ZImage Turbo Sampler Advanced

The plain ZImageTurboSampler picks your sampler and scheduler for you. This one lets you disagree with it. Same inputs, same five mode presets, same outputs - plus a sampling_profile choice and three overrides that let you edit the plan before it runs.

The new inputs:

• sampling_profile - three curated sampler/scheduler pairs:
  • capitan_flow (default) - the pack's euler_flow sampler + capitanZiT linear scheduler. The community-tested favorite.
  • zflow_linear - euler_flow + the zimage_turbo scheduler alias (also a linear ramp, registered by this pack).
  • tongyi_default - res_multistep sampler + simple scheduler, the vanilla pair that tracks the official Tongyi recipe before community tuning.
• steps_override / cfg_override / denoise_override - set to 0 to mean "use the preset." Any nonzero value replaces that part of the plan. So steps_override=12 on base_ultra gives you a 12-step ultra pass; cfg_override=3 turns guidance back on for the people who like a little contrast (it works above 1, but costs steps and can overcook past ~3); denoise_override=0.5 is how you turn a base mode into a half-refresh img2img pass.

The mechanism underneath is unchanged: the node resolves the plan, applies your overrides, and calls common_ksampler with CFG 1 as the baseline. The output echo (steps, cfg, denoise, sampler_name, scheduler_name) is your receipt for what actually ran, which is worth checking once you start overriding.

Where this earns its keep: the refine modes. base_ultra with a denoise_override is how you get a "touch-up this image" pass without dragging out the two-pass refiner; steps_override is how you find your model's personal step sweet spot between 8 and 15 (more than ~9 on the stock pair tends to hurt rather than help, but with euler_flow the relationship is gentler). CFG above 1 is the "more saturated and contrasty" trick from the community tinkering threads - nice, but double the render time and fragile past 3.

Everything else - install, models, the pack's locally-registered sampler names - is shared with the base sampler. ComfyUI Manager → "ZImageTurboHQNodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/marduk191/ComfyUI-ZImageTurboHQNodes

restart, no pip deps, models from Tongyi-MAI's HF repo.

The trap: overrides are easy to leave behind. A cfg_override of 4 from last week's experiment will happily sit at 4 and silently change every render. Zero means "use preset," and if you're not actively experimenting, zeros are what you want. Treat this node like a fine instrument you return to factory between sessions.