Under 5-Minute MiniMax H3 Video on RTX 3090: ComfyUI Latent Upscale Workflow

AI video creator Dead Sun (@deadsun_0) has shared an efficient 'H3 Latent Upscale' ComfyUI workflow that enables rendering high-quality MiniMax H3 video clips in under 5 minutes on a single consumer RTX 3090 GPU. The approach offers an optimal balance between rendering speed and visual fidelity while bypassing the heavy VRAM overhead of native high-resolution diffusion.

Image source: @deadsun_0 on X

Generating native high-resolution 720p or higher video clips directly with MiniMax H3 on a single 24GB GPU typically takes anywhere from 10 to 25 minutes per clip, frequently risking VRAM limits. Dead Sun's pipeline sidesteps this bottleneck by first rendering the video at a low-overhead draft resolution of 480p (0.4 megapixels), then passing the latent representation through a specialized neural upscaler to output an 832p (1.2 megapixel) video—completing an 8-second clip in under five minutes.

Core Architecture and Custom Node Setup

The workflow centers around chaining the open-source custom node Comfyui_Minimax_h3_latent_Upscaler-Plus, developed by xmarre, directly after the initial sampling stage.

1. Install Custom Node: Install the node package via ComfyUI Manager or by cloning it into your custom_nodes directory.
   • Repository: Comfyui_Minimax_h3_latent_Upscaler-Plus (GitHub)
2. Load the Template Workflow: Import the R2V (Render-to-Video) workflow template JSON provided in the repository's workflow_templates directory.
3. Configure the Two-Stage Resolution:
   • Stage 1 (Initial Generation): Set the draft generation resolution to 0.4 mp (approx. 480p) to keep compute load minimal.
   • Stage 2 (Latent Upscale): Configure the latent upscaler node's target_megapixels parameter to 1.2 mp, which produces a final upscale between 780p and 832p.

Much like the proven latent upscaling workflows established in the LTX 2.5 ecosystem, offloading early diffusion steps to a lower resolution dramatically accelerates turnaround without sacrificing output sharpness.

Practical Considerations and Prompting Tips

When putting this workflow into production, creators should keep two key technical nuances in mind:

• Facial Detail Degradation on Wide Shots: Because wide shots position human subjects at a distance, rendering at 480p allocates fewer initial pixels to facial features and eyes. If these areas blur during the initial stage, latent upscaling cannot fully reconstruct lost micro-details. While the workflow excels at close-ups and medium portrait shots, scenes demanding distant full-body clarity may require increasing the initial resolution at the expense of longer render times.
• Prompting for Natural Eye Movement: By default, MiniMax H3 character motion can occasionally exhibit an unnerving, unblinking stare toward the camera. To ensure realistic human expression and eye dynamics, explicitly include the phrase natural blinking in the generation prompt.

Original source

• Dead Sun (@deadsun_0) on X: H3 Latent Upscale Workflow on RTX 3090
• xmarre GitHub Repository: Comfyui_Minimax_h3_latent_Upscaler-Plus