Dreamina Platform Update: Canvas Agent, Elements, and Seedance 2.5 Preview
Not Seedance 3.0 — this is a Dreamina platform update. The canvas and its AI agent got a real overhaul, Elements finally lets characters (and voices) follow you into any prompt, and Seedance 2.5 Preview, still in preview, drafts at 480p and then re-renders the same shot at 1080. I ran three tests, including a Flamethrower Girl stress test, and found where I'd still take manual control.
First, so nobody gets excited for the wrong reason: this is not Seedance 3.0, and it’s not a new model. It’s a big Dreamina platform update. Dreamina is the western home for Seedance 2.0 and 2.5, and this is, to be fair, pretty long overdue. The canvas got an overhaul with an AI agent on top, Elements finally exists, there are new Skills, and the one I think everyone will actually use is Seedance 2.5 Preview: a cheap draft mode that re-renders your exact shot at 1080.
I ran three tests through the new canvas; the short verdict is to let the agent organize, but prompt the shot yourself. The full rundown’s in the video.
The revamped canvas and the AI agent
Dreamina has finally introduced its own murder board. Admittedly, it’s not entirely new: the canvas has been around for a while, but it’s massively improved. I know node-based workflows aren’t everyone’s cup of tea, and you can still do it the old way.
The chatbot inside the canvas got revamped too. The old assistant, in my opinion, was actually pretty stupid. To their credit, Dreamina has upped its game: I wouldn’t call the new one perfect, but it’s drastically improved.
More models inside the canvas
You’re no longer limited to Dreamina’s own models in there.
| Type | What’s in the canvas | Note |
|---|---|---|
| Image | GPT Image 2, Nano Banana Pro, plus the Seedream models | No sign of GPT Image 2.5 yet, which would be a nice addition |
| Audio | Seed Audio 1.0 | |
| Video | All the Seedance versions, plus MiniMax H3 |
H3 is the one that really surprised me — ByteDance hosting the open-source model that fal turned into H3 Max a few weeks ago.
Test 1: the PI office, recast
I took our PI office and Tuesday Sinclair, and since I usually have Malloy in that office, I recast him as James from Death Walks Into a Bar. Three nodes going in, then I gave the agent the transcript of the first version of the scene and said, let’s redo it with these characters. Admittedly, there’s nothing here that we couldn’t do in a non-canvas version, so I pushed it harder.
Test 2: Flamethrower Girl, and why she’s called Lyra
I gave the agent the material from our Flamethrower Girl / jungle girl stress tests from the Seedance 2.0 days. Without being asked, it created transparent PNGs of all our elements and characters, then, after we talked through what I wanted, wrote its own brief, an asset list, and a plan for the generation.
Side note: the agent named Flamethrower Girl Lyra. That’s not her name. It’s the name of the barbarian from The Bridge, which wasn’t generated on Dreamina, so I don’t know where it pulled that from. Nothing conspiratorial. Somewhere in past generations I probably used the name, but I did think it was interesting.
Its first plan was ten separate Seedance 2.5 generations instead of one 30-second generation. I wasn’t keeping an eye on it, and it went with 16:9. Overall, I wasn’t happy with the first three. Technically nothing is super wrong, but it feels a bit bland. At the ten-second mark, where we’re supposed to cut to the next shot, there’s a massive jump cut: the idol’s size completely changes, and as she runs, it vanishes. The third shot picks up with her mid-roll and the idol back, guardian chasing her. Not too much wrong with the back end of that, but it’s not going to cut together.
Test 3: feeding the agent the storyboards
So I gave it the original storyboard breakdown for this stress test, generated way back in GPT Image 1. I told the agent not to use them as storyboards, but as inspiration for its own version. Its response: this is one of the best previs shot breakdowns it had ever seen. The Dreamina agent will gas you up.
What I thought was genuinely smart: once the storyboards were in, the shot planning and text on the board all auto-updated.
