CrunaStudio's First Full Demo — A Year of Building Industrialized AI Comic Drama

A year of work, and today is the first time I’m showing CrunaStudio in full.

Not a demo, not a promo reel. I actually drop a novel in as raw txt and run it through to an 80-episode script, character assets, and copyright-cleared music — one pipeline, start to finish, on camera.

Here are the key stops along that pipeline.

1. Why not build a standalone tool

The AI comic drama space is crawling with tools — text-to-image, text-to-video, one-click script splitting, AI voiceover. Every single capability exists somewhere. But anyone who has actually made an 80-episode series knows standalone tools never add up to an assembly line.

You get to episode 47 and your lead’s face suddenly changes. You swap the BGM and find out the previous track got pulled by the rights holder. Bring in three editors to collaborate and file versions start flying in every direction.

CrunaStudio isn’t another AI tool. It’s a platform built so a team can actually collaborate its way through all 80 episodes.

2. From a 19,000-word novel to a script structure

The first step I demo in the video: drop in the full text of a 19,000-word novel and let the agents break the script out of it.

Not naive chunking. It cuts to short-drama rhythm — every episode gets 3 minutes, 5 plot beats, 1 hook, 1 reversal. The agents finish and output an 80-episode structure table, each episode carrying its characters, settings, and key lines.

I page through it live: the hook in episode 1, the midpoint reversal at episode 30, the landing in episode 80 — they line up structurally.

3. Character assets: from consistency to reuse

The biggest pain in long-form is consistency.

Can your lead’s face hold steady across different scenes? Change the hairstyle, the outfit, the lighting — is it still the same person?

We turn characters into “assets” — define once, call anywhere in the series. In the video I swap three scenes and two outfits, and the lead’s face doesn’t move. This is the solution that fell out of every consistency trap we hit over the past year.

4. Clearing music rights

Getting a short drama pulled offline by a rights holder is routine in this industry.

CrunaStudio has a built-in copyright screening and replacement flow for music. You pick the emotional register (tense, mournful, hype) and it hands you not just the track but a clean provenance record, ready for commercial use.

In the demo I pick one track out of 30-plus candidates to fit a scene, each of them carrying its own provenance chain.

5. Wiring in Seedance 2.0: multiple shots in one pass

The last stretch is generation.

With Seedance 2.0 wired in, a shot is no longer one prompt producing one second of video. A single storyboard panel can generate a multi-shot sequence directly — consistent subject, coherent camera movement. I had it run a continuous three-shot fight sequence live.

Wrapping up

If you run an AI comic drama team, a studio taking production contracts, or you just want to understand industrialized AI video, this video will be worth your 30 minutes.

The rest — let’s talk in the comments.

Source: CrunaStudio live demo, 2026.05.07 (30 minutes, one continuous take) · ASR model gemini-2.5-pro (180-second chunked parallel) · full text of the original recording

[00:00] Hello, hi everyone. Uh, it’s been a long time since I posted anything, and honestly I haven’t really wanted to. My WeChat Channels stuff used to all be made with AI, right. And then the scripts were AI-written too, so posting — it just felt like there was no, no, no point to it, that’s how it felt. So I decided that starting today, all my WeChat Channels content is going to be recorded this way, one continuous take. And I’ll walk you through the things I’ve been doing with AI lately.

[00:27] So the first thing I want to talk about — something I’ve wanted to share with you for a very, very long time — is this thing I built, CrunaStudio. It’s a platform for making AI comic drama. Enough talk, let me just show you directly. So once you log into CrunaStudio, it’s a platform where a team collaborates on comic drama. You can see — top left, you can see the projects currently running, ah, the names of the running projects, and you click to

[00:53] switch between them. Right now it’s already open to some of our sister companies’ teams — you top up your balance and you can start doing comic drama development on it. And those avatars in the top right, those are all people who take part in the comic drama process, whether that’s the screenwriter, or the art people, or the AI, the AIGC video director. They all log in with WeChat to collaborate in this system.

[01:20] So the whole system — you can see it has three columns. The leftmost column is a table of contents, it’s a table of contents. Ah, and you click to jump to different sections. The middle one, you can just think of it as a Feishu doc — everyone can co-edit in there, meaning I, I can see other people editing in it, and I can edit the doc directly myself. It’s very WYSIWYG, very

[01:45] beginner-friendly, I’d say. Now the biggest difference is the right sidebar. The right sidebar is basically like a WeChat group — everyone creating in here can chat in it. But the difference is, there’s actually a character in there called Cruna, and that’s our comic drama agent. It can help the whole comic drama team get a lot of different jobs done during production.

