ZML_高级K采样器

ZML_KSampler_Advanced (ZML_高级K采样器) is the "give me the full controls" version of the pack's sampler family. Where the plain ZML_KSampler covers the everyday seed/steps/CFG path, this one adds the levers people reach for when they're doing partial sampling - inpainting, two-pass refinement, denoise tricks - plus the same built-in live preview and A1111-style metadata that make the family worth using at all.

It's from ComfyUI-ZML-Image, zml-w's 160+ node pack, and it's the sibling of ZML_KSampler in the "sampler related" subcategory. Same ancestry, same borrowings (efficiency nodes for seed/preview, smZNodes for GPU noise, VideoHelperSuite for video preview).

How it works

It wraps the core sampling path with the four controls that define "advanced":

• 开始步数 / 结束步数 (start/end steps) - sample only part of the denoising schedule. Start at 0, end at 10000 (the default) and it runs the full range; set start > 0 to skip the early steps, or set a lower end to stop early. This is the lever for inpainting (where you only want the later, finer steps) and for "sample to 60%, then switch samplers" tricks.
• 添加噪波 (add noise) - enable/disable noise injection. Set to disable when the latent already carries the noise you want (e.g., a noised img2img latent from a prior step) so the sampler doesn't double-noise it.
• 返回剩余噪波 (return leftover noise) - when disabled, the sampler forces full denoise and returns a finished latent; when enabled, it stops mid-schedule and hands back a partially-noised latent for the next stage to continue. That's the "chain two samplers" mechanism.
• 随机种子 - same as the family's seed, -1 for random.

Then the family extras: 预览方式 (auto/latent2rgb/taesd/none) for live sampling preview, 噪波模式 (cpu/gpu) for GPU-side noise, 视频预览 + frame rate for watching frames accumulate, and the 脚本 input for the pack's ZML_SCRIPT nodes (dynamic CFG, extra noise, brightness/contrast, sharpen/blur applied mid-sampling).

The outputs that matter

• LATENT - the sampled latent (or the partially-noised one if you returned leftover noise).
• 图像 - a real decoded image only if a VAE is connected; otherwise a 1×1 placeholder. Connect a VAE to skip the separate VAEDecode.
• 生成信息 - the A1111-style parameter string the pack's save node reads for metadata. Same story as the plain sampler: your settings travel with the PNG.

Installing it

Pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/zml-w/ComfyUI-ZML-Image

restart ComfyUI, or ComfyUI Manager → search "ComfyUI-ZML-Image". No model downloads beyond the optional TAESD decoder for the taesd preview path.

Where people get caught

The start/end step defaults (0 and 10000) are designed to behave like the stock Advanced sampler's "full range," but the exact end cap matters: if you set 结束步数 below your actual step count, the sampler stops there and the output is deliberately unfinished. And the 图像 output's VAE dependency catches everyone once - no VAE, no image, no error. Finally, GPU noise mode changes the noise source, so a locked seed behaves differently between cpu and gpu - same as the plain sampler, worth remembering when you're trying to reproduce a result.

When you'd actually reach for it

When your workflow needs partial schedules - inpainting passes, sampler hand-offs mid-denoise, controlled img2img. If all you do is straight txt2img, the plain ZML_KSampler (or the stock one) is simpler and identical in results. This one earns its "advanced" name exactly when you're sampling in pieces.