MiniMax Workflow
An open-source AI video generation workflow tool built on ComfyUI and optimized for the MiniMax H3 model. It supports mixed multi-reference-source input, video continuation and splicing, and frame interpolation, providing a structured pipeline for AI-driven video production.
Core Facts
Timeline (last 90 days)
Reddit user roycho87 was blocked by the original author after posting critical feedback on a MiniMax workflow, subsequently reconstructed the entire project, and released an improved version on GitHub.
MiniMax Workflow is built on ComfyUI and optimized for the MiniMax H3 model.
The video continuation feature uses an overlap-splicing-based technical approach and synchronizes the audio track accordingly.
The workflow incorporates sparse attention mechanisms and a low-VRAM mode, using chunking technology to split computation tasks into blocks to alleviate VRAM pressure.
The workflow has a built-in RIFE frame interpolation module, allowing users to perform frame interpolation directly after generation to upscale low-frame-rate footage to a higher frame rate.
The workflow provides latent-space upscaling functionality, supporting fine-grained processing of the complete spliced video.
The workflow consolidates scattered control nodes into a unified control panel, and the system automatically adjusts internal routing based on the number of enabled reference types.
An early version had a bug where the final output was the first pass rather than the upscaled version, which the author fixed in V2.
V2 removes the shift parameter and instead provides an optional subgraph for connecting a turbo LoRA.
The workflow requires a placeholder image to be uploaded in order to function properly; this is a design trade-off made to avoid manual bypass operations.
1 more timeline events
All Facts (11)
The video continuation feature uses an overlap-splicing-based technical approach and synchronizes the audio track accordingly.
65%UnverifiedMiniMax Workflow is built on ComfyUI and optimized for the MiniMax H3 model.
50%UnverifiedThe workflow incorporates sparse attention mechanisms and a low-VRAM mode, using chunking technology to split computation tasks into blocks to alleviate VRAM pressure.
50%UnverifiedThe workflow has a built-in RIFE frame interpolation module, allowing users to perform frame interpolation directly after generation to upscale low-frame-rate footage to a higher frame rate.
50%UnverifiedThe workflow provides latent-space upscaling functionality, supporting fine-grained processing of the complete spliced video.
50%UnverifiedThe workflow consolidates scattered control nodes into a unified control panel, and the system automatically adjusts internal routing based on the number of enabled reference types.
50%UnverifiedAn early version had a bug where the final output was the first pass rather than the upscaled version, which the author fixed in V2.
50%UnverifiedV2 removes the shift parameter and instead provides an optional subgraph for connecting a turbo LoRA.
50%UnverifiedThe workflow requires a placeholder image to be uploaded in order to function properly; this is a design trade-off made to avoid manual bypass operations.
50%UnverifiedThe project's open-source address is https://github.com/roycho87/minimax_wf.
50%UnverifiedReddit user roycho87 was blocked by the original author after posting critical feedback on a MiniMax workflow, subsequently reconstructed the entire project, and released an improved version on GitHub.
50%