"30 seconds of video for 17 dollars, pay again if it misses": a Seoul developer works out the math of AI video gacha and open-sources CozyClay, a 3D previz tool. 100 days, 1500 stars

CozyClay is an open-source 3D previs tool that runs in the browser: block out the scene, character actions, camera moves and cuts with white models first, then hand the same shot to AI video models like Seedance, Kling and MiniMax H3 to generate. It never hit Hacker News or the GitHub trending list, and its author has only about six hundred followers on X, yet it picked up roughly 1,500 stars in 100 days. Below: what it is, where the stars came from, who it's competing with, and where the open spots are if you want to build a star-worthy open-source project on the AI video workflow.

"Miss once, and you pay again"

In 2026 August 24 (Seoul time), a Korean developer named Yun (GitHub HaD0Yun, X @Yun_HDY) posted a casual tweet: "Now a 30-second video costs 17 dollars... miss once and you pay again. And one prompt can't hold the whole picture in my head. So I build the scene first, then turn it into video. I made a previs tool, and it's open source." It came with a comparison video: grey-box on top, final render below. The tweet got 39k views and 397 likes.

Two days earlier he had posted something sharper. The AI video platform Higgsfield advertised a 3D feature where you could drive Blender with a prompt by clicking a few buttons. He replied: "They're just throwing a few 3D figures into the render... If this actually worked, I'd interview at their company in my underwear. Even scams should be a little more careful." That post got 23k views.

These two posts were basically CozyClay's entire playbook afterward: a cost everyone can do the math on, a closed commercial platform to argue with, and a comparison video anyone understands at a glance.

What it is

Open cozyclay.org and the homepage's first line reads "A text prompt cannot hold a whole scene," and the second reads "AI video is charged per clip, and a miss means paying again."

The workflow has four steps:

• Build the scene: assemble a street from blocks and cylinders, and place rigged characters or capsule stand-ins;
• Animate: describe what a character does in one sentence (for example, "walk to the ball and push it with both hands"). It generates every frame and checks for clipping, foot sliding, and reach.
• Shoot: arrange shots on a timeline, with dolly and crane moves;
• Export: a "keyframe package" for video models, including first frame, last frame, grey-box reference clip, camera parameters, and a prompt describing the real composition. It can also export depth maps, storyboards, and an OTIO edit list for editing software.

Installation is a single npx cozyclay, with no GPU and no Python. An agent can drive it in three ways: a chat panel beside the interface (sign in with a ChatGPT account or enter your own key), an MCP server (Claude Code and Cursor both connect), and the command line. The repo even includes an installation guide written for AI agents, INSTALL_FOR_AGENTS.md.

It's AGPL-licensed and free. The repo was created in Jul 28, shipped v28 by Oct 3.2.1, and went through 30 versions in under three months. It originally lived under the Seoul open-source organization NomaDamas (the org behind k-skill, which has 8,000 stars), and moved to the author's personal account in early November. The old address redirects automatically.

Where the stars came from

The numbers make it clear. 10-15 had 778 stars, 10-20 came after the 808,3.0 launch plus the cocktailpeanut post, and in one day it went from 10-21 to 871,10-26, passing 1,000. In the second wave, 11-03 added +82 in a single day, the most on record, bringing it to 11-05 and 1,496 (GitHub; daily records start from 10-15). npm downloads were 1,728 in August, 1,772 in September, and 5,195 in October; from 10-22 to 10-26 it ran at 500–600 per day, then fell back to a couple hundred a day (npm, 11-05). So there are real users, but in the thousands of installs, not tens of thousands.

The cocktailpeanut post lays out the workflow most plainly: "A new AI video workflow is taking off: 1. Use an LLM to make 3D animation; 2. Feed it into a diffusion model like MiniMax H3 to generate realistic video. CozyClay makes the first step easy, no need to open Blender separately." A commenter added a practical tip: running H3,1080p locally on 4090 crashes past about 6.5 seconds, so have the LLM split the animation into shots under 6 seconds from the start, preview at 720p, and only render the shots you keep at full resolution. The people following this tool are exactly that crowd.

No paid promotion was found, and it did not appear on HN or Product Hunt.

Why this one

First, the math works. AI video is billed per generation, and you pay again when the result misses. Every AI video maker has felt that pain. "Rehearse in 3D first and burn fewer generations" needs no explanation.

Second, the timing was right. In the second half of 2026, models like Seedance 2.5, MiniMax H3, and Kling started to actually accept reference videos and first/last frames to control camera moves, so the "white model reference → final render" workflow finally worked. At the same time, Higgsfield (which advertised its 3D feature at the end of August) and Morphic (which released 3D Scene 10-01) pushed closed-source versions and educated the market on the need. CozyClay stood alongside and said "free, open source, export stays with you," which caught that demand.

