TunerDiT: Training-free Progressive Steering of Diffusion Transformer for Multi-Event Video Generation

Abstract

Text-to-video (T2V) generation faces challenging questions when generating videos with long horizons containing multiple events. Inspired by the intrinsics of the diffusion process, we probe video diffusion transformers (DiTs) and uncover intrinsic turning points in the DiT denoising trajectory where conditioning text affects generation from global layout to fine-grained details. Building on this finding, we present TunerDiT, a simple yet effective progressive steering method that requires no additional training for multi-event generation. TunerDiT comprises two steering handles: (1) Event-Partitioned Masking that enforces event boundaries while allowing cross-event transition bands; (2) Cross-Event Prompt Fusion that injects neighboring event semantics for late-stage refinement. We contribute a self-curated prompt suite for benchmarking multi-event generation, i.e., MEve. TunerDiT achieves state-of-the-art performance across 8 metrics and offers a tunable trade-off between video consistency and event separation, compared with other training-free methods. The improvement in text alignment increases with the event count, indicating a scaling possibility with increasing event count.

Fig. 1: Different denoising steps utilize text inputs differently. TunerDiT derives this insight by probing the prompt conditioning of video diffusion models. When switching the input prompt at various fractions of the denoising steps (e.g., 0%, 10%, 20%, 40%, 70%, 100%), early steps dominate the global layout while late steps refine fine-grained appearance and motion, revealing an intrinsic turning point τ* where text influence changes.

Method Overview

TunerDiT operates without any training by intervening around the intrinsic turning point τ* of the DiT denoising trajectory. It first generates a coherent shared layout for all events, then unfuses the prompts and enforces event boundaries while preserving transition continuity via two progressive handles:

Cross-Event Prompt Fusion (PF)

In early coarse-layout steps (n < τ_PF), all video segments share the first-event prompt P₁ to establish a unified global scene and camera composition. After τ_PF, prompts are unfused to guide event-specific content.

Event-Partitioned Mask (EM)

Activated at τ_EM, a diagonal attention mask restrains cross-event attention to prevent semantic leakage. Simultaneously, dynamic inter- and inner-event transition bands allow adjacent boundaries to exchange context smoothly following a ReLU decay schedule.

Fig. 2: TunerDiT progressively steers multi-event generation over diffusion steps. Cross-Event Prompt Fusion (PF) first shares a common prompt to build a coherent layout, then gradually sepa- rates event prompts. Event-Partitioned Mask (EM) subsequently enforces event isolation via diago- nal cross-attention blocks and introduces cross-event transition bands around boundaries, enabling smooth and semantically consistent handovers between events.

MEve Benchmark

To evaluate multi-event video generation rigorously, we curated MEve, a dedicated benchmark containing up to 4 sequential events per prompt across diverse domains:

General (60 prompts): LLM-synthesized narratives authored to avoid content biases.
Motion Order Understanding (98 prompts): Sequences specifically evaluating temporal order and motion transitions.
Human Identity (32 prompts): Complex subject identity persistence across multiple distinct actions.
Complex Plot (60 prompts): Narrative sequences with sequential causations and scene changes.
Ego-Exo4D (50 × 2 prompts): Paired first-person (egocentric) and third-person (exocentric) narrations derived from real-world skilled activities.

Multi-Event Qualitative Comparisons

Comparison of TunerDiT against standard base models and training-free multi-event baselines across sequential events.

MEVG

FreeNoise

DiTCtrl

SwitchCraft

TunerDiT

Prompt: "A person is writing an email. Then the person starts baking cookies. The person places the cookie tray into the oven."

MEVG

FreeNoise

DiTCtrl

SwitchCraft

TunerDiT

Prompt: "A person is reading a book. Then the person suddenly starts folding laundry and stacks the folded clothes neatly in a basket."

MEVG

FreeNoise

DiTCtrl

SwitchCraft

TunerDiT

Prompt: "A dog is playing with a ball. Then the dog suddenly lies down on the carpet. Afterward, the dog rests its head on its paws and closes its eyes."

MEVG

FreeNoise

DiTCtrl

SwitchCraft

TunerDiT

Prompt: "A horse is running in the field. Then the horse starts drinking from a stream. Then, the horse lifts its head and shakes water from its muzzle."

Quantitative Evaluation

Comprehensive evaluation across (a) automatic embedding metrics for 2–4 events, and (b) VLM-as-a-judge and human evaluation. Best multi-event method is highlighted in bold, second-best is underlined.

(a) Automatic Metrics over 2–4 Events

Comparison with training-free methods and base models before and after (→) applying TunerDiT.