[02:10] So next let me actually show you. Let’s go through the sections one by one. The first section is called Concept Bible. In here, like when we’re making a comic drama, the core concept, the world-building, the story synopsis — all of it gets done in here, including the character profiles and backstories, which main characters we’re actually using, and what tra-, tra-, traits those characters have.

[02:35] Then through the whole script-writing process, the AI assists the screenwriter too. You can see in here — this is an 80-episode series, and the script, the script in here is written in a fairly standardized format. Now while the screenwriter is writing, it might not be in that standardized form, so the AI cleans it up into the standard form — including the dialogue in it, including the sce-,

[03:00] scenes, including the characters’ actions and the background music. Once the script is done, you move into an asset generation stage. Right now the art style supports anime style, rendered style, and live-action style — once you pick one, all the material after that comes out in that style. And in here it also breaks out into a character design section, where the system’s AI autonomously identifies every character in the script — any character who appears on screen

[03:30] gets this automated character design. Then we can have the system generate the images directly, and the image it produces we can click to enlarge — it does a kind of gacha roll. Uh, after rolling for a character, if we’re not happy with that character, we can also fine-tune it, tweak the prompt right here, or we can just upload a character we’ve already designed — either way works.

[03:58] So after the characters are designed, we also have prop design in here. You can see the system in here, the whole system is globally consistent and cross-linked. Just now we saw what the lead looks like — well, you can also see his employee badge, and his badge references his look, that’s how the badge was made. Completely consistent. Then all the props — once every prop in the series has been generated, there’s also styling, and

[04:26] uh, the main character across different episodes — you can see the episode numbers he shows up in are all listed here. Across different epi-, episodes, his look isn’t quite the same. And each look gets a multi-angle design in here — front, side, back, then some close-ups on details. All of it fits the, uh, style needed for the episode it belongs to. Now let’s scroll down further — there’s also scene design, like some dense-forest scenes, each one gets a

[04:55] wide establishing shot, an image that sets up the big scene where the whole story happens. And then there are also small, small scenes — for example the male and female leads up in a big tree, what exactly does that fork in the branches look like, and then a multi-angle version of that scene. Once the scenes are generated, there’s also VFX design, for example the system pop-ups, plus things like spell arrays or a lone sword, plus some

[05:24] of these vine-wrapping effects — all of these, it generates the VFX animation frame by frame. Uh, one place it’s different from other systems is the voice generation. To chase character consistency, it analyzes and assigns a voice to every character in the show, including the random extras — for each of them there’s an AI judging what kind of character it should be. Right now, look, this lead, Lu Yun,

[05:52] he might be sharp plus sunny plus smooth-talking — there’s actually a recipe behind his voice. And it generates four voices for us to choose from. At this point the director can audition each option. Once you lock in the one you want, all the dialogue generated for the whole show comes out in that voice — and there’s an emotion component in here too. “Kitty, I only ever sell goods. If you want me to sell my life, that’s a different price.”

[06:19] “Are you sure you need it right now?” Right, see, and there’s the female lead’s voice too. This, this is a dragon’s roar. It’s the sound of inner qi and magic surging. Right, once the voices are set, there’s also the soundtrack. In this whole system — one of the bigger headaches when people make comic drama is, how do I keep my BGM from infringing copyright. So after this whole system finishes reading the script, for something like this 80-episode show, it parsed out roughly 40,

[06:46] uh, 40 instrumental BGM tracks for the sound-, soundtrack. It’s based on the kinds of beats that might come up in the plot — like this romantic, heart-fluttering style, then falling for something delicious, and there are also some battle-scene, uh, BGM cues. All of them you can audition right here and download to your, uh,

[07:11] local machine, and then use when you’re editing in post. If there’s something you’re not happy with or want to adjust, you can regenerate right here or tweak the prompt. Then in the actual production process, the AI breaks each episode out into different shot groups.

