Recent advancements in video generation models have significantly improved their ability to follow text prompts. However, the customization of dynamic visual effects, defined as temporally evolving and appearance-driven visual phenomena like object crushing or explosion, remains underexplored. Prior works on motion customization or control mainly focus on low-level motions of the subject or camera, which can be guided using explicit control signals such as motion trajectories. In contrast, dynamic visual effects involve higher-level semantics that are more naturally suited for control via text prompts. However, it is hard and time-consuming for humans to craft a single prompt that accurately specifies these effects, as they require complex temporal reasoning and iterative refinement over time. To address this challenge, we propose P-Flow, a novel training-free framework for customizing dynamic visual effects in video generation without modifying the underlying model. By leveraging the semantic and temporal reasoning capabilities of vision-language models, P-Flow performs test-time prompt optimization, refining prompts based on the discrepancy between the visual effects of the reference video and the generated output. Through iterative refinement, the prompts evolve to better induce the desired dynamic effect in novel scenes. Experiments demonstrate that P-Flow achieves high-fidelity and diverse visual effect customization and outperforms other models on both text-to-video and image-to-video generation tasks. Code is available at https://github.com/showlab/P-Flow.
Figures & tables
Figure 1: It is hard for humans to craft text prompts that precisely control video generation models to generate desired visual effects across diverse scenes, while P - F l o w automatically refines prompts to achieve consistent and realistic visual effects.
Figure 2: Overview of the proposed P - F l o w framework.
Image-to-Video
Text-to-Video
Overall
FID-VID ↓
FVD ↓
Dyn. Degree ↑
FID-VID ↓
FVD ↓
Dyn. Degree ↑
FID-VID ↓
FVD ↓
Dyn. Degree ↑
Wan 2.1 [ 69 ]
34.62
994.70
0.33
42.32
1535.43
0.28
38.47
1265.07
0.31
HunyuanVideo [ 32 ]
36.53
1169.9
0.51
38.35
1362.35
0.34
37.44
1266.13
0.43
VFX Creator ( Training-Based ) [ 44 ]
29.92
752.95
0.63
–
–
–
–
–
–
Wan 2.1 + HF ( Training-Free )
33.12
989.42
0.54
42.21
1632.10
0.59
75.33
1310.76
0.57
HunyuanVideo + HF ( Training-Free )
33.85
1035.49
0.70
36.54
1266.79
0.62
35.20
1151.14
0.66
Table 1: Quantitative comparisons for both image-to-video and text-to-video generation settings. Note that VFX Creator supports only image-to-video generation.
Figure 3: Qualitative comparison on image-to-video generation with different visual effects. The prompts shown beneath each row represent the actual input to each model. As VFX Creator is optimized for short phrase inputs, such prompts are provided to ensure a fair and consistent evaluation.
Model 1
Preference (Model 1 V.S. Model 2)
Model 2
P - F l o w -I2V
80% V.S. 20%
Wan 2.1-I2V
P - F l o w -I2V
84% V.S. 16%
HunyuanVideo-I2V
P - F l o w -I2V
58% V.S. 42%
VFX Creator
P - F l o w -T2V
75% V.S. 25%
Wan 2.1-T2V
P - F l o w -T2V
81% V.S. 19%
HunyuanVideo-T2V
Table 2: Human evaluation results (%) comparing P - F l o w against baseline models in Image-to-Video (I2V) and Text-to-Video (T2V) generation. Note: Model order was randomized during evaluation.
Figure 4: Qualitative comparison on image-to-video generation with different visual effects. The prompts shown beneath each row represent the actual input to each model.
Figure 5: Optimization trajectory of P - F l o w . Starting from a simple base prompt, P - F l o w iteratively refines the text prompt based on the visual feedback (we showcase the iteration 1 → 3 → 5), leading to progressively more accurate alignment between the generated video and the target “squish” visual effect.
Modules
Image-to-Video
Text-to-Video
Overall
Noise-Enhance
Logic-Context
Visual-Context (i-1)
FID-VID ↓
FVD ↓
Dyn. Degree ↑
FID-VID ↓
FVD ↓
Dyn. Degree ↑
FID-VID ↓
FVD ↓
Dyn. Degree ↑
✗
✗
✗
33.42
1089.23
0.64
39.85
1321.70
0.61
36.64
1205.47
0.63
✓
✗
✗
32.80
961.78
0.69
36.74
1182.42
0.66
34.77
1072.10
0.68
✓
✗
✓
30.45
861.52
0.83
34.05
1044.67
0.78
32.25
953.10
0.81
✓
✓
✓
29.32
784.51
0.94
32.93
980.75
0.87
31.13
882.63
0.91
Table 3: Ablation study of P - F l o w on both image-to-video and text-to-video generation.
