in image editing, audio restoration, and video motion transfer.

DiT-Flow denotes the conjunction of Diffusion Transformers with flow-matching objectives, typically in the rectified-flow sense: a Transformer learns a continuous-time velocity field that transports samples from a simple base distribution to the data distribution along a prescribed path, and generation proceeds by integrating the resulting probability flow ordinary differential equation. In current usage, the term names both a general modeling pattern and specific systems. It appears as a generic label for flow-matching Diffusion Transformers in visual generation (Tian, 17 Jun 2026), as the title of a latent-space speech-enhancement framework built on a uDiT backbone (Cao et al., 23 Mar 2026), and in the closely related form “DiTFlow” for a training-free motion-transfer method specialized to video DiTs (Pondaven et al., 2024). Across these usages, the unifying idea is not a single fixed architecture but a family of Transformer-based continuous-time generative models and controls built around velocity regression rather than conventional denoising objectives.

1. Definition and terminological scope

In the most explicit formulation, DiT-Flow refers to Diffusion Transformers trained with flow-matching objectives, where the model directly learns a continuous-time velocity field that transports samples from noise to data along a prescribed path (Tian, 17 Jun 2026). This places DiT-Flow within the broader class of probability-flow models: the generative model is a vector field $v_\theta(x_t,t)$, and sampling amounts to solving an ODE rather than iterating a reverse Markov chain.

The label has broadened beyond a single benchmark setting. In image generation, it includes pixel-space and latent-space DiTs trained under linear or rectified flow objectives, such as SpectralDiT’s flow-matching DiTs on CIFAR-10 and ImageNet-100 (Tian, 17 Jun 2026). In editing, it covers FLUX- and SD3/3.5-class DiT priors used for drag-based manipulation under rectified flow, as in DragFlow (Zhou et al., 2 Oct 2025). In multimodal synthesis, it includes MM-DiT systems that jointly generate audio and facial motion with flow matching, as in JAM-Flow (Kwon et al., 30 Jun 2025). In audio restoration, it denotes a specific speech-enhancement architecture operating in VAE latent space with a uDiT backbone and Mixture-of-LoRA-Experts adaptation (Cao et al., 23 Mar 2026). Related DiT-specific control methods, such as DiTFlow for video motion transfer, use pretrained DiT denoisers rather than retraining the backbone, but still exploit DiT internal structure in a flow-like continuous-time setting (Pondaven et al., 2024).

A common misconception is that DiT-Flow is merely a naming variant for diffusion transformers. The literature instead uses it to indicate a change in training target and sampling formalism: the network regresses a velocity field and is sampled via an ODE, even when the same Transformer backbone family is retained (Tian, 17 Jun 2026).

2. Continuous-time formulation and relation to diffusion

A representative DiT-Flow formulation learns a velocity field through the probability flow ODE

$\frac{d x_t}{d t} = v_\theta(x_t,t).$

Under a common rectified-flow path,

$$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$

the target velocity is time-independent:

$v^*(x_t,t)=x_1-x_0.$

Training then minimizes an $\ell_2$ regression loss on the velocity field,

$$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$

This is the formulation used explicitly for flow-matching DiTs in SpectralDiT, with sampling by Euler integration of the probability flow ODE using 50 deterministic steps (Tian, 17 Jun 2026).

Other DiT-Flow systems use the same principle under different interpolation schedules. DyDiT++ adopts a path of the form

$$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$

and learns a velocity field via

$$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$

retaining the same ODE sampling view while allowing timestep-conditioned dynamic computation (Zhao et al., 9 Apr 2025). JAM-Flow applies conditional flow matching separately to audio and motion branches, summing modality-specific losses while coupling the branches through joint attention (Kwon et al., 30 Jun 2025).

The relation between diffusion and flow matching is more subtle than a simple opposition. Under Gaussian noising, “Score Distillation of Flow Matching Models” shows that diffusion and flow matching are theoretically equivalent, differing mainly in timestep weighting and schedules; scores, $x_0$ estimates, and flow velocities are linear transforms of one another under the Gaussian forward model (Zhou et al., 29 Sep 2025). This does not collapse the methods into a single implementation class, but it does mean that several acceleration and distillation techniques developed for diffusion transfer directly to DiT-based flow-matching teachers.