[07:42] So a shot group like this one might be generated as multiple shots, because right now the whole system is wired into the Seedance 2.0 model, and it’s capable of generating multiple shots within 15 seconds. So you can see, episode 1, shot 1, it automatically generates this card — it’s a six-second shot. And the prompt is already written for you. And the reference images used in here, what kind of reference images — for example it uses this

[08:12] dense-forest reference, or the office reference, and it pulls the matching reference images straight out of the asset library. Including the lead’s look, and the props that will show up. And right here you can, uh, audition the lines that appear in this shot — the lines just get generated right there. Using the lead’s voice. “Yes, yes, yes, cat treat sticks, the hot-selling one. We have to get the price down for Double 11.”

[08:38] And at this point you can actually edit — I can take these lines, you can see there’s an emotion adjustment in here, and you can also change the actual lines, and change the speaking rate. Once it’s generated, this audio becomes the reference for this shot, the reference audio gets pushed straight into Seedance to generate the video.

[09:03] So let me, let me show you the material we generated earlier — you can preview it right here. “How am I up in a tree? Wasn’t I just working late at the office?” Right, you can preview it right here. And all of this, everyone can collaborate on. And you can see, during this process the work gets divided up — who’s producing which episode, it’s all labeled here.

[09:31] Once production is done, there’s a final-cut review stage — people upload the finished cut to this section, and the director gives feedback here, meaning the director watches the video here.

[09:59] Right, and when he pauses at this exact second, he can write an annotation right here saying what needs to change. Then the editor can make the change directly here, and can also add new shots here, and adjust the emotion of the dialogue here. So the whole process, all of it, gets done on this thing. And you can see, the AI on the right, during this process, can

[10:24] during the writing process it can review the script, and it can do an assessment of the script — where exactly is it weak — so you can adjust in time. And during production, if you spot continuity errors between scenes, or a missing character, you can also ask the AI right here, like, does this prop appear in this episode, or what does this character’s hair and makeup look like in this episode.

[10:55] Let’s also create a new project and take a look at the AI’s whole generation process. New project. Say what we’ve got isn’t a script — what we’ve got is a novel, like, say, the novel Demi-Gods and Semi-Devils. I’m going to drop chapter one of Demi-Gods and Semi-Devils straight in and hand it to the agent.

[11:23] You can see it received 19,000 words of material, and it can generate a complete concept bible directly, and here’s how many credits that costs. I can check “don’t remind me again,” and confirm to go into — oh, look, it says it received my work. And then it asks: is this a novel, adapt it into a short drama for me, or is this a concept bible, tidy it up for me. This is obviously a novel.

[11:49] Then it asks whether to adapt following the original story, or whether it can take liberties. Let’s go with adapting the original story first. Now through this whole process, ah, because short drama moves at a faster pace, when we actually run this for real we have the AI rework things directly on top of the original according to what we need. You can see it’s already started building the adaptation plan.

[12:20] It says here’s the adaptation direction: this material is a period light comedy, and it keeps three core threads. There’s Duan Yu with his against-type character setup, and Zhong Ling doing such and such. It extracts the essential content of the novel. How much there is, uh, how many pieces of essential content there are, and then also the suggested adjustments. See, we’re making short drama, so here it suggests cutting the backstory setup about the Wuliang Sword sect’s East-West duel, and then some parts in the novel

[12:46] that maybe aren’t fast-paced and gripping enough — it suggests cutting all of those. Then the episode planning: for the content of the original, ah, estimated total of eight episodes, 60 to 120 seconds each. And it’s already written out a synopsis for every single episode. Including how to handle the characters, right. Which, which characters to keep, which characters can reasonably be trimmed.

[13:12] Mm, so the adaptation plan has already been talked out, and I can give it some feedback on top of that — and if I’ve got no notes at all, I can go straight to generating the concept bible. You can see the agent is running at top speed the whole way. Here’s the concept bible: core concept — Duan Yu, the Dali prince who never learned martial arts, stumbles into the martial world and, with nothing but his mouth and an eagle-raising girl, turns the martial world upside down. Ah, that’s what you’d call the one-line concept.

[13:44] Then our whole project setup, story synopsis, core mechanics, main locations — all of it gets listed out automatically. And of course we can also go right here, because this is just a document, so we can just, uh, the series’ planner or the screenwriter can edit right in here. It’s a mechanism where the AI, the AI and the human team collaborate together.

