Viggle has opened the Viggle-Remake repository on Hugging Face: a pair of LoRA adapters (rank 256, alpha 256) on top of the base video model MiniMax-H3 (approximately 20 billion parameters, wan architecture) reshoots already finished video — in swap mode it replaces characters, in edit mode it changes objects, background, weather, and lighting, while preserving original motion, camera, timing, and sound. Key files (LoRA, sampler, MCP) weigh about 5.5 GB, and int8 and GGUF versions with ComfyUI nodes lower the entry barrier to GPUs with 12 GB of VRAM, without requiring paid API calls. For agents like Claude Code and Codex, there is an MCP server that automatically cuts the clip into windows, draws keyframes, renders, and stitches the result. Main caveats: one render window is calculated for approximately 294 frames, i.e., about 12 seconds, the release materials contain no quality metrics, and the model does not verify the consent of the people depicted.

image

What happened

Viggle published the Viggle-Remake repository on Hugging Face. Technical composition: two LoRA adapters with rank 256 and alpha 256, trained on top of the open video model MiniMax-H3 with wan architecture. The first adapter works in swap mode and replaces characters in a finished clip, the second — in edit mode and changes objects, background, weather, and lighting; in both cases, motion, camera, timing, and sound are taken from the source. In total, key files — the LoRAs themselves, the sampler, and the MCP server — occupy about 5.5 GB. The model card features demonstration examples: a clip based on La La Land and a reshoot of the Burly Brawl scene from The Matrix, which set the frame of expectations for the class of tasks.

Context

The scale of the solution becomes clear against the backdrop of the base: the full version of MiniMax-H3 in bf16 occupies about 144 GB and requires a GPU with 80+ GB of video memory, so Viggle chose the parameter-efficient path — two adapters instead of full fine-tuning of the 20-billion model. The adapters modify the reshoot relative to the original video stream, rather than generating a scene from scratch, which distinguishes the approach from classic text-to-video generation. The entry barrier is lowered by the release packaging: in addition to the main files, there are int8 and GGUF versions (Q4_K_M — 14.3 GB, Q6_K — 17.6 GB, Q8_0 — 22.3 GB) and ready-made nodes for ComfyUI, allowing operation from 12 GB of VRAM. Weights are distributed under the MiniMax H3 Community License, which imposes conditions on use. The role of the MCP server addresses a structural limitation of the base model — short windows: it divides a long clip into fragments and assembles them back, turning the pipeline into a task for coding agents like Claude Code and Codex.

Why this matters for the industry

For the industry, this is a shift in access method: video reshooting was previously a feature of paid APIs, now it is a local layer with open weights, where the main cost is your own GPU. A direct consequence is the automation of post-production by coding agents: replacing actors, changing locations, and weather are performed as a programmable pipeline via MCP, rather than manual editing. By this logic, simple reshoot operations are commoditized, and the protectable layer for products becomes the orchestration of long videos — cutting into windows, seam control, stitching — and distribution through third-party agent stacks. Teams get a window for cheap prototypes: reshooting footage for a client's task can be tested without renting a GPU with 80 GB of memory. The speculative horizon, outlined in release reviews: if the pattern of "open video base plus LoRA plus agent orchestration" becomes established, closed APIs will have to compete with local pipelines in the task class of "preserve motion — replace content."

Why this matters for users

To try reshooting, it is enough to download key files weighing about 5.5 GB or quantized versions, install ComfyUI nodes, and run generation on a GPU with at least 12 GB of VRAM. For manual mode, a prompt in the six-section H3 format and from one to four drawn reference frames are required. The alternative is to connect the MCP server to Claude Code or Codex: then a phrase like "redress the group as politicians" is enough, and the agent will itself divide the video into windows, draw keyframes, render, and stitch the result. A realistic scenario today is short fragments: one window is approximately 294 frames, about 12 seconds, and the longest published run is a two-minute clip, so long videos will require agent stitching or manual cutting. A separate hygiene rule: the model does not verify the consent of the people depicted, so reshooting should only be run on your own or agreed-upon material.

What is currently unknown / limitations

The release materials contain no quality metrics — identity preservation, temporal consistency, synchronization with sound — and no comparison with closed APIs; examples from La La Land and The Matrix are a curated demonstration, not an evaluation on heavy scenes. Image integrity at the seams of multiple render windows during agent stitching is an open technical issue, and the claim that the solution works with full films is confirmed only by a two-minute clip and is closer to marketing in status. The claimed operation of lightweight LoRAs of ranks 128 and 64 has not been verified by independent tests, as has the reliability of the agent orchestration itself: there is no data on artifacts at the seams and the percentage of failed runs in the release. Predictions about community ablations and quantizations, "reshoot by phrase" wrapper services, and standardization of agent pipelines on the horizon of six months to two years remain interpretations, not established facts.

Sources

Author

Look at AI, editorial team