Sampling regimes vary substantially across the literature. Representative systems use Euler with 50 deterministic steps in image generation (Tian, 17 Jun 2026), 32 NFEs for joint audio-motion synthesis (Kwon et al., 30 Jun 2025), midpoint/RK2 with a sigmoid schedule for 5–10-NFE text-to-image generation (Zhuo et al., 2024), four-step score-identity distillation from pretrained text-to-image flow-matching DiTs (Zhou et al., 29 Sep 2025), and one-step Flow Generator Matching for ReFlow teachers and MM-DiT-based Stable Diffusion 3 distillation (Huang et al., 2024).

3. Backbone realizations and conditioning schemes

A representative visual DiT-Flow backbone follows the standard DiT pattern: an image $x \in \mathbb{R}^{C\times H\times W}$ is divided into non-overlapping $\frac{d x_t}{d t} = v_\theta(x_t,t).$0 patches, producing $\frac{d x_t}{d t} = v_\theta(x_t,t).$1 tokens embedded in $\frac{d x_t}{d t} = v_\theta(x_t,t).$2 and augmented with fixed 2D sinusoidal positional embeddings. Transformer blocks use AdaLN-Zero conditioning. Given timestep and class embeddings, each residual branch is modulated by scale, shift, and gate parameters produced by a small conditioning network, with zero-initialized gates so the model starts as an identity residual network conditioned on the embeddings (Tian, 17 Jun 2026).

Within this framework, architectural specialization occurs inside residual branches. SpectralDiT modifies only the final MLP residual update: it reshapes the modulated MLP residual onto the patch-token grid, decomposes it into low- and high-frequency components using a fixed depthwise separable low-pass filter with kernel $\frac{d x_t}{d t} = v_\theta(x_t,t).$3, and applies timestep-conditioned zero-initialized additive gates to the low- and high-frequency components before reshaping back to tokens (Tian, 17 Jun 2026). The design preserves the standard DiT computation and flow-matching training while learning block-wise spectral corrections over generation time.

Multimodal DiT-Flow systems adopt more heterogeneous backbones. JAM-Flow uses a Multi-Modal Diffusion Transformer in which Audio-DiT and Motion-DiT branches are coupled by selective joint attention layers, temporally aligned RoPE embeddings, and localized joint attention masking. Audio self-attention is global, motion self-attention is local-windowed, and cross-modal attention is constrained by aligned temporal indices. The model predicts mel-spectrograms and a mouth-expression subspace from LivePortrait within a single FM-based MM-DiT (Kwon et al., 30 Jun 2025).

Other realizations alter the DiT backbone itself. Lumina-Next replaces Flag-DiT with Next-DiT, introducing 3D RoPE, sandwich normalizations, QK-Norm, tanh gating on the second AdaLN scale, and Grouped-Query Attention; this is still a flow-based large diffusion transformer, but one optimized for training stability, resolution extrapolation, and few-step ODE sampling (Zhuo et al., 2024). In speech enhancement, the DiT-Flow model uses a latent audio compressor followed by a uDiT backbone with extended skip connections, 12 layers, embedding dimension 384, and 6 attention heads, with conditioning implemented by concatenating distorted and perturbed latent features and supplying time $\frac{d x_t}{d t} = v_\theta(x_t,t).$4 to the backbone (Cao et al., 23 Mar 2026).

These differences matter because DiT-Flow is not architecturally monolithic. Patch-token visual DiTs, MM-DiTs, Next-DiT variants, FLUX-style joint image-text blocks, and latent audio uDiTs all instantiate the same flow-matching principle with different conditioning, tokenization, and residual-modulation schemes.

4. Representative systems and empirical profiles

The literature uses DiT-Flow across generation, editing, restoration, and control. The following systems illustrate the range of current practice.

