Hugging Face card · Read 2026-11-06 · No source film
Viggle Remake is a local re-shoot, and this page does not run it
The model card describes two rank-256 LoRAs on MiniMax-H3: one recasts people, one edits objects, weather and light. Nothing here loads those weights or accepts a movie to swap.
This site does not re-shoot a film.
Kling 2.6 makes a new silent clip, not a Viggle Remake
Nano AI does not run Viggle Remake and does not edit an uploaded movie. Kling 2.6 costs 55 credits for one new 5-second clip, with sound: false. It does not load swap or edit LoRAs, keep a source soundtrack, or paint frames. Eligible new accounts receive 64 credits. A 10-second clip is 110 credits.
- Made by
- Viggle
- Run on Nano AI
- Kling 2.6 · from 55 credits per video
- Page updated
- November 6, 2026
Source: huggingface.co
A card image and one edit still
Saved on 2026-11-06. The Hugging Face card has no footage. The second still is the wiki's turbine comparison, not a film frame. The wiki cover that shows a copyrighted scene is not used.
Viggle Remake beside OPENjutsu, without running either
Viggle Remake, on the card read 2026-11-06, is a pair of LoRAs on MiniMax-H3's base transformer, the fl2va weights, not ref2va. swap replaces people and characters. edit swaps, removes or inserts objects and changes background, weather or light. You paint one to four frames. Pins are optional. Sound is the source track, not a new score. The ComfyUI Wiki, updated 2026-11-04, says the project was published that day. The Hugging Face record's last change is the same date. This page does not treat an earlier card timestamp as the release announcement.
OPENjutsu, as that page already states, is D3voz's local app and, separately, Comfy's hosted beta. D3voz documents character and object changes while keeping camera, cuts and soundtrack, with Simple mode on the largest person and Advanced mode for 1 to 4 characters. Its app license is MIT and the model licenses are separate. Remake is not that app, not Comfy's hosted Openjutsu, and not Higgsfield Genjutsu. This site does not add a /genjutsu address.
The button above opens Kling 2.6 for a new clip. It does not take a performance video, does not apply either LoRA, and sets sound: false. The stills are the Hugging Face card image and the wiki's turbine comparison. The wiki's cover still is a film frame, so it is not shown here.
Viggle Remake at a glance
- Card read
- Hugging Face, 2026-11-06
- Modes
- lora/swap and lora/edit, rank 256
- Base download
- About 144 GB, or about 124 GB as the Comfy repack
- Selective LoRA set
- 5.5 GB; whole repo about 210 GB
- One render
- 90–345 frames, 17n+5, recommend up to 294
- 12 GB row
- Q4_K_M or Q8_0, reserve 8; int8 is 124 frames
- Kling 2.6 here
- 5 seconds, 55 credits, sound: false
What the model card says to download
Commands are copied from the Viggle Remake card and COMFYUI.md as read on 2026-11-06. They run on your machine. Nano AI does not install them.
- 1
Use the app if you do not want a GPU
The card's first line says the Viggle app runs Remake with no setup and no GPU. That path is theirs. This page does not sign you into it.
- 2
Do not clone the whole repository
The card says not to git clone the repository: with git-lfs that downloads about 210 GB of weights. Use hf download with an include list. The LoRA and sampler include below is the card's 5.5 GB set. The whole repo stays near 210 GB.
- 3
Fetch the base files the sampler expects
The card says the base model is about 144 GB of MiniMax-H3 files. The command is the one printed on the card. ComfyUI.md separately describes Comfy-Org's single-file repack as about 124 GB. Those are two downloads, not one number.
hf download MiniMaxAI/MiniMax-H3 --local-dir ./MiniMax-H3 --include "modular_model_index.json" "transformer/*" "text_encoder/*" "tokenizer/*" "processor/*" "vae/*" "audio_vae/*" "scheduler/*" "audio_scheduler/*" "docs/*" - 4
Fetch the LoRAs and the sampler, not every quantized file
lora/swap recasts people and characters. lora/edit changes objects, background, weather or light. Each is rank 256 and the card's contents table says 2.67 GB. Rank 128 and 64 live in lora_lowrank/ at 1.34 GB and 0.67 GB. The include below is the card's 5.5 GB set.
hf download Viggle/Viggle-Remake --local-dir ./Viggle-Remake --include "lora/*" "inference/*" "examples/*" "requirements.txt" - 5
Pick a card from COMFYUI.md, do not guess one
sample.py needs one 80 GB+ GPU at bf16. ComfyUI.md's table, also restated by the ComfyUI Wiki on 2026-11-04, is the smaller-card list. At 832×480, 12 GB is Q4_K_M or Q8_0 with --reserve-vram 8, and int8 is 124 frames only. 1280×736 on 12 GB was not tested. 16 GB can use int8, or Q4_K_M or Q6_K with --reserve-vram 6. This page does not add a tier the table does not print.
