Wan 2.2 I2V on a 12GB card only used 8–9GB of VRAM. System RAM was the real bottleneck (measured, 3 renders)
retained text
I wanted to know whether Wan 2.2 image-to-video is realistic on smaller cards, so I watched where the memory actually goes during real renders. **Setup:** RTX 5070 12GB (PCIe 5.0 x16), 64GB DDR5-6000, ComfyUI 0.37 with dynamic VRAM loading. Wan 2.2 I2V A14B fp8 (high + low noise), lightx2v LoRA, 12+12 steps, RIFE interpolation on, no upscale. All runs back to back, models already warm. | Resolution | Frames | Time | Dedicated VRAM | Shared GPU mem | System RAM | |---|---|---|---|---|---| | 832×480 | 49 | 8m 28s | ~9.0 GB | 14.8 GB | 32 GB | | 832×480 | 81 | 17m 30s | 8.0 GB | 13.7 GB | 29 GB | | 1024×576 | 49 | 14m 00s | 8.3 GB | 11.8 GB | 28 GB | What surprised me: - The models are ~34GB in total (high 13.9 + low 13.6 + UMT5 6.4), but the GPU never went above ~9GB. ComfyUI keeps the rest in RAM and streams it in. - Windows showed ~40GB committed during renders. So I'd expect 32GB RAM to start paging to disk, and 16GB to be painful. - 1.65× more frames (49 → 81) took 2.07× longer. RIFE interpolating 80 extra frames took ~10 seconds, so short clips + interpolation is much cheaper. - The high-noise stage was ~2.7× slower per step than the low-noise stage in every run. Not sure why yet. - First render after launch was 3–4× slower per step while models got loaded and pinned. Big caveat: this is a 12GB card. My guess is that 8GB cards work at 480p if you have enough RAM, but PCIe x8 cards (like the 4060) and DDR4 systems should be noticeably slower, since everything gets streamed over that link. **8GB owners: what do you actually get?** GPU, RAM amount/type, resolution, frames and time would be super useful. Next I'm running the same measurements with MiniMax H3 on this card, curious whether it behaves the same way. Disclosure: I'm building Loomance, a free simplified front-end for ComfyUI (one-click install of these Wan 2.2 workflows). The tests above are plain ComfyUI + Wan 2.2, so the numbers apply to any setup.
document fetched 2026-10-20 00:41 · canonical https://reddit.com/r/comfyui/comments/1x9udxg/wan_22_i2v_on_a_12gb_card_only_used_89gb_of_vram