System	Domain	Core mechanism
SpectralDiT	Image generation	Timestep-conditioned spectral correction in the MLP residual branch
DragFlow	Drag-based image editing	Region-level affine supervision, adapter-enhanced inversion, hard background constraints
JAM-Flow	Joint audio-motion synthesis	Audio-DiT and Motion-DiT with selective joint attention
DiT-Flow	Speech enhancement	VAE latent space, uDiT backbone, MoELoRA adaptation
DiTFlow	Video motion transfer	Attention Motion Flow from cross-frame self-attention

SpectralDiT shows how small architectural changes within a DiT-Flow block can affect sample quality and spectral fidelity. On CIFAR-10 pixel-space generation with patch size 1, it improves FID-10K from $\frac{d x_t}{d t} = v_\theta(x_t,t).$5 to $\frac{d x_t}{d t} = v_\theta(x_t,t).$6, increases recall from $\frac{d x_t}{d t} = v_\theta(x_t,t).$7 to $\frac{d x_t}{d t} = v_\theta(x_t,t).$8, and reduces the radial Fourier spectrum distance from $\frac{d x_t}{d t} = v_\theta(x_t,t).$9 to $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$0. In latent diffusion on ImageNet-100, under classifier-free guidance scale 2.0, it improves FID-50K from $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$1 to $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$2, with only $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$3 additional theoretical FLOPs and $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$4 additional parameters; all results are averaged over five seeds (Tian, 17 Jun 2026).

DragFlow addresses a failure mode specific to DiT-Flow editing: naively porting point-based drag supervision from UNet-based DDPMs to DiTs performs poorly because DiT features are finer-grained, more spatially precise, and have narrower receptive fields. DragFlow therefore replaces point-level supervision with region-level affine supervision, integrates pretrained personalization adapters such as InstantCharacter and IP-Adapter family, uses KV injection, and applies adaptive gradient-mask hard constraints for background fidelity. On ReD Bench it reports MD1 $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$5, MD2 $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$6, IF_bg $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$7, and IF_{s2t} $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$8; on DragBench-DR it reports MD1 $$x_t = (1-t)x_0 + t x_1,\qquad x_0 \sim p_{\text{data}},\; x_1 \sim \mathcal{N}(0,I),$$9, MD2 $v^*(x_t,t)=x_1-x_0.$0, IF_bg $v^*(x_t,t)=x_1-x_0.$1, and IF_{s2t} $v^*(x_t,t)=x_1-x_0.$2 (Zhou et al., 2 Oct 2025).

JAM-Flow extends DiT-Flow into coupled continuous-time audio and motion generation. It jointly models mel-spectrograms and mouth-expression trajectories, supports text-to-talking-head generation, audio-driven animation, motion-to-audio generation, and cross-modal infilling, and samples with 32 NFEs. On HDTF it reports, for I2V, FID $v^*(x_t,t)=x_1-x_0.$3, FVD $v^*(x_t,t)=x_1-x_0.$4, LSE-C $v^*(x_t,t)=x_1-x_0.$5, and LSE-D $v^*(x_t,t)=x_1-x_0.$6; for V2V it reports FID $v^*(x_t,t)=x_1-x_0.$7, FVD $v^*(x_t,t)=x_1-x_0.$8, LSE-C $v^*(x_t,t)=x_1-x_0.$9, and LSE-D $\ell_2$0. On LibriSpeech-PC test-clean TTS it reports WER $\ell_2$1 and SIM-o $\ell_2$2, while a variant with frozen Audio-DiT reports WER $\ell_2$3 and SIM-o $\ell_2$4 (Kwon et al., 30 Jun 2025).

The speech-enhancement model titled “DiT-Flow” adapts the paradigm to single-channel SE under compound distortions. It operates on VAE latent features derived from complex STFTs, uses a uDiT backbone, and is trained for robustness to noise, reverberation, and codec compression. On StillSonicSet under Reverb+Noise+Codec-Compression it reports PESQ $\ell_2$5, ESTOI $\ell_2$6, LSD $\ell_2$7, SIG $\ell_2$8, BAK $\ell_2$9, OVRL $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$0, and Spk Sim $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$1, outperforming SGMSE and StoRM on SIG, OVRL, and LSD. Its reported real-time factor is $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$2, versus $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$3 for SGMSE and $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$4 for StoRM. For adaptation to five unseen distortions, MoELoRA trains only $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$5 of parameters and achieves near–full finetuning performance across perceptual and intelligibility metrics (Cao et al., 23 Mar 2026).

