Researchers have unveiled Kandinsky 6.0 Video, a family of foundation diffusion models designed for synchronized text-to-audio-video generation. According to the paper, the family includes two variants: Kandinsky 6.0 Video Lite with 3 billion parameters and Kandinsky 6.0 Video Pro with 29 billion parameters.
Both models produce 5-second video clips paired with synchronized 44 kHz audio, including lip-sync, and support two modes: text-to-audio-video (T2AV) and image-to-audio-video (I2AV). A built-in super-resolution model raises output resolution to Full-HD (1920×1080).
Architecturally, the system builds on the video generation capabilities of Kandinsky 5.0 and introduces a dual-stream CrossDiT architecture that links a pretrained video stream with a newly trained audio stream through bidirectional cross-attention, aiming for temporal and semantic alignment. The training pipeline uses a continuous pretraining strategy: the audio stream is first trained from scratch on large-scale audio corpora, after which both streams are trained jointly on paired audio-video data while preserving unimodal fidelity. This is followed by supervised fine-tuning, reinforcement-learning-based post-training, and distillation.
In side-by-side human evaluation, Kandinsky 6.0 Video Pro reportedly outperforms its predecessor, Kandinsky 5.0 Video Pro, and remains competitive with leading audio-video generation models, particularly in speech quality. The paper also notes that RL post-training reduces Pro's speech word error rate (WER) by 47%, and that on VABench the Pro model achieves the best results among LTX 2.5, Kandinsky Lite, and Kandinsky Pro across speech quality, audio aesthetics, AV alignment, lip-sync, desynchronization, and visual realism.
A separate component, Kandinsky 6.0 Video Super-Resolution, is described as a 1.4B-parameter SR diffusion transformer supporting ×2, ×2.25, and ×4 video upscaling. It operates in KVAE latent space by encoding the input video, upscaling its latents, refining them with tiled diffusion, and decoding the high-resolution video. It is available as a flow-matching model and as a π-Flow DX distilled model requiring just 2 model evaluations per tile.
To accelerate open research and deployment in multimedia generation, the authors state that code, model checkpoints, and Diffusers integration are released under the MIT license.