Third, the content naturally spreads. Every post is "grey-box on top, final render below." The consistency between the two is obvious at a glance, and it can be reposted directly on X, Instagram, and YouTube.

Fourth, it hit the right distribution nodes. Pinokio's one-click launcher targets creators who run AI on their own computers. Once cocktailpeanut shared it and Theoretically Media covered it, 10-21 to 10-26 was the steepest stretch for stars and downloads.

Fifth, it reached the agent crowd. The MCP server, the command line, and the install guide written for agents got passed around among people who were already "using Claude Code for other things."

Sixth, the author kept shipping. Each major release came with a new demo video, and each video was a new round of spread.

Technically, there's nothing here others couldn't copy. It took off because it combined several things at once: open source and free, zero install in the browser, one-sentence motion generation, an export package built for video models, and agent control. The "free open-source alternative to Higgsfield" framing was the line KOLs used most when recommending it.

Who it's competing with

Open-source options exist too, but none are big:

"3D white-model → AI video" is already being built as a category, but apart from CozyClay, no second app has hit 200 stars yet.

If it were you

Verdict: B. The direction is right, and few people know about it yet: a newly emerging AI video workflow where the leading open-source project has only 1,500 stars and the number two is under 200. But building another browser previs right now most likely puts you in the row of dozens-of-stars projects in the table. CozyClay's author shipped 30 versions in three months, each with a demo video, so you're unlikely to beat it on its home turf. Closed-source platforms are also building this feature into their own products.

Who it's for: developers already making AI video who have actually been burned by gacha-style re-rolls; or people already following a creator community (the Chinese AI video scene, the ComfyUI community, the Blender community).

Not for: people who have never touched AI video and just saw a star curve and want to copy it, or anyone who can't do 3D and doesn't plan to make a demo video every week. Projects like this get their stars from a steady stream of side-by-side comparison videos. No videos, no spread.

How to do it: don't build a second CozyClay. Build the stage it doesn't cover.

• Switch the ecosystem: it's a standalone browser app. The ComfyUI side currently has only nodes with a few dozen stars. A genuinely usable "white model → video" ComfyUI workflow or node pack would plug straight into the people running H3 and Wan locally. A Blender plugin version works the same way.
• Switch the audience: Chinese creators mostly use Jimeng and Kling, and its export packs and tutorials are written for the English crowd. A version built for Jimeng / Kling with Chinese prompt templates and tutorials would be an empty space in the Chinese market.
• Build the second step: the cocktailpeanut post has a comment reading "4090 on 1080p runs out of memory past 6.5 seconds, so preview at 720p first, then render," which describes exactly the stage CozyClay doesn't handle. Batch-feed its exported shots into a local or cloud video model, auto-cut, select takes, and reassemble the final cut. An open-source tool that accepts CozyClay's export pack could ride its traffic.

How to get users: copy its playbook. Hook with a concrete money figure ("cut 3 re-rolls per clip, save XX yuan"); pair every release with a "input above, final cut below" comparison video; get listed on one-click launchers like Pinokio; in the README and website, write searchable long-form pieces like "XX software compared" and "motion control tutorial for model X"; for the Chinese version, post on Bilibili, Xiaohongshu, and Jimeng user groups instead of only X.

Before when: The window is probably the next two or three months. Video models have only just started taking reference videos, and workflows haven't settled yet. Once Higgsfield, Jimeng and Kling ship "3D positioning" as a built-in feature and do it well enough, the room for standalone tools will shrink, by inference.

What would show you're wrong: ① one of the follow-on projects in the table crosses 1,000 stars within a month or two, meaning the category has room and you should raise the rating; ② the built-in 3D blocking features on mainstream video platforms become good enough that creators no longer need an external tool, meaning the window has already closed; ③ CozyClay's own npm downloads keep sliding from a couple hundred a day and releases stop, meaning real demand for this workflow is smaller than it looks.

Data and sources

• GitHub: CozyClay repo and releases (as of 11-05/11-06,1,496–1,553 stars); daily star counts tracked from 10-15; star counts for open-source competitors and follow-on projects as of early November.
• npm: monthly and daily downloads for the cozyclay package (api.npmjs.org, as of 11-05).
• X: posts from @Yun_HDY 7-29, 8-22, 8-24, 9-09, 10-20; cocktailpeanut 10-21 post and replies; AI超元域 10-25 post; view counts as captured at crawl time.
• YouTube: videos from Theoretically Media and Shaun Kim (@SkimOnWest). Media and blogs: AlphaSignal (9-09), MindStudio (10-23). Homepage copy on cozyclay.org (11-06).
• Window timing and the judgment on each follow strategy are inferred; the specific content that drove the early-November second wave is unverified; the scale of dimension.so is unverified.