Choose local, hosted or your own cloud
Free code. Real hardware requirements.
The app code is free. You supply compute, memory and storage. The figures below describe the author’s tested configuration; lower-spec hardware is not verified here.
16GB VRAM
RTX 4060 Ti, author’s test GPU. NVIDIA required by the project.
64GB RAM
Author’s system memory. H3 INT8 models stream from RAM.
ComfyUI 0.41.2+
Plus Node.js 22.6+ and Python for the API proxy.
No GPU? Upload your clip and reference photos
Upload your video and reference photos, choose an edit, then review the credit estimate before generating. No local GPU or installation needed.
Countryside Duo
PromptOriginal clips and effect examples from Higgsfield, shown as references. Generation uses the original and your photo; results may vary.
Start with a short clip. Your final quote depends on the task, model, duration, resolution and any image preparation.
Install nodes and place the models
The checked README lists 12 packs across all modes, including optional post-processing. Install the dependencies for your chosen workflow, then restart ComfyUI. The app’s preflight identifies missing registered nodes and files; it cannot validate your available VRAM.
Node packs and their roles12
SolarRenderer
Optional From picture rendering
ComfyUI-MiniMaxH3-Context-Loop
Clip loop and assembly; 0.7.3+
ComfyUI-Fantastic-MiniMaxH3-PromptBuilder
Masked video-edit latent
ComfyUI-KJNodes
Mask growth and VRAM cleanup
comfyui_controlnet_aux
From picture normal maps
ComfyUI-VideoHelperSuite
Load source video
comfyui-h3-oneclick-swap
Swap plans, tracking and roles
ComfyUI-H3-FaceRefine
Optional face repair
ComfyUI-Impact-Pack
Face-repair SAM loader
ComfyUI-Image-Detail-Contrast
Optional colour cleanup
Orion4D_FXMax
Optional tone mapping
comfyui-websocket-reconnect-fix
Live progress after reconnect
Model folders: diffusion_models for H3, loras for speed/control LoRAs, text_encoders for Qwen, vae for video/audio VAEs, and checkpoints for SAM. Optional face repair adds its own detector and recognition models. Use the README’s exact current filenames and download links.
Launch three local services
On Windows, the repository’s start_openjutsu.bat starts the app and proxy after ComfyUI is ready. For a manual setup, follow the three services below. These commands are instructions for your machine, not actions performed by this website.
- 01
Start ComfyUI
Keep ComfyUI listening on 127.0.0.1:8188. On Windows, set AIOHTTP_NOSENDFILE=1 in the shell that starts ComfyUI; the README documents a video-seeking stall without it.
- 02
Get the app
In a terminal with Git and Node.js installed, clone the repository. Copy .env.example to .env.local, then install the app dependencies.
git clone https://github.com/D3voz/openjutsu.git cd openjutsu # Copy .env.example to .env.local npm ci - 03
Start the API proxy
From the cloned repo, use your Python environment for both installation and the progress patch. Keep this terminal open. The proxy listens on port 8189.
python -m pip install comfy-api-proxy==0.1.6 python patch_proxy.py comfy-api-proxy run --max-upload-mb 500 - 04
Build and open the app
In another terminal in the same repo, check .env.local, build and start. Open http://127.0.0.1:3000. Keep this local; follow the project’s password and HTTPS guidance before exposing it on a network.
# .env.local COMFY_BASE_URL=http://127.0.0.1:8189 COMFYUI_URL=http://127.0.0.1:8188 COMFY_API_KEY= # Terminal npm run build npm start
Your first character swap
Choose one short shot
Use one visible person and avoid mirrors, cuts and heavy occlusion for the first test. Keep an untouched copy of your source.
Add the character photo
Choose Simple motion transfer and upload the source and character image. Name the target by role, not clothing, to avoid leaking clothing words into the result.
Select a short part
Set the start and length under Part to use. Begin at 480p before trying a longer or higher-resolution run.
Check the mask
Run Check mask. Inspect the tracked person and the regenerated area frame by frame. Adjust Mask margin if limbs are cut off or too much background is selected.
Generate once, then inspect
Follow job progress instead of repeatedly resubmitting. Check identity during turns, hand contact, edges and audio timing across the entire output.
Save a useful result
Download the video, keep the input and settings, then change one variable for the next attempt. A good still frame is not a full-video quality test.
What the other modes do
Advanced motion transfer
Map 1-4 characters to roles. Review person matching before a multi-person render.
Object swap
Change, remove or add an object, or change the background with the relevant controls.
From picture & Relight
Prepare a posed character image and optionally change lighting before rendering. This is a different preparation path.
Fix faces / Fix colours
Optional post-processing for a finished video. Review face-model licensing before commercial work.
Render time needs context
About 5 minutes
Genjutsu AI operator estimate. Varies by model, clip and queue; not a matched benchmark.
6-13 minutes
Local Openjutsu: 6s / 480p on a 4060 Ti, author report.
About 28 minutes
Local face repair: 17s / 768p, four chunks on a 4060 Ti, including loading. Author report.
These are different tasks on different systems. They cannot establish which tool is faster or produces better results. Compare the same input, settings and acceptance criteria before choosing for quality.
Comfy hosted and RunPod routes
Comfy’s hosted app avoids installing the local stack. Its published app workflow selects 5-15 seconds from an MP4/MOV up to 100MB, with one continuous shot and a person visible in the first frame. A fade from black or hard cut can break the workflow.
