`SANA-WM`, a 2.6B open-source world model for 1-minute 720p video

A single image plus a 6-DoF camera path can produce controlled long video on one GPU. Useful signal for product mockups and scene previews, but still closer to R&D than a plug-and-play SaaS feature.

[ KEY POINTS ]

Input is one image + 6-DoF camera trajectory, so the value is controlled scene movement rather than text-to-video prompting.
Hybrid Linear Diffusion Transformer mixes frame-level Gated DeltaNet with periodic softmax to keep long rollouts coherent.
Single-GPU 720p 1-minute generation lowers experiment cost, but integration still depends on weights, license, and inference setup.

Originalnews.hada.io/topic?id=29572Read original →

// related

#0001
#0001Generative media GeekNews5 days ago
`Supertonic 3` launches ultra-light on-device TTS with 31 languages and emotion tags
60radar
Supertonic 3On-device TTS engine — multilingual with emotion tags
A small-footprint TTS now handles expressive cues like laugh and scream, while improving pronunciation and voice cloning. Strong fit for offline narration, in-app voice UX, and cost-sensitive shipping.
- Supports 31 languages including Korean, which lowers the friction for multilingual voice features without a cloud dependency.
- Adds 10 emotion tags such as laugh, breath, and scream, so scripted dialogue can sound less flat with simple text markup.
- Pronunciation accuracy improved, and repetition or omission failures were reduced; this matters more than flashy demos in production TTS.
- Voice cloning quality also improved, making it more usable for character voices, guided audio, or branded app narration on-device.
Source: news.hada.io/topic?id=29522Read original →
FIG-0011:1
60radar
FIG-0011:1
#0002
#0002Generative media together_ai6 days ago
`Violin`: open-source AI video translation stack
50radar
ViolinOpen-source video translation tool — unifies `ASR`, translation, and `TTS`
An end-to-end stack for multilingual video localization, not just a demo. Useful when you need ASR + translation + TTS without stitching vendors together.
- Combines speech recognition, LLM translation, and text-to-speech in one flow, reducing glue code for video localization.
- Being open-source matters more than model novelty here: you can inspect, swap, and self-host parts of the pipeline.
- Best fit is repurposing existing video assets across languages; less compelling if you only need subtitles or basic dubbing.
Source: www.together.ai/blog/violin-open-source-translation-skilRead original →
FIG-0021:1
50radar
FIG-0021:1
#0003
#0003Generative media together_ailast week
`Voice Finder`: search and audition **600+** TTS voices faster
50radar
Voice FinderTTS voice search tool — matches by prompt or audio
Voice selection moves from manual browsing to prompt or reference-audio search. Useful if your app ships spoken UX, though it matters more for workflow speed than model quality.
- Search spans 600+ voices across Together AI TTS models, cutting the time spent comparing presets by hand.
- Natural-language prompts let you filter by tone or style, which fits rapid prototyping before wiring custom voice settings.
- Audio-sample matching is the more practical hook: upload a reference clip and shortlist similar voices faster.
- This is a discovery layer, not a new speech model. Shipping impact depends on whether voice choice is your current bottleneck.
Source: www.together.ai/blog/introducing-voice-finder-a-new-toolRead original →
FIG-0031:1
50radar
FIG-0031:1