Can't use H3 in the EU, so I made Wan 2.2 twice as fast instead: 15:31 → 8:49 on a 12GB RTX 5070 (same-seed comparisons + real user data)
retained text
Everyone here is playing with MiniMax H3, but I'm in Lithuania, and H3's license excludes the EU (also the UK, US and South Korea). So I've been squeezing everything I can out of Wan 2.2 image-to-video instead, which is Apache 2.0 and usable anywhere. Same PC, same workflow, same seeds, times straight from ComfyUI's log (RTX 5070 12GB, 64GB RAM, ComfyUI 0.40): \- 832×480, 81 frames: 15:31 → 8:49 \- 960×544, 53 frames: \~23 min → 10:47 \- 1536×864, 57 frames: used to stall for hours on the old ComfyUI build → now 29:55 What made the difference: 1. SageAttention (--use-sage-attention): about 1.6× faster per step. On Windows portable you need triton-windows, the Python include/libs folders and a wheel that matches your torch/CUDA. 2. CFG 1, but only in the later high-noise steps. At exactly CFG 1 ComfyUI skips the negative prompt, so each step is 2× faster. CFG 1 everywhere killed the action in my test (a woman never took the sip I prompted), because the first steps decide what happens. So: steps 0–6 at CFG 1.7, steps 6–12 at CFG 1.0 (add\_noise disabled), low-noise steps unchanged. 3. ComfyUI 0.40: its dynamic VRAM loading fixed the 1536×864 stall on 12GB with no extra flags. Does it look worse? Not that I can see. With a fixed seed you get the same scene, characters, style and colours. Not pixel-identical (both SageAttention and the CFG split drift late frames by \~5/255 on average), but not worse. "Just use a 4-step LoRA": yes, lightx2v-style 4-step setups are about 10× faster. In my side-by-side tests they cost fidelity (faces and characters drifted, fine detail got mushy), so this is for people who want the full 24-step quality, just faster. Comparison: [https://loomance.app/blog/same-wan-2-2-fast-mode-vs-full-workflow.html](https://loomance.app/blog/same-wan-2-2-fast-mode-vs-full-workflow.html) Not just my PC: across \~170 renders from users of my app on different GPUs over two weeks, the average render time dropped from 22.6 min to about 10 min. Mixed settings, so it's an average, not a benchmark. SageAttention passed a self-test on RTX 3060 Ti, 3070 Laptop, 4060 Laptop, 4060 Ti, 5060, 5070 and 4090. Haven't tried torch.compile or the new --use-ck-attention yet. If you have numbers for those, I'd love to compare. Full write-up with logs, exact KSampler settings and frame comparisons: [https://loomance.app/blog/wan-2-2-2x-faster-sageattention-split-cfg.html](https://loomance.app/blog/wan-2-2-2x-faster-sageattention-split-cfg.html) (Disclosure: I build Loomance, a free local app that uses these settings. Everything above works in plain ComfyUI too.)
document fetched 2026-10-29 00:43 · canonical https://reddit.com/r/comfyui/comments/1xhp7j6/cant_use_h3_in_the_eu_so_i_made_wan_22_twice_as