Everyone’s waiting on Seedance 2.5 news, and there’s a little of that below — but this one’s mostly a roundup of everything else going on, starting with ByteDance’s new image model. Full breakdown’s in the video.
Seedream 5.0 Pro: the layer trick
Seedream 5.0 Pro is another entry in the thinking-image-model category, with native text support, claimed skill at dense infographic-style layouts, and up to 10 reference images — which I think beats Nano Banana. The genuinely new ideas are interactive precision editing (lasso and point selection, sketch rendering) and layer separation: generating within layers, which in principle lets you lock in your compositions. That was the most exciting thing about it on paper. It is not quite fully baked in practice.
As a standalone generator, through the standard tests:
- Man in a blue business suit: native 4K, and at first glance we got what we asked for — nice city detail, not much to complain about. Then you punch in, and it’s the return of the blobby face people. Even our man himself isn’t full blob, but he’s on his way to the transformation. A re-roll cleverly faced most people away from the camera, but the few facing us: blobby. There’s also some trademark territory on the signage — Starbucks, CVS, and something that might be a Nando’s.
- The UK living room detail test (Fofr’s prompt): honestly pretty good. Blue walls, red carpet, white wicker furniture, the vintage computer — the identifiable landmarks are all there.
- References: two inputs — a location, plus the Oni stuntmen from Dragon Blue — came out looking like Blade Runner 2049 directed by Quentin Tarantino, which is a movie I’d watch. But Flamethrower Girl in her jungle adventure outfit exposed real weaknesses: identity drift in the face, one pose that looked genuinely painful, and one shoulder-mounted flamethrower unit nobody ordered (though now I kind of want it). Tuned prompts got outputs that are technically fine — consistent lighting, correct flamethrower, five fingers, the tattoos — and just kind of bland.
- Where it shines: medium close-ups, genre and style work (it has a very good handle on Giallo — full Argento/Suspiria vibes), and especially environments. A Wong-Kar-wai-by-way-of-cyberpunk frame came out legitimately cool.
Then there’s the layer editor, which lives on Dreamina (I didn’t see it in the API). The marketing implies you click a button and your image explodes into editable layers. That is not it at all. It’s a fairly frustrating, brush-based experience: you can arrange, flip, and delete layers, but for some reason the inpainting models on offer are Image 3 and 3.1, not 5.0 Pro — so a new layer has no contextual awareness of your base image. Remove-background does a strange flipped-masking thing where the checkerboard shows up on the foreground object. And my final compositing test — dropping in the Flamethrower Girl image from the Muse video and asking for a relight — simply refused to generate all day. I did get refunded three credits.
The verdict from the video: as it stands, it’s kind of a mess — but Seedream 1.0 wasn’t very good either, and ByteDance is dominating everywhere else. If they get the editor humming with 5.0 Pro doing the inpainting and compositing, this becomes a very interesting addition next to GPT Image 2 and Nano Banana.
Seedance 2.5 watch: delayed again, leaks keep improving
The release has slipped again — to at least July 20 — but new samples leaked in the meantime, continuing the run. These focus on one-shot camera output:
- The train station reunion: a 30-second single shot with very nice camera movement and a very Seedance look. (“You’re really here. You’re actually here.” — how many times did this guy stand her up for that reaction? I’m guessing at least three.) Some compression crunch and a slightly-off frame-rate feel, possibly a 60fps upscale.
- The kitchen: dynamic acting, great early camera movement, a delicious-looking scallop, and a chef who nearly loses an eyebrow. I just finished The Bear, so this one’s close to my heart.
- “We made it”: one of those scenes you can’t place — end of the movie, or the beginning before a flashback? Either way it’s starring Andrew Garfield.
Mid-edit, ByteDance also dropped an official short film made with Seedance 2.0, built around Michael Owen’s 1998 World Cup goal against Argentina. It’s a nice little short — though if you’re paying close attention, the physics aren’t perfect and there are a couple of small morphs. That’s not a knock; these are the best video models in the world right now, and the story is what matters. It does have a bit of that GPT-Image grain to it, though.
LingBot-World 2.0: a real-time world model you can actually run
From Reactor, with weights on Hugging Face: a real-time, open-source world model in the Genie 3 mold, with text-triggered actions, events, and weather, plus an agentic “brain and cerebellum” system that continuously proposes new events as you play. I ran the Siege of the Keep preset — WASD movement, triggerable flaming volleys and rally cries — and the interesting part is how it handles the edges of what it knows. Where Genie 3 goes blobby and decoherent when you push past its seed, LingBot holds together longer, and when you do hit the limit it just respawns you to another location. That’s a smart containment trick. Beyond the pseudo-game angle, treat that castle as a virtual set you can fly a camera through and this gets interesting for filmmaking. There’s an online sandbox if you don’t want the local install.
CARA-4: the voices are getting faces
Anam’s CARA-4 does real-time AI avatars, and the two-avatars-talking demo is pretty good. With voice models as good as they are now, we’re close to most AI voices having a face. Real-time Flamethrower Girl co-hosting the show? Probably going to happen — maybe even this year.
One admin note: I’m doing a free one-hour live session with Teachable’s AI Academy (they sponsored the video) — including how the recurring-character roster (blue-suit man, the FBI agent, Flamethrower Girl) works as a stress-test system for new models. Sign-up link is in the video description.