Viggle AI became visible to many people through a wave of viral character swaps. Behind those memes is a Canadian AI startup trying to make AI video more controllable, and a training-data debate that is becoming harder for the industry to avoid.
From viral remixes to controllable motion
One of Viggle's most recognizable moments came through videos built from footage of the rapper Lil Yachty bouncing onstage at a summer music festival. Users replaced him with other figures, including Joaquin Phoenix's Joker and Jesus, creating countless variations from the same basic motion.
The central idea is simple: Viggle lets a character take on a specified movement. In one workflow, a user uploads an original video, such as Lil Yachty dancing onstage, and also uploads an image of the character that should perform that motion. The system then generates a new clip using that movement as the guide.
Viggle also supports other creation paths. Users can upload a character image and pair it with a text prompt describing how the character should move. They can also create animated characters from text prompts alone.
That focus on motion is the company's main pitch. Viggle trained a 3D-video foundation model called JST-1, which the company says has a "genuine understanding of physics," according to its press release. CEO Hang Chu argues that this makes Viggle different from AI video generators that can produce unrealistic character movement.
Why creatives are using Viggle beyond memes
The meme cycle gave Viggle public attention, but Chu says memes represent only a small percent of its users. According to him, filmmakers, animators and video game designers are using the model as a visualization tool.
The appeal is not that the videos are finished production assets. The source article notes that Viggle's clips are still shaky and that faces can look expressionless. The value, at least for now, is speed: a person with an idea can turn a character motion concept into something visible without building the whole sequence manually.
That makes Viggle less like a conventional video editor and more like an early-stage ideation tool. A creator can test a movement, pose or character concept quickly, then decide whether the idea is worth developing further.
Chu described the company's direction in broader terms: "We are essentially building a new type of graphics engine, but purely with neural networks," he said. He also said the model is different from existing video generators because it is designed around structure and physics rather than mainly around pixels.
The business around the model
Viggle currently offers a free, limited version of its AI model through Discord and its web app. It also sells a $9.99 subscription for increased capacity and gives some creators special access through a creator program.
The company is also looking beyond individual users. Chu says Viggle is talking with film and video game studios about licensing its technology. At the same time, he says the tool is being adopted by independent animators and content creators.
On Monday, Viggle announced a $19 million Series A led by Andreessen Horowitz, with participation from Two Small Fish. The startup says the funding will help it scale, accelerate product development and expand its team.
Viggle told TechCrunch that it partners with Google Cloud, among other cloud providers, to train and run its AI models. The source article notes that Google Cloud partnerships often include access to GPU and TPU clusters, but typically not YouTube videos for AI model training.
The YouTube training-data issue
The most sensitive part of Viggle's story is not the memes. It is the data used to train the model.
When TechCrunch asked Chu what data Viggle's AI video models were trained on, he said: "So far we've been relying on data that has been publicly available." Asked whether the training dataset included YouTube videos, he replied: "Yeah."
That answer matters because YouTube CEO Neal Mohan told Bloomberg in April that using YouTube videos to train an AI text-to-video generator would be a "clear violation" of the platform's terms of service. His comments were made in the context of OpenAI potentially having used YouTube videos to train Sora.
Mohan also said Google, which owns YouTube, may have contracts with certain creators to use their videos in training datasets for Google DeepMind's Gemini. But according to Mohan and YouTube's terms of service, harvesting video from the platform is not allowed without obtaining permission from the company.
After the interview, a Viggle spokesperson first tried to walk back Chu's statement, saying he "spoke too soon in regards to if Viggle uses YouTube data as training." The spokesperson added that Hang/Viggle was unable to share details of the company's training data.
TechCrunch then pointed out that Chu's earlier comments were on the record and asked for a clearer statement. In response, Viggle's spokesperson confirmed that the startup trains on YouTube videos, saying the company uses a variety of public sources, including YouTube, to generate AI content.
A wider gray area for AI video
Viggle is not alone in facing questions about YouTube as a training source. The source article says other AI model developers, including Nvidia, Apple and Anthropic, have reportedly used YouTube video transcriptions or clips for training.
That places Viggle inside a broader unresolved tension. AI video systems need large amounts of training material, while major platforms have rules about how their content can be accessed and reused. The unusual part of Viggle's case is how directly the issue surfaced.
For users, the immediate product story is still about controllable AI characters: upload a motion source, upload or describe a character, and generate a new video. For the industry, the larger question is whether tools built from publicly available online material can keep scaling while satisfying the platforms that host that material.
Viggle's next challenge is therefore twofold. It needs to improve the quality of its character videos, moving beyond shaky clips and expressionless faces. It also has to navigate the training-data scrutiny that follows any AI video company willing to say YouTube out loud.