	TA ↑			TIS ↑			BC ↑			IC ↑			MS ↑
Shot Number	2	3	4	2	3	4	2	3	4	2	3	4	2	3	4
Training-free Multi-Event Methods
MEVG	0.201	0.206	0.205	0.271	0.270	0.272	0.839	0.842	0.821	0.711	0.718	0.709	0.941	0.932	0.928
DiTCtrl	0.186	0.207	0.216	0.259	0.271	0.278	0.872	0.888	0.881	0.767	0.765	0.770	0.972	0.979	0.976
FreeNoise	0.197	0.199	0.206	0.272	0.267	0.275	0.863	0.870	0.854	0.725	0.747	0.709	0.962	0.967	0.958
SwitchCraft (Wan2.1)	0.184	0.194	0.205	0.267	0.266	0.275	0.943	0.928	0.935	0.916	0.891	0.903	0.976	0.974	0.972
Base Video Generation Models & With TunerDiT
Open-Sora 1.2	0.174	0.179	0.182	0.270	0.269	0.276	0.941	0.940	0.945	0.904	0.908	0.901	0.985	0.983	0.988
↳ + TunerDiT	0.202	0.211	0.213	0.277	0.284	0.281	0.943	0.938	0.937	0.899	0.891	0.894	0.981	0.984	0.983
Open-Sora 2.0	0.194	0.198	0.199	0.270	0.276	0.281	0.929	0.932	0.931	0.871	0.867	0.861	0.991	0.990	0.984
↳ + TunerDiT	0.210	0.211	0.219	0.280	0.277	0.287	0.924	0.928	0.927	0.848	0.851	0.850	0.981	0.982	0.984
Wan2.1	0.182	0.184	0.183	0.265	0.267	0.263	0.944	0.945	0.947	0.922	0.918	0.919	0.979	0.977	0.981
↳ + TunerDiT	0.210	0.212	0.209	0.275	0.272	0.281	0.940	0.939	0.937	0.918	0.910	0.913	0.977	0.981	0.980
Wan2.2	0.185	0.201	0.201	0.270	0.271	0.275	0.921	0.916	0.919	0.849	0.847	0.851	0.987	0.985	0.983
↳ + TunerDiT	0.201	0.208	0.211	0.273	0.277	0.279	0.913	0.914	0.916	0.844	0.850	0.848	0.981	0.982	0.984

(b) Human-Preference and VLM-as-Judge Alignment

Left: VLM-judge Event Isolation (EI) and Text-Video Alignment (TVA) across 2–4 events.
Right: 18-person human study on a 5-point Likert scale (Q1: Preference, Q2: Motion Naturalness, Q3: Transition Smoothness, Q4: Text Alignment).

	EI ↑			TVA ↑
Model	2	3	4	2	3	4
Training-free Baselines
MEVG	0.360	0.435	0.515	1.335	1.375	1.451
FreeNoise	0.362	0.436	0.469	1.360	1.400	1.514
DiTCtrl	0.305	0.375	0.449	1.385	1.425	1.479
SwitchCraft (Wan2.1)	0.401	0.482	0.563	1.412	1.517	1.521
Ours
TunerDiT Open-Sora 1.2	0.423	0.499	0.580	1.450	1.492	1.524
TunerDiT Open-Sora 2.0	0.442	0.572	0.639	1.490	1.533	1.631
TunerDiT Wan2.1	0.419	0.542	0.628	1.437	1.617	1.648
TunerDiT Wan 2.2	0.399	0.474	0.553	1.462	1.503	1.549

Model	Q1	Q2	Q3	Q4
Training-free Baselines
MEVG	1.82	1.91	1.82	2.05
FreeNoise	1.99	2.01	1.99	1.79
DiTCtrl	2.11	2.03	2.18	2.35
SwitchCraft (Wan2.1)	3.20	3.60	3.40	3.10
Ours
TunerDiT Open-Sora 1.2	2.66	2.63	2.62	2.94
TunerDiT Open-Sora 2.0	3.16	3.03	3.03	3.34
TunerDiT Wan 2.1	3.75	3.60	3.50	3.40
TunerDiT Wan 2.2	3.01	2.97	2.82	3.30

BibTeX

@misc{liao2026tunerdittrainingfreeprogressivesteering,
      title={TunerDiT: Training-free Progressive Steering of Diffusion Transformer for Multi-Event Video Generation}, 
      author={Ruotong Liao and Guowen Huang and Qing Cheng and Guangyao Zhai and Lei Zhang and Xun Xiao and Thomas Seidl and Daniel Cremers and Volker Tresp},
      year={2026},
      eprint={2605.31590},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.31590}, 
}

🎛️ TunerDiT: Training-free Progressive Steering of Diffusion Transformer for Multi-Event Video Generation

Abstract

Method Overview

MEve Benchmark

Multi-Event Qualitative Comparisons

Quantitative Evaluation

BibTeX