ZImage Turbo Two Pass Refiner Advanced

The plain ZImageTurboTwoPassRefiner gives you three strength presets and calls it done. This one is the same upscale-and-refine machine with the sampling plan left open: you keep strength, upscale_by, and the seed, and you add a sampling_profile plus two overrides for the steps and denoise of the refine pass itself.

The mechanism is identical to the base refiner - Lanczos upscale by upscale_by, VAE encode, a CFG 1 partial-denoise pass with the pack's sampler, VAE decode, IMAGE out. What's new:

• sampling_profile - capitan_flow (default; euler_flow + capitanZiT linear), zflow_linear (euler_flow + zimage_turbo), or tongyi_default (res_multistep + simple). If you've already got a favorite pair from the sampler nodes, this is where you make the refiner agree with it.
• steps_override - 0 means "use the preset" (the mode gives you 8–10 steps); any nonzero value replaces it. Handy when a refine pass feels like it's not converging and you want a couple more steps.
• denoise_override - 0 means "use the strength preset" (0.25 / 0.32 / 0.40); otherwise it's your number. This is the knob people actually reach for: dropping to ~0.2 makes the refiner a gentle texture polish, raising to ~0.6 turns it into a real "reinterpret this image" pass.

Note there's deliberately no cfg_override. The refiner always runs at CFG 1.0, which is the right call for this guidance-distilled model - guidance above 1 during a refine pass tends to burn contrast and fight the denoise. If you want CFG drama, do it on the base sampler, not here.

The inputs you'll still touch are the ones from the base version: model, positive/negative (negative is the zeroed formality at CFG 1), vae, image (the VAEDecode output of your base pass), seed (pair with ZImageTurboSeedControl's refine_seed), upscale_by (1.0–4.0, 1.5 default), and strength. Output is a single image.

Install is pack-level: ComfyUI Manager → "ZImageTurboHQNodes", or git clone https://github.com/marduk191/ComfyUI-ZImageTurboHQNodes into ComfyUI/custom_nodes, restart. No pip dependencies; the models (z_image_turbo_bf16.safetensors, qwen_3_4b.safetensors, Z-Image VAE) come from Tongyi-MAI's HF repo, not the pack.

Honest take: most people never need this over the base refiner. The base version's presets are well-chosen, and strength + upscale_by covers 95% of use. Reach for the Advanced only when a refine pass is coming out too soft (denoise_override up) or too crunchy (down), or when you've standardized on a sampling profile and want every pass to match it. And like all the override fields in this pack, remember zero means "use the preset" - if you're not experimenting, keep them at zero.