The generation is all there, but maybe not quite as kinetic as Seedance 2.0, which I’ve noticed a lot with 2.5. The glowing-eyes problem is back, and obviously stood out most, plus another weird jump cut.
Edit segment, and the snootiest possible fix for glowing eyes
The new edit-segment feature is pretty handy: scroll to the glowing eyes and prompt your way out of them. When I ran that edit, Seedance’s solution was to have her eyes closed the entire time. I’ll admit that was funny. The AI as the ultimate snooty waiter makes me laugh: “Oh, you wanted her eyes open and not glowing? Well, sir, that is not what you ordered.”
The verdict: let the agent organize, prompt the shot yourself
The canvas has significantly improved, and the agent does a great job keeping everything organized and pulling it together. But when it comes time to prompt your shot, take over manual control, or use your LLM of choice. I took all of this material and ran it through my own prompt, and I’m biased, but I do think that looks better. (My approach to those prompts is in the Seedance 2.5 prompt guide.)
Elements: characters and voices you can @ in any prompt
A quick quality-of-life feature: there’s now an Elements folder. It’s very simple, and Kling has had this for ages, but I can put Tuesday Sinclair in there, give her a custom voice, and then @ her in any prompt, across the canvas or standard generation. There are additional angles for her too. Not a big leap, but sorely overdue and very much appreciated.
PJ Ace’s Skyline Heist board
For a longer-form piece, friend of the channel PJ Ace recently released Skyline Heist, a chapter of his Nexus project: a steampunk-flavored heist film in the sky. No notes. The canvas workflow lives on Dreamina, and there’s a lot on that board worth stealing:
- Character sheets. His version is a pretty good format.
- A note to the animator, essentially the board itself. A little outdated: it says 2.5 is limited to 720, and it’s 1080 now.
- An acting system. Possibly outdated too, since Seedance now has a micro-acting skill, but still worth cribbing.
- The generated videos at the bottom, which show what he used: a character reference sheet and a location or backplate, then very detailed prompts for the action rather than a first frame.
Skills on the generate side
On the generate side, there are new Skills: an iconic film style, a film storyboard skill, cinematic long takes (which allows generations longer than 30 seconds; I don’t mess with it much, because it gets expensive), and a micro-expression prompt, which allows multiple emotional ranges within one prompt. There’s a whole bunch of others too. I kind of read them as pre-built markdown files, so they’re worth a look, or use them just by hitting them through chat.
Seedance 2.5 Preview: 480p drafts, then a 1080 re-render
This is the one I think is super exciting, and again, it’s in preview right now, so expect a release at some point soon. Whatever you prompt in Seedance 2.5 Preview generates at 480p, at a significantly lower price than 720 or 1080.
I gave it a quick whirl with channel fan favorite Sniper Girl. Admittedly it’s a weird generation, with a cross dissolve in the sniper scope, but I think it looks kind of cool. It’s 480p, so upscaling isn’t much of an option: take it to 1080 or 4K and it would probably look a bit of a mess.
What you can do is generate in full quality, and Seedance re-renders that clip at 1080. It’s not upscaling. It’s a full re-render. That lets you hone in on a prompt, rather than using one of the fast models, where the same prompt in a higher-level model isn’t guaranteed to give you the same output. This time, it is.
I’m not sure exactly when this releases, but it’s definitely past beta, so hopefully you’ll have access within a week or so.
BytePlus Dramagic and a possible ByteDance world model
I do think we’re on the verge of some pretty big ByteDance and Seedance updates. BytePlus just introduced Dramagic, which is more on the enterprise side: script analysis, storyboarding and video review, essentially all in one platform. It’s in beta, with a link to request access — request, not get. I’ve signed up, which doesn’t mean I’ll get it.
The other interesting thing is the rumors going around that ByteDance is nearing the release of its own world model. That would be very interesting after a loaded week for world models, and I’ll let you know as soon as I hear more.