MiniMax H3 on RTX 4070 Ti SUPER: Generates 5-Second Video Clips in ~5 Minutes
EdenAlon · reddit · 2026-08-05
A developer shared an optimized ComfyUI workflow for running MiniMax H3 image-to-video generation efficiently on an RTX 4070 Ti SUPER (16GB VRAM).
Performance
By utilizing the official pruned INT8 ConvRot model, quantized Qwen3-VL text encoder, and SageAttention, the workflow performs exceptionally well on 16GB VRAM. Generating a 5-second video (158 frames) at 0.7 megapixels (720p class) takes just under 7 minutes without triggering CUDA OOM.
Workflow Features
The JSON workflow is deliberately clean, removing upscaling and interpolation stages. Users simply provide a starting image and a prompt to generate videos, with an optional LoRA slot included. The author notes that the output quality is high enough that post-processing is currently unnecessary.
Related event: MiniMax H3 Local Tests: Performance and Optimization Across GPUs(13 posts)→
More from Multimodal
- Tessellation-free geometric queries on displaced surfaces cut caustics variance 10x — ssh4net · 2026-10-16
- Complex multi-link CAD mechanism finally animates, dev celebrates the breakthrough — zetalyrae · 2026-10-16
- OpenMOSS releases MOSS-Image-SFT: Apache 2.0 text-to-image model on Hugging Face — bdsqlsz · 2026-10-16
- Hyperflow open-sourced: 49 sampling steps down to 8, ~3x faster video generation — mhdfaran · 2026-10-16
- Cinematic 3D fantasy character prompt template with salvaged-robot armor — azed_ai · 2026-10-16
- Midjourney vintage woodcut style: --sref 2446102396 beaver with a cleaver — michaelrabone · 2026-10-16