DiTFlow, in the video literature, uses a pretrained video DiT rather than retraining a flow-matching backbone. It extracts Attention Motion Flow from cross-frame self-attention maps and guides denoising by optimization on latents or positional embeddings. On DAVIS with CogVideoX-5B across all prompts, it reports Motion Fidelity $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$6 and Image Quality $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$7, compared with $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$8 for SMM and $$L(\theta)=\mathbb{E}_{x_0,x_1,t}\left[\|v_\theta(x_t,t,y)-v^*(x_t,t)\|_2^2\right].$$9 for MOFT. For zero-shot motion transfer, positional-embedding optimization yields a motion-preservation drop of $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$0 and a prompt-adherence drop of $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$1, whereas latent optimization yields $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$2 and $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$3, respectively (Pondaven et al., 2024).

5. Efficiency, acceleration, and distillation

A major research direction within DiT-Flow is computational reduction without relinquishing the continuous-time formulation. DyDiT++ introduces timestep-wise dynamic width and spatial-wise dynamic token routing, so that attention heads, MLP channel groups, and selected tokens are activated conditionally on the generation timestep and, for tokens, on spatial content. Under flow matching this yields “DySiT” when applied to SiT-like DiT architectures. On DySiT-XL with ODE sampling, FLOPs drop from $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$4G to $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$5G at $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$6 with FID improving from $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$7 to $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$8; at $$x_t = \alpha_t x_0 + \sigma_t x_1,\qquad \alpha_t+\sigma_t=1,$$9, FLOPs drop to $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$0G with FID approximately $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$1. For DyFLUX-Lite at $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$2, FLOPs are reduced from $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$3T to $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$4T with approximately $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$5 wall-clock speedup, and integrating TeaCache yields an extra approximately $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$6 speedup (Zhao et al., 9 Apr 2025).

Lumina-Next pursues efficiency at the sampler and positional-encoding levels. It analyzes flow ODE truncation behavior, introduces rational and sigmoid time discretization schedules, and advocates a midpoint solver under a sigmoid schedule with $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$7, $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$8, and $$L_{\text{velocity}}=\mathbb{E}_{t,x_0,x_1}\left[\left\|v_\theta(x_t,t)-\frac{d x_t}{dt}\right\|^2\right],$$9. It also proposes Frequency- and Time-Aware Scaled RoPE for extrapolation and a time-aware Context Drop mechanism that merges redundant visual tokens by average pooling Keys and Values but not Queries. The reported result is high-quality text-to-image generation in 5–10 NFEs and a $x_0$0 inference-time speed-up at 1K resolution with $x_0$1 downsampling, while preserving or improving quality (Zhuo et al., 2024).

Distillation methods now treat DiT-Flow teachers directly. “Score Distillation of Flow Matching Models” shows that Score identity Distillation transfers to pretrained text-to-image flow-matching models including SANA, SD3/3.5, and FLUX without teacher finetuning or architectural changes. With a four-step student, SiD-DiT reduces teachers using 28–40 steps to 4 steps while maintaining or improving quality on several models: for SD3.5-Medium, teacher FID/CLIP/GenEval is $x_0$2 versus $x_0$3 for SiD-DiT; for SD3.5-Large, teacher performance is $x_0$4 versus $x_0$5 for SiD-DiT. The FLUX.1-dev result is weaker, with teacher $x_0$6 versus SiD-DiT $x_0$7, and the paper attributes part of this to FLUX’s learned guidance embedding rather than standard CFG (Zhou et al., 29 Sep 2025).

Flow Generator Matching goes further by targeting one-step generation. It defines a generator-induced path and proves that a tractable objective based on an auxiliary online flow model has the same gradient as an ideal intractable flow-matching objective under stated regularity assumptions. On CIFAR-10 unconditional generation, the one-step FGM model achieves FID $x_0$8, outperforming the teacher at 50 steps, which reports FID $x_0$9. For Stable Diffusion 3 distillation with an MM-DiT backbone, the resulting one-step MM-DiT-FGM model reports GenEval overall $x \in \mathbb{R}^{C\times H\times W}$0 at NFE $x \in \mathbb{R}^{C\times H\times W}$1, compared with $x \in \mathbb{R}^{C\times H\times W}$2 for the 28-step SD3 teacher, $x \in \mathbb{R}^{C\times H\times W}$3 for Flash-SD3 at 4 steps, and $x \in \mathbb{R}^{C\times H\times W}$4 for Hyper-SD3 at 4 steps (Huang et al., 2024).

