EU users can't run H3, so this dev made Wan 2.2 2x faster: 15:31 to 8:49 on RTX 5070
noise_to_motion · reddit · 2026-10-28
- Why: MiniMax H3's license excludes the EU (plus UK/US/South Korea), so a Lithuania-based dev optimized Apache-2.0 Wan 2.2 image-to-video on an RTX 5070 12GB instead.
- Speed gains (same PC, same seeds, ComfyUI logs): 832×480/81 frames from 15:31 to 8:49; 960×544 from 23min to 10:47; 1536×864 previously stalled for hours, now 29:55.
- Three key techniques:
- SageAttention (--use-sage-attention): 1.6x faster per step; on Windows portable you need triton-windows and matching torch/CUDA wheels.
- Split CFG: full CFG 1 skips the negative prompt and kills prompted action, so steps 0–6 at CFG 1.7, steps 6–12 at CFG 1.0.
- ComfyUI 0.40's dynamic VRAM loading fixed the 12GB high-res stall.
- Quality: same scene/characters/style with fixed seeds; 5/255 average late-frame drift, visually not worse. Unlike 4-step LoRAs (10x faster but mushy faces), this keeps full 24-step fidelity.
- Validation: 170 user renders over two weeks averaged 22.6min to 10min; SageAttention self-tested on GPUs from 3060 Ti to 4090. Author builds Loomance but all settings work in plain ComfyUI.
More from Multimodal
- VEED ships dubbing model API: accurate translation, lip-sync, multi-speaker support — hey_abusiddik · 2026-10-28
- VEED launches dubbing API with built-in lip sync that preserves emotion and accent — hey_abusiddik · 2026-10-28
- Mundo turns photos of real places into explorable, editable interactive worlds — Scobleizer · 2026-10-28
- Google's new AI world model reportedly simulates entire worlds in real time — heykahn · 2026-10-28
- NVIDIA open-sources Nemotron 3.5 Transcribe 30B ASR model; fine-tuning cuts word error rate 45.8% — NVIDIAAI · 2026-10-28
- NVIDIA open-sources Nemotron 3.5 Transcribe: 30B MoE ASR model with ~4B active params — NVIDIAAI · 2026-10-28