What if you could feed a movie scene into an open source tool and get the same scene back as a Pixar-style cartoon, with the same characters, the same camera moves, and the same story beats? That is exactly what Soap2Soap does. Built by Show Lab at the National University of Singapore, it is an open source pipeline that remakes long videos in a new style or with recast actors, instead of generating a few seconds of footage from a text prompt.
Key Takeaways
>
Soap2Soap remakes full video segments, up to 42 consecutive shots, not just short clips.
Three AI agents split the work: one understands the video, one generates the remake, and one verifies every shot for consistency.
Styles in the code include Pixar, Disney, LEGO, anime, clay, and Family Guy.
The code is on GitHub at showlab/soap2soap, with 120 stars as of October 25, 2026.
You need a Gemini API key and a BytePlus key for Seed Dance 2.0 to generate real video.
What Soap2Soap actually does
Most AI video tools start from a prompt and produce a few seconds of footage. Soap2Soap starts from an existing video and remakes it. The official code implementation describes itself as video-to-video generation powered by Google Gemini and Seed Dance 2.0, and it supports two scenarios from the accompanying research paper, Soap2Soap: Long Cinematic Video Remaking via Multi-Agent Collaboration.
The first scenario is real-to-stylized remaking: take a live action scene and render it as a Pixar-style cartoon, a LEGO movie, a clay animation, or an anime sequence. The second is live action re-casting: keep the scene but replace the actors, using reference images of the new cast. In both cases the goal is the same: preserve the narrative, the motion choreography, and the identity of every character across the whole segment.
The demo examples in the README show how far this goes. A Huaqiang Watermelon Meme clip becomes a clay animation across 18 shots and 6 characters, with the original Mandarin dialogue preserved. A Titanic scene becomes a Pixar-style remake across 12 shots and 5 characters, with English dialogue intact. Other showcases turn a Doctor Strange scene into LEGO and a palace drama into a Family Guy style 2D cartoon.
Watch the 30-second repo signal: Turn Any Video Into Pixar With Soap2Soap.
How three agents remake a movie
The hard part of remaking a long video is consistency. A character who looks one way in shot 1 must look the same in shot 40, and the story must still make sense when the shots are stitched back together. Soap2Soap solves this with three agents working under what the paper calls Dual-Bridge Consistency.
The understanding agent reads the movie first
Before anything is generated, the understanding agent watches the source video and writes a detailed screenplay in JSON format. It breaks the video into scenes and shots, and for each shot it records the characters, the setting, the camera language, the action, and the dialogue. Transcription comes from Whisper, and the analysis runs in roughly 60-second chunks in parallel, so long videos are processed in manageable pieces.
The generation agent draws in grids, not one frame at a time
The generation agent does not render shots in isolation. It creates keyframes for several shots together in 2x2 or 3x3 grids, so characters and environments stay visually consistent across shots. A unified Design Sheet, built from character reference images, anchors what each character looks like. Those keyframes then become short video clips through an image-to-video model.
The verification agent checks every shot
The third agent audits the keyframes and clips against the shared screenplay and the visual anchors. When something drifts, a character changes face, a setting shifts, a story beat breaks, only the affected shots or time windows are regenerated. This selective regeneration loop is what lets the pipeline hold together across dozens of consecutive shots instead of falling apart after the first few.
The paper versus the code: two different model stacks
One detail worth stating plainly: the research paper and the public code do not use identical models, so do not treat them as the same stack.
Aspect | Research paper | Public code |
|---|---|---|
Understanding and verification | Gemini 3 Flash | Gemini (via GENAI_API_KEY) |
Keyframe generation | Nano Banana 2 | Gemini image generation, Imagen 3 |
Video generation | Veo 3, 4 to 8 second clips | Seed Dance 2.0 by default, Veo 3 as an alternative |
Output | Benchmark results on SoapBench | Final video merged with ffmpeg at 1280x720 |
The paper's benchmark, SoapBench, covers 10 movies with 607 shots in its understanding track, and over 95 percent of shots carry explicit character anchors. In the remaking track, each movie contributes one continuous segment of 1.5 to 5 minutes, with the longest sequence reaching 42 consecutive shots. On the paper's 10-point VLM scores, Soap2Soap reached 9.17 for character identity, 8.84 for scene consistency, and 8.67 for plot, ahead of Kling O1 at 8.11, 8.37, and 8.60. One caveat comes from the paper itself: the VLM judge is Gemini 3 Flash, from the same model family used for generation, with CLIP scores reported as a model-agnostic cross-check.
What you need to run it
Soap2Soap is a Python project, and the README lists clear requirements: Python 3.10 or newer, ffmpeg, a Gemini API key, and a BytePlus API key for Seed Dance 2.0, which is required for real video generation. Optional keys cover Veo 3 and Runware. The default style is Disney, the default maximum is 10 shots, and the default keyframe mode is consistency mode, which favors character stability over variety.
A few practical notes before you try it. First, the repository currently states no license: the GitHub API returns a null license field and the repository root contains no LICENSE file, so check the terms before building a product on top of it. Second, video generation through commercial APIs is not free, and a full remake of a multi-minute segment means many model calls, so costs add up fast. Third, the repository was created in March 2026 and last pushed in May 2026, with 120 stars and 13 forks as of October 25, 2026, so treat it as research code rather than a polished product.
Why Soap2Soap matters
Short AI clips are already common. Long, coherent remakes are not. Soap2Soap is interesting because it treats a video as a story to be understood first and rendered second, and because it shows its working: a screenplay, a design sheet, and a verification loop. For creators, it points toward a future where restyling a scene is an editing decision, not a reshoot. For developers, the multi-agent pattern, understand, generate, verify, is reusable far beyond video, anywhere long-form output has to stay consistent.
Read this article in Indonesian: Soap2Soap: Ubah Video Menjadi Remake Gaya Pixar Dengan AI Open Source.
FAQ
Is Soap2Soap really open source?
The code is publicly available on GitHub under showlab/soap2soap, but no license is currently stated in the repository, so its legal reuse terms are unclear until the authors add one.
Can it remake a full movie?
The paper demonstrates segments of 1.5 to 5 minutes, up to 42 consecutive shots. Full feature-length remakes are not demonstrated, and cost and compute would grow with length.
Which styles does it support?
The code lists Pixar, Disney, LEGO, anime, clay, Family Guy, and realistic styles, with Disney as the default.
Do I need paid API keys?
Yes, for real video generation. You need a Gemini API key and a BytePlus key for Seed Dance 2.0, with Veo 3 available as an alternative video model.
How is it different from a text-to-video tool?
Text-to-video tools invent footage from a prompt. Soap2Soap remakes existing footage, preserving the original shot order, camera language, characters, and dialogue structure in a new style.