OmniOPD: Stable On-Policy Distillation for Joint Audio-Video Generation
Organizations: University of Science and Technology of China
Abstract
Reinforcement learning (RL) elevates joint audio-video generation, but optimizing heterogeneous, modality-specific audio and video rewards often introduces cross-modal interference. On-policy distillation (OPD) alleviates this issue by transferring the capabilities of specialized audio and video RL teachers into a single student model. However, a significant imbalance in cross-modal token counts leaves audio with sparse supervision, causing unstable distillation. We observe that naively increasing the audio loss weight fails to solve this problem since audio learning relies heavily on visual context, indicating richer visual cues are crucial to offset sparse audio signals. To address this, we propose OmniOPD, a modality-wise on-policy distillation framework designed for balanced and stable audio-video generation. First, we construct privileged visual information using asymmetric noise scheduling, allowing the audio teacher to generate more informative targets from cleaner video context. Second, we introduce grouped multi-noise sampling to expand audio supervision coverage and stabilize training. Third, we introduce timestep-wise importance weighting based on teacher-student transition divergence, preserving reliable supervision while down-weighting unreliable learning signals, to stabilize fluctuations in audio learning. Evaluated on JavisBench and VBench demonstrate OmniOPD preserves audio quality while significantly enhancing video generation, achieving effective joint distillation.
Figures & tables
| AV-Quality | Text-Consistency | AV-Consistency | AV-Synchrony | |||||||
| Model | VQ | AQ | TV-IB | TA-IB | CLIP | CLAP | AV-IB | AVHScore | JavisScore | DeSync |
| – T2A+A2V | ||||||||||
| TempoTkn ( Yariv et al., 2024 ) | – | – | 0.084 | – | 0.205 | – | 0.139 | 0.122 | 0.103 | 1.532 |
| TPoS ( Jeong et al., 2023 ) | – | – | 0.201 | – | 0.229 | – | 0.124 | 0.129 | 0.095 | 1.493 |
| – T2V+V2A | ||||||||||
| ReWaS ( Jeong et al., 2025 ) | – | – | – | 0.123 | – | 0.280 | 0.110 | 0.104 | 0.079 | 1.071 |
| AV-Quality | Text-Consistency | AV-Consistency | AV-Synchrony | |||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Model | VQ | AQ | TV-IB | TA-IB | CLIP | CLAP | AV-IB | AVHScore | JavisScore | DeSync |
| LTX-2.3 | 2.714 | 5.252 | 0.279 | 0.178 | 0.308 | 0.429 | 0.256 | 0.251 | 0.222 | 0.584 |
| LTX-2.3+GDPO | 2.713 | 5.577 | 0.280 | 0.181 | 0.313 | 0.439 | 0.251 | 0.249 | 0.221 | 0.563 |
| Cascade | 3.131 | 5.235 | 0.273 | 0.179 | 0.306 | 0.463 | 0.274 | 0.269 | 0.241 | 0.432 |
| Vanilla OPD | 3.174 | 5.342 | 0.275 | 0.180 | 0.309 | 0.433 | 0.256 | 0.250 | 0.226 | 0.583 |
| Ours OmniOPD | 3.347 | 5.449 | 0.275 | 0.192 | 0.310 | 0.447 | 0.262 | 0.258 | 0.232 | 0.457 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| No. | Prompt |
|---|---|
| 1 | ”A fixed close-up shows two domestic cats sitting opposite each other on a windowsill in soft afternoon light. One is an orange tabby and the other is a gray British Shorthair. The camera remains perfectly still while both cats maintain stable seated postures. The orange cat looks toward the gray cat and produces a gentle meow. After a brief pause, the gray cat slowly blinks and answers with a shorter, lower-pitched meow. Only subtle head, ear, and whisker movements occur. The audio contains the two clearly separated cat vocalizations with almost no other environmental sound.” |
| 2 | ”Amidst twisted metal and rubble under a hazy, sunlit sky, five figures stand arranged in a loose semi-circle. Strong backlighting from the upper right casts dramatic golden highlights across their forms, while a slightly low camera angle looks up at the group, emphasizing their presence. On the far left, a man in a red-and-gold suit with glowing blue accents stands next to a younger man in a red-and-blue suit. To their right, a muscular, bald man with reddish-brown skin and tattoos faces a woman in black leather armor with antennae-like headgear, while on the far right, a man in a brown jacket holds a futuristic weapon at his side. The man in the red-and-gold suit turns his body slightly to look directly at the resolute-looking, muscular man opposite him and says quietly, “I can’t shake the feeling that this isn’t over yet.” In response, a faint smile appears on the muscular man’s face, and he replies in a calm, steady voice, “As long as we’re still standing together, we can get through anything.” The other three figures remain standing, their gazes shifting subtly as they listen in silence to the exchange. The audio shows a quiet background ambience featuring the sound of soft wind moving through the ruins, punctuated by the faint, intermittent clinking of metal debris. The dialogue is clear, with the first man speaking quietly and the second replying in a calm, steady tone.” |
| 3 | ”At night beside a mountain lake, four friends sit around a small campfire beneath a clear sky filled with stars. One person holds an acoustic guitar while the others sit on logs wrapped in blankets. The camera begins with a close-up of burning wood and glowing embers before slowly pulling back to reveal the group and the dark lake behind them. The guitarist begins playing a gentle chord progression, and after several measures, a woman quietly starts singing. The others gradually join her during the chorus. The audio features detailed fire crackles in the foreground, occasional pieces of wood collapsing into the flames, and a naturally recorded acoustic guitar. The first singing voice is clear and close, while additional voices gradually create a small group harmony. Faint insects and distant water sounds remain audible behind the music.” |
| 4 | ”SpongeBob and Patrick are sitting on a rock underwater, talking deeply. Patrick looks at Patrick and says clearly: ”Knowledge cannot replace friendship, SpongeBob!” with an emotional expression. SpongeBob smiles warmly and responds: ”I agree, but I also really want that Krabby Patty!” The scene features professional lip-syncing and facial muscle movements. Cinematic 2D animation, high-detail CGI. Soft underwater lighting, floating sea particles, 8k resolution, photorealistic textures for sponge and starfish skin.” |
Audio-Video Generation
Audio-video generation creates or edits synchronized sound and moving images from text, media, or other multimodal inputs.
See all 162 papers