⚓ Capitan Presets

If you've generated with Flux, Z-Image-Turbo, or SD3 and gotten images that look burned, washed out, or just wrong no matter what you do to the sampler dropdown, you've hit the exact thing this node exists for. Flux, SD3, Lumina2, and Z-Image are all rectified flow models. They are not SDXL under the hood, and the sampling habits you carried over from SDXL - DPM++ 2M Karras, 30 steps, CFG 7 - actively hurt them.

Why flow models make you guess

Flow matching trains a model to follow a near-straight line from noise to image, instead of walking back down a curved noise schedule the way SD 1.5/SDXL do. Two things follow. First, you don't need many steps: 8 is normal for a distilled model like Z-Image-Turbo. Second, aggressive schedulers like Karras, which concentrate denoising effort in the middle of the trajectory, distort a straight path rather than correcting it. The knob that replaced the scheduler for flow models is shift - how much of your step budget gets spent at high noise (composition) versus low noise (detail). And shift is model-specific: the value that makes Z-Image-Turbo sing will overcook something else.

That's the problem Capitan Presets solves. It's a table of "right steps, right shift, right CFG" per model and quality tier, wrapped in a single dropdown. One reviewer on the launch thread called the presets "ingenious for beginners," and that's the accurate review: the whole point is that you don't need to know what shift is to get it right.

The one input, three outputs

The entire node is one preset dropdown with ten options: Fast/Balanced/Quality tiers for Lumina2 and Flux, Speed/Quality for Z-Image-Turbo, and Balanced/Quality for SD3. Pick one and you get three outputs:

• sigmas - the actual noise schedule, computed from the preset's steps and shift.
• recommended_steps - an INT telling you how many steps the preset assumes.
• recommended_cfg - a FLOAT with the CFG that tier was tuned at.

That CFG output is the sneaky-useful one. Guidance-distilled flow models run at CFG 1.0 (Z-Image-Turbo) or low values, and cranking CFG 7 doesn't buy you prompt adherence - it applies the guidance correction twice and burns the image at double the render time. The presets bake the right CFG in so you don't have to remember which models are distilled.

Wiring it up

sigmas goes into a SamplerCustom node's sigmas port, and you select capitan_flowmatch via a KSamplerSelect into SamplerCustom's sampler input.

CapitanShiftPresets --sigmas--> SamplerCustom.sigmas
KSamplerSelect (capitan_flowmatch) --> SamplerCustom.sampler

Install

One install gets you this node, its two siblings, and the three samplers. In ComfyUI Manager, search for ComfyUI-CapitanFlowMatch and install, then restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-CapitanFlowMatch

Restart ComfyUI. No model downloads, no pip requirements - it's a single pure-Python file that only imports torch and ComfyUI internals.

The honest take

Here's the part the README won't lead with: the three samplers this pack registers are Euler in a trench coat. They're all the same first-order update on the flow ODE - the "advanced" variant is algebraically identical to the flagship, and "turbo" just adds a guard for the near-zero sigma step. A reddit reviewer called this out within hours of launch, and they weren't wrong.

That's less damning than it sounds. Euler on a conservative, slightly-shifted schedule is exactly what flow models want, and the real value of this pack isn't a secret sampler - it's the schedule knowledge. The dropdown encodes steps + shift + CFG per model, which is the part people get wrong constantly. The pack's own FAQ admits you can reproduce it with built-in ModelSamplingAuraFlow + BasicScheduler; you just need to know the shift values, which is precisely the knowledge this dropdown hands you.

Common issues

• Sigmas into a KSampler? Doesn't work - KSampler has no sigmas port and builds its own schedule. Use SamplerCustom, or use KSampler and pick capitan_flowmatch from its dropdown (the schedule will be KSampler's, not the preset's).
• Z-Image-Turbo at 36 steps? Yes, seriously. The Quality tier runs Turbo at 36 steps, shift 6.0, CFG 2.0. Turbo is optimized for few steps, not limited to them - that's the pack's hidden quality mode, and arguably the reason it exists.
• Wrong model preset. The dropdown is keyed to model names, not to what you loaded. Pick "Flux (Quality)" with an SD3 checkpoint loaded and you get Flux's numbers, not SD3's.
• Burning at high steps. That's the failure mode the whole pack is built to avoid - stock euler with a reshaped schedule accumulates errors on flow models; the shifted linear schedule here doesn't.