Jakkanna SCAIL-2 Flow-UniPC

SCAIL-2's default ComfyUI workflow samples with the plain euler/simple topology and six steps - fast, but that's the "Fast" path. When you want the higher-quality render, the Jakkanna Production workflow swaps in Jakkanna SCAIL-2 Flow-UniPC, and this node is that swap: it builds a Flow-UniPC sampler and its matching sigma schedule in one go.

If you've used UniPC (the UniPC sampler from the diffusers/SD world), this is the same family, reworked for flow matching and for SCAIL-2's Wan 2.1 base. The pack vendors its own implementation of the scheduler, converted for flow prediction, rather than depending on your ComfyUI install having the right one. That's the whole point of shipping it as a node: the sampler and the sigma schedule are generated together so they can never drift out of alignment.

What you set

Two inputs, and honestly the defaults are the settings:

• steps - 40 by default. That's the Production workflow's chosen quality/speed point.
• shift - 3.0 by default. Shift warps the flow-matching sigma schedule; higher shift concentrates sampling steps near the start of the denoise where the big structural decisions get made. The default 3.0 is what the validated workflow uses, so treat it as tuned unless you have a reason to move it.

The node returns sampler (a SAMPLER) and sigmas (a SIGMAS). Wire them into a SamplerCustomAdvanced alongside the latent and the noise. That's the whole graph contribution - you're replacing what a KSampler would normally build internally.

Why a dedicated node instead of KSampler

Mechanism matters here. The production workflow wants the exact 40-step Flow-UniPC topology recorded in its reproducibility manifest, and it wants the sampler and schedule to be provably the ones validated. The built-in sample_flow_unipc checks that the sigma schedule you feed it matches the one the sampler expects - shape and every value - and raises if it doesn't. That's the "matching upstream sigma schedule" guardrail, and it's exactly the class of silent drift this pack was built to kill.

The companion nodes are the rest of that story: Jakkanna SCAIL-2 Upstream Noise supplies the seeded noise tensor separately (so the seed isn't buried inside a KSampler), and Jakkanna SCAIL-2 Validate Sampler cross-checks the whole sampling configuration against the manifest before the run starts.

Install

This is a single node in the teenu/ComfyUI-Jakkanna pack, so install once and it's there. ComfyUI Manager: search Jakkanna → Install → restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/teenu/ComfyUI-Jakkanna.git
cd ComfyUI-Jakkanna
pip install -r requirements.txt

Then restart. Jakkanna replaces the upstream vnccs-utils package - don't install both. The heavy lifting is SCAIL-2's own model files (the ~30 GB of Wan-based weights the Production workflow loads), not this node.

Gotchas

The one real trap is wiring the wrong schedule in. If you grab sigmas from a different scheduler - say, reusing one from the Fast workflow - sample_flow_unipc will throw the "requires the matching upstream sigma schedule" error rather than produce a garbage render. That's the node working as designed, so just take both outputs from this node together. Also expect the Production path to be slow: this is the 40-step topology on a 14B Wan-based model, so budget minutes per clip, not seconds.