[14:15] Look, all the character backstories have been worked out automatically too. Mm, and once it’s done, it actually runs an audit on itself — in case anything got missed, because we know the context of this novel

[14:40] is a whole lot of words, and for the AI’s atten-, for its attention span, that’s a pretty big challenge. Now that it’s finished organizing, all the characters are out too, and next we’re going to move into the script-writing stage. Since I’m short on time, I’ll have it write the scripts for the first five episodes. The scripts come out in one pass too — or rather,

[15:05] normally, because of the context limits, we write five episodes at a time or one episode at a time — meaning the screenwriter still has a human role to play in this process. It first works out the global outline, then starts writing those first five episodes. Let’s, let’s just write episode one. The compute this burns is pretty substantial.

[15:38] Mm, it’s finished working it out, and now it’s starting to plan the script structure for episodes one through five.

[16:04] Right, a system like this, mm, hugely raises the efficiency of the whole team’s collaboration. And in this process, if you’ve already, if you’ve already got a script done outside the system, you can also import the script directly. The AI can tell right away — hey, is what you put in a novel or a script. If it’s a script, it’ll import the script episode by episode, and during the import we can actually choose whether to

[16:29] keep the script’s original text, or do some cleanup on the script’s formatting. You know how, the thing is, in the screenwriting industry, there isn’t currently a domestic standard for writing a script the way Hollywood has one, so — you can see it’s already started writing episode one, it writes one episode at a time. It’s, uh, very clear: the white triangle, what follows it is a

[16:54] description of the shot, and the black triangle is a description of the character’s action, and the music symbol is the sound-effects part — including which location the scene is in, which day it’s on, whether it’s day or night, all of it laid out pretty clearly. And on the left, you can also see how much length this, this episode actually took,

[17:19] how many words it took to finish. If there’s anything you want to talk about, type it in the comments — I mean, I actually

[17:46] I’m going to try to post daily, and share with you the interesting things I do day to day, or some of the lessons I’ve picked up. Okay, now it’s writing episode two. Honestly this production process is pretty hard, because

[18:13] this AI has to be able to, first, handle this much context, and second, it has to understand how to communicate with a lot of different people — including what structure all the documents get stored in, how, how to take this, mm, content and put it in a — because when the AI goes to, say, produce assets, there’s no way it can shove all eighty-something episodes’ worth of content in there.

[18:41] How it chunks that content, keeps it organized, and then loads it into context on demand — there’s actually a huge amount of thinking behind that. Now across the industry, everybody’s building similar platforms. There’s a term I, I use a lot lately, uh, parallel invention. Everybody’s basically in something like the Industrial Revolution right now — once people, uh, had the steam engine, everybody went, whoa, so that’s how energy and mechanical energy can convert into each other.

[19:11] So everybody’s, uh, inventing products like this off the needs that already exist, so I don’t think it matters much who’s borrowing from whom, who’s copying whom, that kind of feeling. Sometimes it really is pretty demoralizing though — uh, like, this platform, the first version was already built around December last year, and at that point everybody on the market was still doing it the canvas-workflow way. That approach isn’t bad, by the way — you can get extremely fine-grained with it, and it, uh, meets creators’ needs well, but what it

[19:47] uh, is convenient for is short films. If you’re making a few-minute commercial or a short film, it’s really quite good. But the moment you’re dealing with a hundred-plus characters, then dozens of locations, even a hundred-plus props — a seventy- or eighty-episode show with hundreds of assets like that — imagine doing that inside a canvas like that: it’s really hard to pull off rigorous project coordination and organization, and it’s pretty easy to end up with continuity errors of all kinds.

[20:21] So this, I think, is a fairly, fairly new model. And also, also I feel like — because we’re not just heads-down building the system, we’re also making all kinds of short films and series ourselves, including the WeChat Channels content. Like this MBTI Floor 16 we’re about to launch soon, it’s like a, uh, 16-personality-types version of Friends, a show like that, uh, uh, and it was actually made with this same system.

[20:53] It makes the whole creative process very — how do I put it? Uh, ah, you’d call it — you can’t call it easy, you’d call it orderly. Because we’ve got a lot of J types on our team, so they really enjoy this process of collaborating with a system.

