Wan 2.2 I2V at 720p on 24GB — the settings that actually rendered, and why the "faster" 4-step graph is slower

Disclosure first: I built and sell this. I am not posting benchmarks I did not run. Every number below came off my own machine, and the graphs are the ones I tuned.

Most Wan 2.2 ComfyUI examples I found were repackaged "example workflow" JSON with default values — 832x480, 33 frames, whatever shipped in the repo README. This is the opposite: each parameter is set to a value I verified, with the reasoning written next to it.

WHAT IS IN THE PACK

wan22_i2v_720p_high_low — Wan 2.2 I2V, 1280x720, 81 frames at 16fps (5.06s of video), dual-model HIGH to LOW sampling across two KSamplerAdvanced stages. Core ComfyUI nodes only, no custom node packs to install and debug.

wan22_i2v_720p_4step_lightning — the same graph plus the lightx2v 4-step LoRA at cfg 1.0. I included this one specifically because it is counterintuitive: on my card it is SLOWER than the 8-step graph, and the README explains why, and when it is still the right pick.

infinitetalk_talking_head_480p — audio-driven talking head at 832x480 on ComfyUI-WanVideoWrapper, blocks_to_swap=20 so it fits inside 24GB.

scripts/run_workflow.py — queue any workflow from the CLI, no UI required.

samples/ — real renders from each workflow, so you can compare the 4-step and 8-step output yourself before you commit to either.

README — the exact verified download source for every model file.

WHY THESE PARAMETERS

81 frames is 4n+1, which is what Wan's temporal packing requires. The HIGH to LOW handoff splits at exactly half the step budget because the LOW model is trained on low-noise steps. cfg 3.5 is where motion stops crawling on I2V. shift 8.0 on ModelSamplingSD3. The README explains each one and what to change when you want a different look, including what to touch first if you are not on 24GB.

MEASURED, NOT GUESSED (RTX 5090 Laptop, 24GB, ComfyUI 0.30.x, PyTorch 2.11 cu128)

720p 8-step HIGH to LOW: 1135s for 81 frames (5.06s of video at 16fps). 720p 4-step Lightning LoRA: 1731s. Yes, slower than 8-step on this card. VRAM is saturated so the sampler is memory-bound; fewer steps does not mean faster here. InfiniteTalk 480p: 1080s for a 15.72s talking-head clip, windowed with frame_window_size=212.

That 4-step result is the single most useful thing in this post. If you have been reading that 4-step Lightning is "the fast one," that is only true when you are not VRAM-bound.

REQUIREMENTS

ComfyUI 0.30.x, PyTorch 2.11 (cu128 build used here). About 24GB VRAM for the 720p workflows — drop to 832x480 below that, the README says how. Sage attention recommended: python main.py --use-sage-attention --disable-smart-memory.

Models are NOT included. They are multi-GB and separately licensed. Every one is listed with its exact Hugging Face repo and filename.

Full pack ($12, itch.io): https://adel93.itch.io/wan22-comfyui-workflow-pack

I would rather answer a question here than have you buy something that does not fit your card. If you tell me your GPU and resolution, I will tell you whether this is worth it to you, including if the answer is no.