Skip to content
AI Field Notes

Weekly roundup

AI model drops: SwitchX 3.0 and Seedream 5.5

By Nigel Camp ·

A good week if you shoot real plates and finish with AI. Beeble and ByteDance both put out video tools worth an afternoon, ByteDance Seed sorted out the colour drift that creeps into repeated image edits, and ElevenLabs gave editors a better transcriber. The language models, plus a few voice and image releases we didn't have room for, are listed at the bottom.

VideoBeeble11 Nov 2026

SwitchX 3.0: a new world behind your handheld plate

Beeble's video-to-video model now runs on MiniMax H3, retrained on Beeble's own data, and it's built for one job: replacing the world around a live-action shot. You can give it up to four references, say a wide of the set, a prop and a storyboard frame, and it pulls them into one shot. Backgrounds move far more than in 2.0, so traffic, crowds, fire and snow, and Beeble says the new world stays locked to the camera even when you're handheld, orbiting or running with your subject. It's in the Beeble app and Canvas now, with output up to 4K.

VideoByteDance (Volcano Engine)12 Nov 2026

Seedance 2.5 flash: the same tricks, twice as quick

Volcano Engine's faster tier of Seedance 2.5 keeps most of what the flagship does: single clips up to 30 seconds, and up to 10 images, 5 videos and 5 audio clips as references in one go. The catch is resolution, 480p or 720p only. ByteDance says it generates twice as fast and costs less, so it's the one for blocking out ideas and batch runs before you spend on the full model. The listing is on the Chinese Ark console for now; our Model drops line has the link.

ImageByteDance Seed9 Nov 2026

Seedream 5.5: edit the same frame 100 times without drift

Photorealistic sample stills from ByteDance Seed's Seedream 5.5 announcement, including a woman and child on grassland, a restaurant, a cat, a fisherman and a subway

ByteDance Seed trained this image model to leave untouched areas alone at the pixel level, so colour doesn't shift and detail doesn't fall apart as you keep editing one image. Their demo runs 100 edits in a row on the same person. It also makes native RGBA images with full and partial transparency, which means cut-out elements for comps and titles with no background removal or edge clean-up. Skin, lighting and small text look more natural too, and Seed says usability in its film and TV tests was more than double Seedream 5.0 Pro's.

AudioTranscriptElevenLabs10 Nov 2026

Scribe v3: transcripts that get the speakers right

ElevenLabs cover card reading Scribe v3 in white on a soft blue and white blur

Not a generator, but anyone logging interviews will feel it. ElevenLabs says its new speech-to-text model makes 27% fewer word errors than Scribe v2 across 102 languages, and 17% fewer speaker-label mistakes on recordings with three or more people. You can feed it up to 1,000 key terms, handy for character names and odd place names, and choose whether numbers come back as digits or words. It's now the default in the ElevenLabs app, which sits on our ElevenLabs page.

Also dropped

Every claim here comes from the maker's own page, linked above. Licence tags are what the maker's licence or terms said on the day we filed, not legal advice, so check the linked terms before client work. The full day-by-day list is on Model drops, and older roundups are in the weekly archive.