SuperUltimate VACEFUN Long Video

If SuperUltimate VACE Long Video is the all-in-one long-video node, this is its sibling for people running Wan 2.2 - and "Fun" is doing a lot of work here. Wan 2.2 is a Mixture of Experts: a high-noise expert that handles motion and composition, and a low-noise expert that refines detail. Most ComfyUI workflows hide that behind a loader. This node doesn't hide it. It gives you both models as separate inputs and separate sampler settings, which is precisely what the community settled on as best practice for 2.2.

Why you'd reach for it

Two reasons. First, you're on Wan 2.2 VACE (or the 2.2 family generally) and you want the long-video machinery - multi-round extension, refine-init quality preservation, control videos, loopback - that the pack's other node gives you on a single model input. Second, you want the two-pass control: the Wan 2.2 speed-LoRA PSA was a whole saga about how speed LoRAs destroy the high-noise pass but are fine on the low-noise one, and this node's split is exactly what that advice needs. High-noise stays full quality, low-noise gets the fast LoRA.

How it works

The plumbing is identical to VaceLongVideo - same vace_prompt_list round chain, same control list handling, same refine of crossfade frames, same loopback crossfade. The difference is the sampling core. Instead of one model and one sampler call, it runs a high-noise pass on model_h for steps_h steps, then a low-noise pass on model_l for steps_l steps starting from that result (the second pass disables noise and force-runs to completion). Total steps are steps_h + steps_l.

That means:

• model_h - the high noise model (tooltip says so). Motion, camera, composition.
• model_l - the low noise model. Detail, texture, faces.
• Separate steps_h / cfg_h / sampler_h / scheduler_h and the _l twins, one set per expert.
• model_override on the Prompt Combine nodes is explicitly "no use for fun vace" - the tooltip says it, and the code ignores it, so don't expect per-round model swaps here.

Everything else you already know from the sibling: vace_prompt_list, loopback_crossfade, optional vace_control_list, optional nag_params, and an IMAGE output that's the full stitched video.

Install

Same pack, same drill:

cd ComfyUI/custom_nodes
git clone https://github.com/bbaudio-2025/ComfyUI-SuperUltimateVaceTools

Restart ComfyUI (or install via Manager by searching "SuperUltimateVaceTools"). No extra dependencies beyond your Wan 2.2 VACE weights in the ComfyUI model folders.

Troubleshooting

• You wired the wrong model into the wrong socket. model_h/model_l are labeled, but if the high-noise socket gets the low-noise expert you'll get soft, drift-heavy video. Check the input, not the workflow.
• "identical frame_position detected" / overlapping control errors - same control-list validator as the sibling node; fix the positions, don't fight it.
• It's still VACE-slow. Two passes on two experts isn't double speed, it's double work if you run both at full quality. The whole reason the two-expert split exists is so you can cheapen the low-noise pass.
• Update left you with a stale node? Right-click → Fix node (recreate) - the author's standard fix.

If you're on Wan 2.1 VACE (a single 14B checkpoint, no MoE), don't reach for this - use the plain SuperUltimate VACE Long Video instead. This one needs two experts to shine.