MiniMax H3 Workflow
An open-source ComfyUI-based workflow for MiniMax video generation. It supports mixed multi-reference types, video continuation, frame-level control, and LoRA loading, and provides a centralized control panel to simplify operations.
Core Facts
Timeline (last 90 days)
When extending generation to the second segment, the motion context must be enabled and the latent data numbered 1 must be loaded, with the second generation output changed to number 2.
This ComfyUI workflow is adapted from the earlier 'dual-ten sampling' workflow, and raises the 15-second limit to 30 seconds, 40 seconds, or even longer through a segmented continuation approach.
At 700,000 pixels running 10 seconds, it cannot run without tiling enabled; after enabling tiled processing and splitting into two tiles, it can scale up to 800,000 pixels and 10 seconds.
The author recommends splitting into two tiles when VRAM allows, as this is relatively faster.
Through tiled processing and nodes that free up VRAM occupied by the CLIP model, this workflow can run on a graphics card with 6GB of VRAM.
The most core module of the workflow is the motion context, which determines whether the video segments generated in sequence can be seamlessly connected.
When generating the first video segment, the motion context module must be disconnected and disabled, an FL-prefix model must be selected for image-to-video generation, and the save latent node number must be set to 1.
This workflow can generate continuous videos up to 60 seconds long on a 12 GB GPU.
Normal mode runs the full-precision base diffusion model with 20-step rendering, while Turbo mode reduces the scheduler to 8 steps and simultaneously scales the LoRA distillation weights.
The author extensively employs int8 and int4 quantization and pruning techniques to trade precision for reduced VRAM usage, enabling a 12 GB GPU to handle long video generation.
4 more timeline events
All Facts (14)
This workflow can generate continuous videos up to 60 seconds long on a 12 GB GPU.
65%UnverifiedWhen extending generation to the second segment, the motion context must be enabled and the latent data numbered 1 must be loaded, with the second generation output changed to number 2.
50%UnverifiedThis ComfyUI workflow is adapted from the earlier 'dual-ten sampling' workflow, and raises the 15-second limit to 30 seconds, 40 seconds, or even longer through a segmented continuation approach.
50%UnverifiedAt 700,000 pixels running 10 seconds, it cannot run without tiling enabled; after enabling tiled processing and splitting into two tiles, it can scale up to 800,000 pixels and 10 seconds.
50%UnverifiedThe author recommends splitting into two tiles when VRAM allows, as this is relatively faster.
50%UnverifiedThrough tiled processing and nodes that free up VRAM occupied by the CLIP model, this workflow can run on a graphics card with 6GB of VRAM.
50%UnverifiedThe most core module of the workflow is the motion context, which determines whether the video segments generated in sequence can be seamlessly connected.
50%UnverifiedWhen generating the first video segment, the motion context module must be disconnected and disabled, an FL-prefix model must be selected for image-to-video generation, and the save latent node number must be set to 1.
50%UnverifiedThe author extensively employs int8 and int4 quantization and pruning techniques to trade precision for reduced VRAM usage, enabling a 12 GB GPU to handle long video generation.
50%UnverifiedThe author states that this long-video workflow is not intended to replace the existing 30-second solution, as generating 60-second videos on a 12 GB GPU is more time-consuming.
50%UnverifiedThis workflow trades generation time for longer and smoother finished videos, representing a time-for-time tradeoff.
50%UnverifiedA minimum generation time of 35 minutes means this workflow is not suitable for rapid iteration scenarios, and is better suited for users with a clear need for long-shot content creation.
50%UnverifiedNormal mode runs the full-precision base diffusion model with 20-step rendering, while Turbo mode reduces the scheduler to 8 steps and simultaneously scales the LoRA distillation weights.
50%UnverifiedThe workflow has been published on the CivitAI platform, with a Mega backup download link also provided.
50%