- 6
Keep one window inside the frame rule
One render is 90 to 345 frames, where the count is 17n+5, at 24 fps. The card recommends up to 294 frames, about 12 seconds. Longer films are windows. The MCP server cuts those windows, paints keyframes, writes prompts, renders and splices. The soundtrack is carried from the source. A silent source stays silent.
- 7
Write the six-section prompt
PROMPTING.md requires MiniMax-H3's six sections, in order: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music. Write a description of the target video, not an instruction to replace someone. retention_analysis is obeyed literally. Timestamps must fit the frame count. Free-form prompts are untested. Paint 1 to 4 frames. Pins are optional and, when used, one per painted frame.
Quantized transformer sizes on the card
Sizes are from COMFYUI.md, read on 2026-11-06. The edit twin of each file is the same size. Nano AI does not download them.
| File | Size | Loader they name | What Nano AI does |
|---|---|---|---|
| viggle_remake_swap_pruned_int8_convrot.safetensors | 21.0 GB | Load Diffusion Model | Nothing. Not loaded here. |
| viggle_remake_swap-Q8_0.gguf | 22.3 GB | Unet Loader (GGUF) | Nothing. Not loaded here. |
| viggle_remake_swap-Q6_K.gguf | 17.6 GB | Unet Loader (GGUF) | Nothing. Not loaded here. |
| viggle_remake_swap-Q4_K_M.gguf | 14.3 GB | Unet Loader (GGUF) | Nothing. Not loaded here. |
Choices the card actually prints
App, or your own GPU
The Viggle app is the no-setup path. sample.py is the 80 GB+ bf16 path. ComfyUI, int8 and GGUF are the smaller-card path, down to the 12 GB row.
Swap is not edit
lora/swap recasts people and characters. lora/edit is objects, background, weather and light. The ComfyUI sigmas differ by mode. Do not load a Remake LoRA on top of an already merged int8 or GGUF file.
144 GB is not 5.5 GB
The base include is about 144 GB. The LoRA and sampler include is 5.5 GB. Cloning with git-lfs is the about-210 GB path the card tells you to avoid.
A new clip is not a re-shoot
Kling 2.6 here is 55 credits for 5 seconds, sound: false. It does not keep a soundtrack from a file you upload, because it does not take that file.
Viggle Remake next to OPENjutsu and a new clip
Remake figures are from the card and COMFYUI.md on 2026-11-06. OPENjutsu figures are only what that page already states from READMEs checked on 2026-11-03. Kling prices are Nano AI's. This is not a quality ranking.
| Viggle Remake | OPENjutsu, as that page states | Kling 2.6 here | |
|---|---|---|---|
| What changes | People in swap; objects, weather and light in edit | Characters and objects, with SAM 3.1 tracking in the D3voz app | Nothing in an uploaded film |
| Where it runs | Viggle app, or your GPU | Your NVIDIA GPU, or Comfy's separate hosted beta | This site, as a new clip |
| A hardware line they print | 80 GB+ for sample.py; 12 GB row for Q4_K_M or Q8_0 | Tested RTX 4060 Ti 16 GB and 64 GB RAM, not a universal minimum | No local GPU |
| Sound | Source soundtrack, not generated | D3voz README: soundtrack retained | sound: false |
| Length | 90–345 frames, recommend up to 294 | Comfy hosted beta: 5–15 seconds in that README | 5 or 10 seconds |
| License of the weights | MiniMax H3 Community License | Model licenses separate from the MIT app | Not those licenses |
Pages this note was read from
Opened on 2026-11-06. No weight file was downloaded and no GPU job was run.
How to use this page
- 1
For a re-shoot, leave this site
Follow the card's hf download includes, or the Viggle app. Do not git clone the 210 GB tree. Paint only frames from footage you have the right to change.
- 2
For a new silent clip, open Kling 2.6
The button selects 5 seconds and 55 credits, with sound: false. It will not recast a person or remove an object from a file you already have.
- 3
Read OPENjutsu as a different guide
That page is the D3voz app and the Comfy hosted beta. It is not a second copy of these LoRAs, and this site still has no /genjutsu page.
Prompts for a new silent clip
Scenes for Kling 2.6. They are not six-section Remake prompts and they do not name a person to recast.
Empty field
An empty green field and a dirt track under moving clouds, one continuous shot, no people, no turbines, no readable text, no logos.
