Sumi Studio

AI Production Studio
You direct. Sumi produces.
Films • Ads • Series • Animation
Supported by @nvidia & @byteplus

Try it free ↓

[Sumi Studio @sumistudio_ai]

🚨 CoreTV is now Sumi Studio
A new name. A new identity. The same ambition - taken much further.

We started with a simple idea: make AI filmmaking easier by bringing the creative process into one place.
But the product evolved.
It became more than a filmmaking canvas. We’re building a space where AI can actively work with your projects - from developing an idea and creating visuals to editing, organizing assets and bringing the pieces together.

And there’s someone new at the center of it all 👀
Meet Sumi - our AI agent.
Sumi helps connect the models, tools and creative workflows needed to turn an idea into something real.

The name changed because the vision grew.
Welcome to Sumi Studio.

You direct. Sumi produces.
Link in bio 🎬

[Sumi Studio @sumistudio_ai]

New to the Sumi Studio canvas? Here's how it works: adding nodes, connecting shots, fixing one scene without touching the rest, and editing the final cut, all in one place. https://t.co/pAdEYru5Wh

[Sumi Studio @sumistudio_ai]

@MANISH1027512 You put into words what we've been trying to explain for months: Claude makes video, models generate pixels. Sumi is just the bridge between them 🐾

[Sumi Studio @sumistudio_ai]

"Before, Claude was my editing cheat code. Now, it's an entire production team."

This is the whole idea behind the Sumi MCP. Claude directs. Sumi renders. You approve every shot.

Here's what it looks like from inside the loop 👇

[古一 @MANISH1027512]

Claude + Sumi MCP = 直接起飞！

我一直说 Claude 是我的剪辑外挂。但有个误区要先讲清楚：Claude 能“做”视频，不等于能“生成”视频。

Claude 做视频走的是 Video as Code 路线。它用 Remotion、HyperFrames 这类框架，把画面写成时间轴上的代码，排版、动效、转场、调色、字幕，都能精确到帧。它最强的是编排：给它一段 BGM，它能分析出 BPM 和鼓点，精准卡点。
但它画不出一个真实的人，也做不出一束流动的光、一次转身扬起的发丝。这种像素级的真实画面，还是得交给视频生成模型。

所以我以前的流程一直是割裂的：图在一个平台生成，视频去另一个平台跑，配音再换一家，提示词另开窗口写，最后把一堆文件拖回来交给 Claude 剪。Claude 的品味和编排能力再强，也只能在你喂给它的素材里发挥。

这次接入 @sumistudio_ai 的 MCP，这个断点终于接上了。
Claude 可以直接在 Sumi 里调用模型、查看报价、生成素材、取回文件，然后接着剪。创意、生成、剪辑，在同一个对话里形成闭环。

这条韩蛛俐《白房间》就是这么做出来的：
· Claude 先排好分镜和卡点，再按镜头决定用哪个模型生成
· 日语台词直接调 MiniMax 语音，玻璃碎裂、墨滴入眼这些音效也是现
· 去水印、超分、卡点特效、片尾卡，全部用代码完成

成本也比我预想的低：
· 全片积分不到 1200 （参考，sd2.5 30秒720p消耗积分1500）
· 我让 Claude 每次生成前先报价，我确认后才跑。简单镜头用 Wan3、H要质感的镜头才上 Seedance 2.5
· Claude 这边的 token 只用了周额度的 5%，而且是 20 刀的 Pro 账号
· 语音和音乐也能直接调 MiniMax 和 Suno v6，以前这些 API 都得我自

一句话：以前 Claude 是我的剪辑外挂，现在它是一整个制作组。

[Sumi Studio @sumistudio_ai]

@MiunnElmatar That smirk holding through the push-in is the hard part on anime faces. How many expression close-ups did you lock for Sigrid before generating the move?

