Hardware guide
Shotcoda was built and measured on an ordinary desktop, not a render farm. Here is what it needs, what each part does and how fast it runs, using measured numbers only.
Same quality as the reference. Motion takes roughly twice as long.
48 min of machine time for a 62.6 s Short (fast profile).
The most headroom. Best when you publish every day.
Not suitable: GPUs under 12 GB of VRAM for the main motion model, AMD and Apple GPUs (the stack is CUDA), and Windows or macOS without a Linux install. Not sure about your machine? Send your specs and you'll get a straight answer.
What each part does
Measured
Measured on the reference machine (2 × RTX 5060 Ti 16 GB, Ryzen 5 7600X, 32 GB RAM) on 14 September 2026. Motion is the bottleneck, at about 3.5 seconds per generated frame.
| Stage | GPU 0 | GPU 1 | System RAM | Time |
|---|---|---|---|---|
| Keyframes (Z-Image, 1088×1920) | ≈ 14.6 GB | — | ≈ 12 GB | 26 s / image |
| Narration (Qwen3-TTS 1.7B) | — | 5–7 GB | ≈ 4 GB | 5–9 s / take |
| Music (ACE-Step XL turbo) | — | ≈ 12 GB | ≈ 10 GB | 24–37 s / track |
| Motion, high-noise (Wan 14B fp8) | ≈ 13.7 GB | — | ≈ 27 GB with both | 86–237 s / clip |
| Motion, low-noise + decode | — | ≈ 13.5 GB | shared | 118–272 s / clip |
| RIFE interpolation (iGPU) | — | — | < 1 GB | 12–28 s / clip |
| Render (4 Chrome workers) | — | — | ≈ 3 GB | 2.3–3.5 min |
Fast profile, 62.6 s Short, whole pipeline
Of that, motion: 12 clips, 584 frames
Model storage for video, image, voice and music
Consumer GPUs, no NVLink, second card on PCIe x4
Hardware FAQ
Yes. One NVIDIA card with 16 GB of VRAM gives the same quality as the reference machine. The high-noise and low-noise motion experts then share one card, so motion takes roughly twice as long, and 64 GB of system RAM is recommended.
Not for the main motion model. Wan 2.2 A14B needs about 14 GB per expert. Smaller cards can run lighter models, such as Wan 2.2 TI2V-5B or quantised image models, with clearly lower motion quality.
Not today. The video, voice and music stack is built on NVIDIA CUDA with fp8 weights. ComfyUI itself runs on Apple Silicon, but fp8 Wan and the CUDA TTS stack do not, so macOS is untested.
Linux. The reference machine runs Ubuntu 26.04 with the NVIDIA 595 driver and CUDA 13.2. A recent Ubuntu with a current NVIDIA driver is the target for new setups.
About 89 GB for the video, image, voice and music models, plus about 10 GB for the local director LLM and 1–3 GB per episode. Plan for at least 150 GB free on an NVMe SSD.
No. Only a latent file passes between the two motion experts, so the second card works fine on a PCIe 4.0 x4 slot. The reference machine uses x8 + x4.
Electricity for the hours the GPUs work, which is under an hour per Short on the reference machine. There are no per-video fees. If you use Claude or ChatGPT as the director instead of the local LLM, that plan is the only subscription.
Build my video pipeline
Tell me what you make and where you publish. I'll check your hardware, build the pipeline around your format and brand, and set it up on your machine. No pricing tables: every setup is different, so you get a straight answer.