Open genjutsu on Genjutsu AI is a hosted editor, independent of Higgsfield and the GitHub projects below, so you need no NVIDIA card, ComfyUI or API key: upload an MP4 or MOV, pick 5 to 30 seconds, add a character photo. A swap costs from 39 credits (about $1.45); new accounts get 3 free credits, not enough for one. The 480p–1080p MP4 has the clip’s original sound, watermarked until your first purchase.
How open genjutsu works here
- 01
Upload the video and pick the seconds
Upload my video is already selected: add an MP4 or MOV up to 5 minutes long and 50 MB, then slide to the 5 to 30 seconds you want. Nothing is installed or downloaded.
- 02
Add the character photo and say who it replaces
Keep the change type on Character, add a photo or drawing of the new character (up to 9 images, 10 MB each) and edit the instruction so it names who in the clip each image replaces, for example “the woman in the yellow coat”.
- 03
Check the price and generate
Seedance 2.5 edits the clip itself on our servers. The button shows the exact credits first, and the finished MP4 carries the selected seconds’ original sound.
Character swap examples: original clip and result
Each pair shows an original clip on the left and its character swap on the right: one person, a couple and a trio, each replaced by new characters while the setting and the framing match. The clips are silent, landscape and 13 to 14 seconds long.
原片
换角色后
A woman in headphones talking to a selfie camera on a forest road is swapped for an aviator in a leather flying helmet, goggles and a sheepskin collar. In the swapped clip the head turns, the forest and the framing match the original. 14 秒 · 横屏 · 无声。
原片
换角色后
A ballroom couple practising in a mirrored studio is swapped for two flamingo characters, one in a white shirt and one in a coral gown. In the swapped clip the hold, the turns and the studio match the original. 14 秒 · 横屏 · 无声。
原片
换角色后
Three dancers under stone arches are swapped for two toy soldiers and a ballerina doll. In the swapped clip the routine, the three positions and the arches match the original. 13 秒 · 横屏 · 无声。
Which open genjutsu GitHub project is which?
Open genjutsu is not one program: the name covers several GitHub projects that copy Higgsfield Genjutsu’s video character swap. D3voz/openjutsu runs open model weights on your own NVIDIA card through ComfyUI; Anil-matcha/open-genjutsu and abdullayevziyodulla/open-genjutsu are local apps that send each job to Seedance 2.5 or Wan through paid API accounts you bring; Comfy-Org/comfy-openjutsu is the code behind the hosted Openjutsu app on Comfy’s platform.
| Project | Where the video is made | What you need | Clip limits | Cost per video |
|---|---|---|---|---|
| D3voz/openjutsu (about 2,500 stars) | Your PC: MiniMax H3 open weights, SAM 3.1 tracking, Qwen3-VL captions, all in ComfyUI | An NVIDIA card (tested on a 16 GB RTX 4060 Ti with 64 GB of RAM), ComfyUI, custom nodes and model downloads | Whole clip or a chosen part; 6 s at 480p took about 6–13 min in testing | No per-video fee; electricity and your hardware |
| Anil-matcha/open-genjutsu (about 1,200 stars) | A hosted API: Seedance 2.5 Omni Reference, or Wan 3.0 for the Lite option | Python 3.10+, FFmpeg and an API key with generation credits | Source 4–15 s, 50 MB; up to 8 character images; output 4–30 s | Paid to the API account; preparation steps (depth, frames, character sheets, prompt) are billed too |
| abdullayevziyodulla/open-genjutsu | Hosted providers you pick: Seedance 2.5 or Wan 2.2 Animate | Python 3.12+, Rubber Band and your own provider keys | 480p draft or preview, then a 720p or 1080p render depending on provider | App free; preparation and every render billed to your provider accounts |
| Comfy-Org/comfy-openjutsu | A Comfy cloud deployment running MiniMax H3 | A Comfy Developer Platform account, an API key and a GPU deployment | MP4 or MOV up to 100 MB; a 5–15 s part | Deployment credits while the GPU runs |
| Genjutsu AI (this page) | Our servers: Seedance 2.5 edits your clip, then the original sound is put back | A browser and an account; no install, GPU or key | MP4 or MOV up to 50 MB; a 5–30 s part; up to 9 photos | 39–928 credits (about $1.45–$34.50), shown before you generate |
Facts for the four projects are taken from their own README files and repository pages, checked 19 November 2026; star counts change daily. All four are MIT licensed, but the model files have their own licences: D3voz/openjutsu notes that its face-fix step uses InsightFace buffalo_l (non-commercial research only) and Ultralytics YOLO weights (AGPL-3.0). Forks and copies with similar names, such as other “openjutsu” and “open-genjutsu” repositories, are not listed here. Genjutsu AI is not one of these projects and is not affiliated with their authors.
