RelayVSR + DLSS5 ComfyUI workflow upscales MiniMax H3 video to 2K in 7 min on 6GB VRAM
cgpixel23 · reddit · 2026-11-10
A detailed ComfyUI tutorial for fixing MiniMax H3 video upscaling detail loss using RelayVSR combined with NVIDIA DLSS5 nodes, runnable on GPUs with only 6GB of VRAM.
- How RelayVSR works: a large generative model creates high-quality reference latents for keyframes, then a lightweight Dual-Memory Video Transformer reconstructs remaining frames from these references, preserving original structure and character appearance instead of hallucinating new details.
- Timings: 35 min to generate at 0.7MP with MiniMax H3, 7 min to upscale to 2K with RelayVSR + DLSS5, versus 45 min for MiniMax latent upscaling.
- Stack: optimized MiniMax H3 workflow, Pika H3 6-step LoRA, attention optimization nodes, DLSS5.
Project page, Hugging Face weights, workflow files and a video tutorial are all linked for direct reproduction.
More from Multimodal
- MiniMax H3 Image generates a stunning exploded-view layout of an FPV racing drone — aftahi_ai · 2026-11-10
- Cyber Pastel Sanctuary: a reusable AI image prompt template with glitch and VHS aesthetics — LudovicCreator · 2026-11-10
- Dev builds pickleball game with Fable 5.2 and Unreal Engine 5.9's AI-first workflow — chongdashu · 2026-11-10
- Adobe adds Generative Remove to Premiere, plus AI cleanup tools across Photoshop and Lightroom — The Verge AI · 2026-11-10
- One video, 24,000 flying pages: a creator showcases Fable 5.2's cinematic take on a moonlit library — oran_ge · 2026-11-10
- RTX 5060-generated prank video of Nvidia CEO Jensen Huang — samwick47 · 2026-11-10