Setting
FID-VID ↓
FVD ↓
Dyn. Degree ↑
w/o SVD Projection ( ρs=0.0,ρm=1.0 )
33.25
1052.80
0.58
Random Noise Only ( α=0.0 )
32.74
923.67
0.73
Enhanced Noise ( α=0.001,ρs=0.1,ρm=0.9 )
29.32
784.51
0.94
Enhanced Noise ( α=0.01,ρs=0.1,ρm=0.9 )
29.21
803.35
0.88
Table 4: Analysis of noise prior enhancement components.
Visual effects (VFX) are essential for enhancing the expressiveness and creativity of video content, yet producing high-quality effects typically requires expert knowledge and costly production pipelines. Existing AIGC systems face significant challenges in VFX generation due to the scarcity of effect-specific data and the inherent difficulty of modeling supernatural or stylized effects. Moreover, these approaches often require per-effect fine-tuning, which severely limits their scalability and generalization to novel VFX. In this work, we present EffectMaker, a unified reasoning-generation framework that enables reference-based VFX customization. EffectMaker employs a multimodal large language model to interpret high-level effect semantics and reason about how they should adapt to a target subject, while a diffusion transformer leverages in-context learning to capture fine-grained visual cues from reference videos. These two components form a semantic-visual dual-path guidance mechanism that enables accurate, controllable, and effect-consistent synthesis without per-effect fine-tuning. Furthermore, we construct EffectData, the largest high-quality synthetic dataset containing 130k videos across 3k VFX categories, to improve generalization and scalability. Experiments show that EffectMaker achieves superior visual quality and effect consistency over state-of-the-art baselines, offering a scalable and flexible paradigm for customized VFX generation. Project page: https://effectmaker.github.io
Autoregressive video models can generate minute-long videos in real time, but they produce generic subjects from text rather than specific subjects from user-provided images. Existing customization methods either require costly per-subject optimization or use pretrained conditioning networks that jointly process all video frames with bidirectional attention. Neither approach is designed for causal streaming. We present Custom Forcing, a training-free method that stores reference-based anchor frames in the persistent KV cache of a frozen autoregressive video model. However, fixed anchors face two limitations: simple conditioning allows identity to drift, and the text prompt continues to favor a generic subject. To address these problems, drift-adaptive value amplification (DVA) scales reference influence with the degree of identity drift, while anchor contrast guidance (ACG) steers generation away from the generic class prior. Over two-minute rollouts, fixed anchors fall from 0.58 to 0.42 in DINO-I, while Custom Forcing keeps it between 0.58 and 0.62 without reducing motion. Custom Forcing also achieves higher subject similarity than bidirectional customization methods and better preserves identity over 30s than causal image-to-video and reference-to-video models, while generating each frame 9.5-28.5 times faster than these long-video baselines.
Yunseung Ok, Hyunsoo Kim, Minseo Kim +1
Kyung Hee University · The University of Texas at Austin
Despite recent advances in video generation, motion customization remains challenging due to content leakage, where appearance attributes from the reference video unintentionally propagate into the generated output. We identify this issue as a consequence of the generative process collapsing toward the reference video, which arises from formulating the learning objective as a direct regression on the reference. To address this, we propose Control-based Motion Customization (CMC), a principled training framework that is structurally robust to content leakage. Our key idea is to steer generative dynamics toward desired motion while avoiding collapse toward the reference video, which we formalize using Stochastic Optimal Control (SOC). Under this formulation, customized videos acquire the target motion yet remain within the pre-trained model's prompt-conditional distribution, where appearance is determined by the text prompt rather than the reference video. Furthermore, to improve efficiency, we tailor the SOC formulation to motion customization by eliminating the need for an explicit reward and introducing a timestep-adaptive motion cost that focuses only on early generative stages, accelerating training by 2.5 times. Extensive experiments demonstrate that CMC effectively mitigates content leakage and achieves competitive motion fidelity while preserving the diversity of the base model across diverse scenarios.