Workshop bench
A quiet workshop bench, tools still, one lamp, the camera holds, no people, no readable text, no logos.
Night window
Rain on a dark window over an empty room, camera still, no people, no readable text, no logos.
First frame
The camera slowly pushes toward the object in the first frame. Keep the object and the background. No captions, no logos, no people.
Consent, license, and what this page will not do
The limits below are copied. This site did not test a GPU and cannot check who is in a frame.
- Identity is the picture you paint
- The card says the model cannot verify identity or consent. The intended use is footage you have the right to change. It says to label the result as AI-generated, and points commercial hosts to Section V.5 of the MiniMax agreement. Do not use it on people who have not agreed.
- Weights are not Apache
- Code in inference/, examples/, comfyui/ and mcp/ is Apache 2.0 and covers scripts and prompts only. The weights, including int8 and GGUF, are a Model Derivative of MiniMax H3, so the Community License applies. Neither licence grants the third-party footage or the likenesses.
- US$20 million and the name on the UI
- Section IV.1 requires prior written authorization from MiniMax if commercial products and services generate more than 20 million US dollars a year. The card says to write to api@minimax.io with the subject MiniMax H3 licensing - authorization request. Section IV.2 requires the commercial UI to display MiniMax H3 prominently. The card also asks products to show Powered by Viggle Remake and the model URL.
- The license names excluded territories
- The Community License, read on 2026-11-06, defines Applicable Territory as worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States. It says people in those territories can contact MiniMax about a separate license. That sentence is in the license. It is not an extra rule added on this page.
- OPENjutsu is a different project
- The OPENjutsu page, from its README as checked on 2026-11-03, describes a MIT app over ComfyUI and MiniMax H3, with SAM 3.1 tracking, tested on an RTX 4060 Ti 16 GB and 64 GB of system RAM. That test is not a universal minimum. Remake's own table starts the bf16 sampler at 80 GB+ and the 12 GB row at Q4_K_M or Q8_0. This site has no /genjutsu page. Higgsfield Genjutsu stays the hosted product already described on the OPENjutsu page. No new Genjutsu claims are added here.
This site does not re-shoot a film.
Kling 2.6 makes a new silent clip, not a Viggle Remake
Nano AI does not run Viggle Remake and does not edit an uploaded movie. Kling 2.6 costs 55 credits for one new 5-second clip, with sound: false. It does not load swap or edit LoRAs, keep a source soundtrack, or paint frames. Eligible new accounts receive 64 credits. A 10-second clip is 110 credits.
FAQ
Viggle Remake FAQ
Download size, the two modes, the license, and the clip this site actually prices.
Does Nano AI run Viggle Remake?
No. It does not load the LoRAs, accept a film for swap or edit, or keep a source soundtrack. The button opens Kling 2.6 for 55 credits and 5 seconds, sound: false.
Should I git clone the repository?
No. The card says a git-lfs clone downloads about 210 GB. Use hf download with an include. The LoRA and sampler include is 5.5 GB. The base MiniMax-H3 include is about 144 GB.
What fits on 12 GB?
COMFYUI.md says, at 832×480, Q4_K_M or Q8_0 with --reserve-vram 8. int8 is 124 frames only. 1280×736 on 12 GB was not tested. Q4_K_M is 14.3 GB, Q8_0 is 22.3 GB, Q6_K is 17.6 GB, and int8 is 21.0 GB.
What is the six-section prompt?
subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, then non_diegetic_music. PROMPTING.md says free-form prompts are untested. Paint 1 to 4 frames. Pins are optional.
How is this different from OPENjutsu?
OPENjutsu's page describes D3voz's MIT app and Comfy's hosted beta, including a tested RTX 4060 Ti 16 GB with 64 GB RAM. Remake is two H3 LoRAs with its own VRAM table. Higgsfield Genjutsu is only the hosted product already noted there. This site has no /genjutsu page.
What does the license require?
Apache 2.0 covers the named code, not the weights. The Community License, dated August 2, 2026, requires prominent MiniMax H3 on a commercial UI, a copy of the license and the NOTICE file, and written authorization above 20 million US dollars a year. It also excludes the EU, the UK, Korea and the United States unless MiniMax grants a separate license. Label AI output. The model cannot check consent.
More from Nano AI
Related pages
Nano AI is an independent service and is not affiliated with Viggle, MiniMax, D3voz, Comfy or Higgsfield. Weights and commands are from the Viggle Remake card, COMFYUI.md, PROMPTING.md and the MiniMax H3 license as read on 2026-11-06. The linked generator runs Kling 2.6 and does not re-shoot footage.