MiniMax H3 Singularity V2 on a 6 GB GPU: Text2Video, Img2Video, MultiRef + Audio and 4K upscale (free workflow)
retained text
I’ve been testing **Singularity V2**, the MiniMax H3 fine-tune, on a **6 GB VRAM** card. It runs better than I expected, so I put everything into one workflow. Free Workflow in the Youtube video description: [https://www.youtube.com/watch?v=opoqT6fB\_Nk](https://www.youtube.com/watch?v=opoqT6fB_Nk) **What’s inside:** * **Text2Video:** structured H3 prompt (style, subject, setting, camera, action, dialogue, soundscape) * **Img2Video:** animate any image with native audio * **MultiRef + Audio2Video:** character + environment images plus your own voice clip, with solid lip sync * **Ref2Video:** up to 4 reference images, each with its own role (great for product ads) * **4K upscale + frame interpolation:** 720p → 4K, 24 → 48 fps **My times on 6 GB VRAM:** * Text2Video: 0.6 MP in \~18 min * Img2Video: 8 s at 0.6 MP in \~25 min * MultiRef + Audio: 10 s at 0.6 MP in \~35 min * Ref2Video: 8 s at 0.8 MP in \~45 min **What makes it fit in 6 GB:** * Pruned int8 / w4a8 / GGUF builds of Singularity V2 * H3 SLA Attention + Low VRAM Attention + Chunk FeedForward * TaoMate 3-step LoRA for speed, plus a realism LoRA (trigger: `R34L1SM`) **Heads-up on missing nodes:** if some nodes show red and the Manager says nothing is missing, the source is just mislabeled. H3SLAAttention comes from **ComfyUI-PlagueKind-Nodes**, and Low VRAM Attention / Chunk FeedForward come from **ComfyUI-KJNodes**. Install both and restart. Happy to answer questions or help troubleshoot in the comments. I’d also love to see what you make with it!
document fetched 2026-11-12 00:43 · canonical https://reddit.com/r/comfyui/comments/1xtj5xr/minimax_h3_singularity_v2_on_a_6_gb_gpu