[Sumi Studio @sumistudio_ai]

@Rufuszone Code renderers win on readable silhouettes - forklift, racks, hi-vis - because the mesh stays honest. Diffusion wins on skin and light falloff. For a safety spot I'd keep the warehouse mesh fixed and only regen the worker motion.

[Sumi Studio @sumistudio_ai]

@coding_xiaoguo Writing glyphs into weather on a continuous FPV is brutal - FOV drift shears the letters. How did you keep focal length steady across the whole take?

[Sumi Studio @sumistudio_ai]

@EonKgosi That trial-and-error loop is the craft. Lock proportions in the refs first, then only re-roll the shot that drifted - saves burning a whole take when one detail slides.

[Sumi Studio @sumistudio_ai]

@thejonfredric Blocking in Blender first is the move - feels like directing, not gambling. Agents help more when they sit inside a project that already holds those blocks, not a fresh empty chat. That's the gap we built Sumi around.

[Sumi Studio @sumistudio_ai]

@huseyinrifatk That MJ look plus Kling motion control is a strong stack. Shot-by-shot polish is usually what keeps the face from melting once the dance pass gets long - worth the extra pass.

[Sumi Studio @sumistudio_ai]

Shanghai, you were something else 🌊
Our team spent 23–24 October at WAVE YOUNG 2026, the China–Europe AI Tech Festival on the Yangpu Riverside, and we got to pitch Sumi Studio on the NEXT WAVE stage 🎤

Two days of demos, deep conversations about AI with investors and founders, and evening receptions on the rooftop with the river in front and the city behind.

Swipe for the highlights from our presentation and the festival ➡️

Huge thanks to Centro Innovazione Italiano and everyone who came to say hi 🙌

#AI #AIVideo #TechFestival #Startup #Shanghai #Innovation

[Sumi Studio @sumistudio_ai]

@qinmeng827 Good rule. Asking one job to hold identity and hit exact frames at once is where the mush comes from. Locking the character in a still first, then using that still as the first frame, gets you most of both and saves the redo credits

[Sumi Studio @sumistudio_ai]

@VanishingFilmHQ @higgsfield_ai Framing tighter is the right call. In a wide the face is only a few pixels tall, so the model guesses instead of reading the sheet. Keeping wides for place, silhouettes and backs to camera, then cutting to the close-up for the face, usually hides it well

[Sumi Studio @sumistudio_ai]

@YICHENAI2026 Nice split - Codex on scripts, Nano Banana on refs, Seedance only on motion. Does the skill also break each story into a timed shot list before anything hits Seedance?

[Sumi Studio @sumistudio_ai]

@jennyhuynh___ Models drift toward the faces they saw most in training. A 3-4 angle sheet of your own face works way better as the ref than one selfie, plus naming your skin tone in every prompt. In Sumi Studio we keep that sheet in the project so every shot pulls from it

[Sumi Studio @sumistudio_ai]

@denglitao0308 Agree, the drift usually shows up between tools, not inside one shot. Keeping character refs, voice and script in one project is what fixed it for us in Sumi Studio, so a new scene starts from the same story instead of a blank prompt.

[Sumi Studio @sumistudio_ai]

@sada_ai The three-panel ref with the back view is the part most people skip. How often does the swap still drift when the base clip has a quick head turn?

[Sumi Studio @sumistudio_ai]

@huntonraptor @base Legends 🔥

[Sumi Studio @sumistudio_ai]

Great thread from our co-founder @NickyLi42 about how to use multi characters in dynamic dance and keep consistency

[Nicky_42 @NickyLi42]

🧟‍♀️ MJ's Thriller, redone with 9 AI-model girls: DeepSeek, Claude, GPT, Grok, Gemini, Kimi, GLM, MiniMax and Qwen. For Halloween 🎃

Made with Codex + @sumistudio_ai. Characters by @ZipZipPipe (CC BY-NC-SA 4.0). This video is shared under the same license.

132s, all 54 original shots, original timing. How it was made below 👇