The decision: use native ComfyUI with the Wan 2.2 TI2V-5B workflow. It is the best balance of quality, VRAM fit, and maintained setup for this machine.
why it fits
The desktop has an RTX 4070 with 12 GB of VRAM, a Core i7-14700KF, 32 GB of RAM, and FFmpeg already installed. ComfyUI's native Wan 2.2 5B workflow is documented to fit within 8 GB of VRAM through model offloading, leaving useful headroom on this card.
The workflow produces image-conditioned 720p video at 24 FPS without custom nodes. That makes it a more practical starting point than current 14B–22B video models, which would require aggressive quantization, heavier offloading, and much longer generation times.
the two blockers
- NVIDIA driver mismatch. Updated 595.84 userspace components were paired with a running 595.71.05 kernel module, so
nvidia-smifailed. Reboot first, then verify the driver before installing anything. - Disk space. Only about 11 GB was free. The three required model files total roughly 18.15 GB before ComfyUI, PyTorch, downloads, and outputs. Free at least 35 GB; 50 GB is comfortable.
the short-clip workflow
- Install ComfyUI and its native Wan 2.2 TI2V-5B model, text encoder, and VAE.
- Open the Wan 2.2 5B video template and enable its bypassed
Load Imagenode. - Use 73 frames for about three seconds or 121 frames for about five seconds at 24 FPS.
- Write a motion-first prompt describing subject movement, environmental movement, and camera behavior.
- Start with the template's 1280×704 landscape or 704×1280 portrait size and expect minutes—not real time—per result.
alternatives
FramePack has the simplest dedicated upload-and-prompt interface and supports relatively low VRAM, but its model downloads exceed 30 GB and its design is aimed more at progressively generating longer clips.
LTX-Video 2B Distilled is smaller and faster, making it useful for previews, but the expected quality is below Wan 2.2 5B. HunyuanVideo 1.5 lists a 14 GB minimum with offloading, so it misses this card's 12 GB ceiling.
result
No software was installed during the conversation. The correct next move was to reboot, confirm nvidia-smi, free storage, and only then install ComfyUI and the three Wan files.