ComfyUI-Mova node
The ComfyUI custom node is not finished and still needs real work. What it already does well: it can run FP8 weights and has a workable low-VRAM / high-VRAM strategy (offload, streaming loads, expert swap).
Official MOVA defaults are about cfg 5 and 50 steps. On my hardware that was roughly 19 minutes for a 7 second clip.
LightX2V isn’t an official MOVA stack, it’s a Wan-style 4-step video distill. For experiments I baked it into the FP8 models anyway.
That path is much faster for me (45 seconds for 7s), but quality and audio are the weak spots. After a lot of back-and-forth (local Qwen3.6 coding assistant + Qwen3.7-plus) and still didn’t get a solid audio path (joint speech + sync, not just video with sound).
The video below is current results with the FP8 LightX2V models, take it as 'what this node can do today, not a polished fully working node'.
If you want a headache (or a starting point), this custom node is a reasonable backbone: Comfy wiring + memory plumbing are there,
generation quality, audio, and LightX2V behavior still need love. If you fix something, please share it.
Screenshot
360p and 720p
| Pair | Same architecture? | Same weights? |
|---|---|---|
| 360p high vs 720p high | Yes (1095 tensors) | No |
| 360p low vs 720p low (BF16) | Yes (1095 tensors) | No |
shared across 360p and 720p
These components are byte-identical between variants:
- Video VAE
- Audio VAE
- Tokenizer (
tokenizer/) - Scheduler (
scheduler/) - Text encoder
/text_encoder/umt5_xxl_bf16_ComfyUI.safetensors(Load CLIPwantype)
Keep separate per resolution
| Component | 360p | 720p |
|---|---|---|
| High-noise video DiT | mova_360p_high_noise | mova_720p_high_noise |
| Low-noise video DiT | mova_360p_low_noise | mova_720p_low_noise |
| Dual-tower bridge | mova_360p_dual_tower_bridge | mova_720p_dual_tower_bridge |
| Audio DiT | mova_360p_audio_diffusion | mova-720p-audio-diffusion |