[21:18] Of course our team also has, also has P types, and P types have great creativity. It’s not like it’s purely NTJ personalities. Mm. Right, so I mentioned everybody in the industry has built one, right — like iQiyi’s Natto, right, and ByteDance, ByteDance built two, they built Xiao Yunque, and then they built Volcano Juchuang. And now, since the model is ByteDance’s, on Volcano Juchuang it can

[21:46] generate live-action-style output directly through the API. Otherwise you have to go the enterprise route and sign an annual framework contract, which basically means committing to a yearly, uh, compute spend, before they’ll open up that interface for generating live-action likenesses directly. Like, mm, among the earlier ones there’s Jurilu, which is also done really well, and then there’s Mulan — right, I’m a loyal Mulan user, and Mulan is also

[22:15] a really solid, uh, canvas-workflow tool for short video, and it’s an agent tool too. And my, uh, approach to agent collaboration, the earliest inspiration for it actually came partly from Mulan’s Little Mulan. And in the industry there’s also Ououyi, and LipTV, and TapNow — all, all of them really solid platforms.

[22:42] They’ve also put a lot of, uh, money into marketing, and into, mm, subsidizing users — on their platforms you really can get a very cheap price. But I do put a question mark on that: users you pull in with, with subsidies like that, in the end — like recently when Doubao started charging, everybody went, hey, you and I aren’t friends anymore, right. So when you actually want to turn a profit, when you want to go back, to go back to a normal price,

[23:12] the users you built up before, they’re just not that kind of paying user. Ah, our scripts for the first five episodes are done. It asks if there’s anything that needs adjusting — and you can adjust all of it through natural-language conversation. Let’s, let’s just go run a character detection pass. Let’s just go run a character detection pass.

[23:44] Mm, it’s identified 16 characters now, and then it runs a second review pass. Okay, it’s already, already, already done identifying them. Once that’s done, we can run the character breakdown. Let’s just go break down Zuo Zimu.

[24:12] Break down Duan Yu. See, even a character like the Lightning Marten is in there. It analyzes based on all the scripts. And then what this character should look like, plus the era the story happens in — because as you know, Demi-Gods and Semi-Devils is actually set in

[24:37] it should be a story set in the Northern Song period, so the character looks and costumes in there will fit that era. Anyway, coming back — right now this tool is given to some of our partners, like some production teams,

[25:06] or some individual OPC creators to use. We also haven’t been considering going broadly to C right now. Overall, my thinking is that for an ordinary C-, C-side user, this tool is actually expensive, and it doesn’t deliver that much value either — because if a consumer user is just making things for their own fun, making little short films, then the kind of

[25:37] canvas workflow, or, or canvas agents on the market are already a perfectly good option, and there’s no need to create with a tool this heavy. The people who genuinely need a tool like this genuinely need the efficiency gain: they’ve already, already landed the production contract, meaning I’ve already got the money in — maybe 800, 1,000, 1,500 per minute,

[26:03] that kind of fee, and once you’ve got that fee in, I know exactly where my revenue is. And my cost might be 200 or 300, that kind of cost per minute. Then it makes total sense for me to use a system like this to raise overall output efficiency. So that it can — I’m on rendered style right now — so that it can put out this, this content faster.

[26:28] So we’re on rendered style right now, so let’s look at the rendered style first, wait for the image. Zuo Zimu — the rendered style version is already, already done. This is basically the casting stage, and it’s produced the character’s look. And we can also have it generate live-action — Duan Yu is out too. Duan Yu’s look, and then also the

[26:53] Lightning Marten’s look. Right, I forgot to pick live-action a moment ago, let’s select live-action now. Let me see. Regenerate it. Because it might need to re-run the breakdown. Let’s test the system too, because we’ve never had a case where, with a

[27:18] series half-produced, you go back and change the entire art style. Right, I was halfway through a point earlier — now if you, if you’ve already landed the contract, being able to use this system to boost efficiency is really great, because after all — ah right, with a system like this, a skilled, skilled AIGC director can put out roughly five to eight minutes of fairly high-quality video content in a day.

[27:49] That’s a fai-, uh, fairly solid output capacity. Mm. Oh, okay, it actually — it really works, no need to re-run the breakdown. Oh, so it’s already produced Zuo Zimu’s live-action look.

[28:15] You can see there’s a lot of detail in here — including the skin texture, and the fact that it’s not a perfectly symmetrical look. Okay, that’s it for today’s share for now. As for the rest, since we’re out of time, anyone who’s interested, we can talk more one-on-one.