Is open genjutsu free?
Open genjutsu is free to download but not always free to run. D3voz/openjutsu has no per-video fee once you own a capable NVIDIA card; the two API-based open-genjutsu apps are free and MIT licensed, yet every preparation and render step is billed to your own provider account. The browser editor on this page charges credits per swap instead, with the price shown before you generate.
On Genjutsu AI, the cheapest swap is 39 credits (about $1.45) and the 3 sign-up credits do not cover one on their own. Uploading, trimming, adding photos and seeing the price cost nothing. Swaps from accounts without a purchase carry a moving genjutsu-ai.app watermark; any purchase unlocks clean downloads, including earlier videos.
| Question | What happens here |
|---|---|
| Price per swap | 39 to 928 credits (about $1.45 to $34.50 at the lowest credit price), set by the model, the resolution and the seconds you select. A 10-second swap on Seedance 2.5 at 480p is 78 credits (about $2.90). |
| Free credits | 3 sign-up credits, valid for 7 days. Not enough for a swap on their own: the cheapest is 39 credits. Uploading, trimming and seeing the price cost nothing. |
| Smallest purchase | The $5 Trial pack (one-time) has 24 credits, not enough for one swap. The smallest one-time pack that covers one on its own is $12 for 220 credits. The cheapest start from zero is the $9 Starter plan, a monthly subscription with 240 credits a month. |
| Watermark | Swaps from accounts without a purchase carry a moving site watermark. Any purchase unlocks clean downloads of earlier and later videos. |
| Failed or rejected | If the model fails, or its review turns down a photo or the clip, the credits go back automatically; an upload stopped by the safety check is never charged. If the model has delivered and only adding the sound back or exporting fails, the credits are kept and support finishes the video. A finished swap is not refunded. |
Is there an open genjutsu app or download?
Open genjutsu has no official phone app: the GitHub projects are downloaded from GitHub and run on a computer as a local web page in your browser (D3voz/openjutsu has a Windows start script; the open-genjutsu apps start from Python). Genjutsu AI needs no download at all: open this page in a phone or desktop browser, tap Upload my video, pick the clip from your gallery and download the finished MP4 to post it. A long 4K phone video may be over the 50 MB limit, so trim or export it smaller first.
Downloads that promise a packaged “Openjutsu APK” or installer outside GitHub are not from any of these projects and can carry malware.
Pricing and credits for a browser swap
A swap here costs 39 to 928 credits, about $1.45 to $34.50 at the lowest credit price; the resolution and the seconds you select set the price. A 10-second swap is 78 credits at 480p (about $2.90), 175 at 720p and 310 at 1080p. The smallest one-time pack that covers one swap is $12 for 220 credits, or the $9 Starter plan (a monthly subscription) gives 240 credits a month.
| 模型 · 分辨率 | 5 秒 | 10 秒 | 15 秒 | 30 秒 |
|---|---|---|---|---|
| Seedance 2.5 · 480p | 39 | 78 | 117 | 233 |
| Seedance 2.5 · 720p | 88 | 175 | 262 | 523 |
| Seedance 2.5 · 1080p | 155 | 310 | 464 | 928 |
Seedance 2.5 bills for the seconds of your clip it reads and again for the seconds it outputs, which is why a swap costs more per second than a text-to-video clip. If the model fails or its review turns down an input, the credits come back automatically; if the model has delivered and only our sound and export step fails, the credits are kept and support finishes the video. More packs and plans are on the pricing page.
What the browser version does not do
The browser editor trades the local projects’ control for no setup: it does not run offline or on your own GPU, it does not let you swap in your own model weights or LoRAs, and it renders at most 30 seconds in one go. The prompt asks the model to follow the clip’s movement and camera; results vary with the clip and the photo, so try a short section first. The usual problem spots, with what helps:
| What you may see | Why | What helps |
|---|---|---|
| The new face changes during a head turn or a spin | The face is hidden or side-on for a few frames and has to be redrawn | A sharp front photo plus a three-quarter view; a slower or shorter section |
| The wrong person is replaced | The instruction did not single the person out | Name them by side of the frame and clothes at the start of the clip |
| Small background people look odd | Tiny figures and reflections are hard to redraw | A clip where the person fills more of the frame; avoid mirrors |
| The model refuses before it starts | Its review turned down a photo or the clip; the credits come back | A clearer photo of an adult |
It does not swap live on a camera, change the voice or write new dialogue or music. For one built-in clip that only needs your photo, see motion transfer; for a fuller guide to replacing a person, see AI character swap; and for what is and is not free across the editor, see the guide Genjutsu AI free.
最近更新: 2026-11-19