For your own cloud GPU, community RunPod Pod and serverless projects exist. Budget for compute and storage, follow their authentication instructions and stop idle compute. We have not deployed those templates for this guide.
Common failures and practical fixes
MemoryError on a long 1080p file
Issue #44 reports native-resolution decoding before scaling. Trim and downscale the source file before importing, then check the issue for the current fix.
Missing nodes or model files
Use the preflight’s exact missing-node names and destination folders. Restart ComfyUI after installation. A successful preflight does not guarantee sufficient memory.
The reflection keeps the original person
The README documents this limitation. Choose a shot without reflective glass or mirrors instead of expecting a prompt to fix it.
Small faces lose detail
Try a closer, clearer source shot first. Optional face repair adds another pass and separate model-license considerations.
Progress freezes or video seeking stalls
Check ComfyUI and proxy logs before resubmitting. Apply the documented proxy patch; Windows users should check AIOHTTP_NOSENDFILE.
Code licenses are not model licenses
MIT app code
The local Openjutsu app is MIT-licensed. Check separate licenses for the models, weights and custom nodes you actually enable.
Optional face repair
The local README identifies InsightFace buffalo_l as non-commercial research only and the Bingsu/adetailer YOLO weights as AGPL-3.0. This is not a blanket ban on all Openjutsu workflows.
Hosted services
A paid subscription does not replace permissions for your inputs. Read the service and model-provider terms for your specific use.
What does one video cost?
USD examples, not fixed per-video prices. See the exact quote before generating; taxes and regional prices appear at checkout.
≈210 credits
Standard edit · 10s / 720p. A published pricing example, not every model’s rate.
≈510 credits
Kling 4.0 Reference · 10s / 1080p. Video background, source audio on, no prepared image. Editor estimate.
≈770 credits
Scene or multi-person · 10s / 720p. Model and preparation can change the quote.
Collect up to 210 credits over a 7-day check-in streak. Packs start at $7.90 for 220 credits. Free and paid-credit exports have no watermark.
What to upload
Your existing video
MP4, WebM, OGG or MOV. Up to 200 MiB. Select a 4-30 second segment; the model can impose a shorter limit.
Your reference photos
JPG, PNG, WebP or HEIC/HEIF. Up to 30 MiB each. Up to 30 images across supported workflows; some models accept fewer.
For clothing edits
Clothing images: up to 20 MiB, at least 300 px on each side. Use the dedicated clothing workflow for garment controls.
A front-facing portrait is not mandatory. Clear, well-lit references help; image preparation is workflow-dependent. This edits an existing clip, not a video from only one photo. Download completed videos within 7 days.
Questions, answered
Is there a free way to start?
Yes. Genjutsu AI offers up to 210 daily check-in credits over a 7-day streak, with no purchase required and no watermark. Rewards expire after 30 days. Local Openjutsu code is free, but hardware and compute are not.
What GPU does local Openjutsu need?
An NVIDIA GPU is required by the local project. The author tested an RTX 4060 Ti with 16GB VRAM and 64GB system RAM. This is a tested configuration, not a certified minimum.
How long does a video take?
Plan for about 5 minutes on Genjutsu AI, an operator estimate rather than a measured service guarantee. Model, clip and queue affect the wait. Local Openjutsu’s author reports 6-13 minutes for a 6-second 480p clip on a 4060 Ti.
Why does a long 1080p clip raise MemoryError?
Issue #44 reports a loader decoding the entire long video at native resolution before scaling, exhausting system RAM. Trim the file before importing and check the issue for fixes; lowering output resolution alone may not fix that loading path.
Can I run Openjutsu on RunPod?
Yes, community deployment projects exist, including rodamora/openjutsu-runpod and zociwhite/openjutsu-serverless. They are self-managed routes with GPU and storage costs, not a free hosted service.
Does Openjutsu handle mirrors and small faces?
Not reliably. The local README warns that reflections can retain the original person and contaminate the replacement. Small faces may need an optional repair pass, which has its own cost and license conditions.
What does Comfy’s published app accept?
Its documented workflow uses MP4/MOV files up to 100MB and a selected 5-15 second part. Use one continuous shot with a person visible in the first frame. Check the hosted app for current limits and prices.
Can I use the result commercially?
It depends on the input rights and tool terms. Local Openjutsu’s optional face repair uses separately restricted models. Genjutsu AI does not claim ownership of your output rights, but permission, applicable law and provider terms still determine commercial use.
Can I use an Openjutsu alternative in a browser?
Yes. Genjutsu AI and Comfy’s hosted Openjutsu are online options. Here you upload an existing clip plus reference photos; no local GPU or model installation is needed.
How many credits does 1080p cost?
It depends on the model. A 10-second Kling 4.0 Reference motion edit at 1080p is estimated at about 510 credits with video background, source audio on and no image preparation. Do not apply this number to every 1080p model.
Is there a Chinese interface?
The repository has an English/Simplified Chinese interface contribution in PR #40. Check whether your installed revision includes it, and update according to the project’s instructions.
Sources & review notes
Specifications and reported timings come from the linked sources, reviewed on November 2, 2026. Prices and limits can change. This guide does not present a matched cross-tool performance benchmark.
Genjutsu AI is independently operated. It is not Higgsfield, Comfy or the Openjutsu project.
Start creating