6. Limitations, misconceptions, and open directions

Several recurrent limitations show that DiT-Flow is not a uniformly solved design space. In SpectralDiT, gains vanish at patch size 4 on $x \in \mathbb{R}^{C\times H\times W}$5 CIFAR-10, with FID changing from $x \in \mathbb{R}^{C\times H\times W}$6 to $x \in \mathbb{R}^{C\times H\times W}$7, consistent with aliasing on an $x \in \mathbb{R}^{C\times H\times W}$8 token grid; unoptimized PyTorch also introduces a wall-clock increase of approximately $x \in \mathbb{R}^{C\times H\times W}$9–$\frac{d x_t}{d t} = v_\theta(x_t,t).$00 despite only approximately $\frac{d x_t}{d t} = v_\theta(x_t,t).$01 theoretical FLOPs increase (Tian, 17 Jun 2026). In DragFlow, point-based supervision is brittle for DiT features, inversion drift is aggravated by CFG distillation in modern DiT-Flow models such as FLUX, and severe non-rigid changes or strong occlusions can exceed what region-level affine schedules capture (Zhou et al., 2 Oct 2025).

In multimodal DiT-Flow, architectural coupling can become unstable. JAM-Flow reports that fully joint attention yields slightly better numbers but unstable training in practice, so a half-joint configuration is adopted; lip-sync also degrades when localized masks are removed, and robustness depends on data quality, caption noise, and the LivePortrait renderer (Kwon et al., 30 Jun 2025). In dynamic-compute DiT-Flow, over-pruning heads, channel groups, or tokens can degrade quality, which is why DyDiT++ uses warm-up, magnitude-based ranking, and FLOPs-aware losses to stabilize routing (Zhao et al., 9 Apr 2025).

Task-specific systems expose domain-dependent trade-offs. In speech enhancement, DiT-Flow leads on SIG and OVRL under realistic compound distortions, but in the single-distortion reverb-only setting its PESQ trails SGMSE; the paper also notes that synthetic training cannot cover all real-world artifact modes (Cao et al., 23 Mar 2026). In video motion transfer, DiTFlow still struggles with out-of-distribution motions such as complex acrobatics, heavy occlusions, and drastic camera changes, and latent optimization can leak semantics from the optimization prompt (Pondaven et al., 2024). In Lumina-Next, FT-Scaled RoPE reduces repetition and preserves detail, but extreme extrapolation factors or aspect ratios may still exhibit localized artifacts (Zhuo et al., 2024).

A second misconception is that the diffusion–flow distinction is absolute. Under Gaussian assumptions, score and flow-velocity parameterizations are linearly convertible, which is precisely why SiD can distill rectified-flow and TrigFlow DiT teachers without architectural changes (Zhou et al., 29 Sep 2025). A plausible implication is that future DiT-Flow research will continue to blur the historical boundary between diffusion-model tooling and flow-matching tooling, while still preserving practically important differences in timestep schedules, guidance mechanisms, and solver design.

Open directions stated in the literature are correspondingly heterogeneous: JAM-Flow points to stronger video diffusion decoders, more modalities, and periodic long-range cross-modal attention (Kwon et al., 30 Jun 2025); the speech-enhancement DiT-Flow paper proposes broader real-recorded corpora, richer codec and reverberation simulators, and libraries of experts for domain-specific adaptation (Cao et al., 23 Mar 2026); DiTFlow suggests multi-scale AMF and stronger positional-embedding constraints (Pondaven et al., 2024); and score distillation for flow-matching DiTs identifies model-specific guidance integration, especially for FLUX-style learned guidance embeddings, as an unresolved issue (Zhou et al., 29 Sep 2025). Taken together, these directions indicate that DiT-Flow is best understood as an active research program centered on continuous-time Transformer generative modeling, rather than as a closed or singular architecture.