SkyReels-V2 是什么?
SkyReels-V2 是由昆侖萬(wàn)維開源的一款視頻生成大模型,首創(chuàng)“擴(kuò)散強(qiáng)迫(Diffusion-forcing)”訓(xùn)練框架,具備生成高保真、連貫、長(zhǎng)時(shí)長(zhǎng)(30s 以上)的影視級(jí)視頻的能力。模型融合了多模態(tài)大語(yǔ)言模型(MLLM)、強(qiáng)化學(xué)習(xí)、視頻理解與擴(kuò)散優(yōu)化等多項(xiàng)技術(shù),適用于故事生成、圖像到視頻(I2V)、運(yùn)鏡控制和多元素一致性生成等豐富場(chǎng)景。
SkyReels-V2 的主要特點(diǎn)
- 無(wú)限時(shí)長(zhǎng)生成能力:基于擴(kuò)散強(qiáng)迫(Diffusion-forcing)技術(shù),實(shí)現(xiàn)連貫、穩(wěn)定、長(zhǎng)達(dá)數(shù)十秒甚至理論無(wú)限時(shí)長(zhǎng)的視頻輸出。
- 影視級(jí)視頻質(zhì)量:在清晰度、色彩、構(gòu)圖和鏡頭語(yǔ)言等方面接近真實(shí)影視內(nèi)容,適用于專業(yè)創(chuàng)作場(chǎng)景。
- 精準(zhǔn)指令遵循與一致性控制:通過(guò) SkyCaptioner-V1 理解模型和強(qiáng)化學(xué)習(xí)機(jī)制,提升提示詞響應(yīng)能力與跨幀一致性。
- 高質(zhì)量運(yùn)動(dòng)生成:解決傳統(tǒng)模型運(yùn)動(dòng)扭曲問(wèn)題,生成流暢、自然、物理合理的動(dòng)作與鏡頭變換。
- 多種生成范式支持:支持文本生成視頻(T2V)、圖像到視頻(I2V)、多元素合成(E2V)、運(yùn)鏡模擬等復(fù)雜場(chǎng)景。
- 全開源與低訓(xùn)練成本:核心模型及子模塊(包括 A2 和 Captioner)均已開源,支持中低算力高效訓(xùn)練與部署。
SkyReels-V2 的視頻demo
Prompt:A graceful white swan with a curved neck and delicate feathers swimming in a serene lake at dawn, its reflection perfectly mirrored in the still water as mist rises from the surface, with the swan occasionally dipping its head into the water to feed.
Prompt:A graceful sea turtle glides through the ocean currents, its powerful flippers moving rhythmically as it swims. The turtle's smooth, greenish-brown shell catches glimpses of sunlight filtering through the water, creating a mesmerizing play of light and shadow. Schools of small fish dart around the turtle, adding life to the underwater scene. The background showcases vibrant coral reefs and swaying seaweed, creating a rich, dynamic marine environment. Wide-shot underwater scene with gentle camera movement following the turtle's path.
Prompt:A jellyfish floats in deep blue waters, its translucent bell rhythmically contracting and expanding. Its tentacles drift like ribbons in the water. As the jellyfish gracefully swims, we see it passing through a colorful coral reef, where the orange, pink, and purple corals create a striking contrast against the blue-white hues of the jellyfish. The jellyfish continues forward, entering a forest of seaweed, where slender green fronds sway gently with the ocean currents. Finally, the jellyfish drifts toward an open sandy area, scattered with a few glistening starfish and shells. Throughout the journey, sunlight penetrates from the surface, creating beams that cut through the water, adding shifting patterns of light and shadow to this dreamlike underwater world.
SkyReels-V2 的應(yīng)用場(chǎng)景
- AI 故事視頻生成:通過(guò)連續(xù)文本提示生成具有敘事連貫性的長(zhǎng)鏡頭影視內(nèi)容。
- 圖像到視頻合成(I2V):支持以圖像為條件生成高一致性視頻,適用于素材動(dòng)化。
- 導(dǎo)演級(jí)運(yùn)鏡模擬:精細(xì)控制攝像機(jī)視角與運(yùn)動(dòng)軌跡,適用于短劇、廣告腳本制作。
- 多角色合成視頻(E2V):將人物、物體、背景元素統(tǒng)一協(xié)調(diào)生成視頻,廣泛用于虛擬電商、創(chuàng)意內(nèi)容制作。
如何使用SkyReels-V2 ?
- 在線體驗(yàn):https://www.skyreels.ai/
- GitHub倉(cāng)庫(kù):https://github.com/SkyworkAI/SkyReels-V2
- 模型地址:https://huggingface.co/collections/Skywork/skyreels-v2-6801b1b93df627d441d0d0d9
- 論文鏈接:https://arxiv.org/abs/2504.13074
