MinusZone - StylizePhotoKSampler

You know the claymation filter - the one that turns a photo of your dog into a cute little stop-motion figurine? MZ_StylizePhotoKSamplerNode is that whole pipeline in a single node. Drop in an image, pick a checkpoint, hit run, and you get back a stylized photo. No checkpoint loader, no VAE encode, no LoRA loader, no prompt builder, no sampler - the node does all of it, then stamps "ComfyUI / Generated by AI" across the bottom unless you tell it not to.

This is a "results, not knobs" node from the small pack MinusZoneAI/ComfyUI-StylizePhoto-MZ (the author is MinusZone, a Bilibili/Xiaohongshu creator who publishes his Chinese tutorials off GitHub). The whole point is simplicity: it's img2img plus a style LoRA under the hood, which is the standard recipe for photo-restyle. Reach for it when you want the clay look without hand-building a workflow, or when you're porting someone's one-click stylize graph and just want the thing to run.

How it works

Mechanically it's a normal SDXL img2img pass. The node loads your checkpoint, resizes the input so its long edge matches resolution (rounded to a multiple of 8 so the VAE stays happy), VAE-encodes, samples at your denoise (0.75 default - high enough to restyle, low enough to keep your composition recognizable), and decodes. The clay effect comes from a LoRA, clay_v1.pt, that the node downloads itself on first run from ModelScope into ComfyUI/models/minus_zone_models/ - you never place a file by hand. It also auto-downloads a Chinese font for the watermark.

What's fixed and what isn't: the sampler is hardcoded to euler_ancestral with a normal scheduler, the LoRA strength is locked at 0.88, and each style_type carries its own baked-in prompt ("clay, clay_model, photography, macro, cute..."). Your positive_prompt gets appended to that; your negative_prompt to the style's negative. Coming from a full KSampler graph, this feels restrictive - but that's the deal. It's a stylize button, not a sampler playground.

The inputs that actually matter

• xl_ckpt_name - pick any SDXL checkpoint from your list, or set it to "none" and wire model/clip/vae in yourself. The author's examples use Juggernaut XL.
• image - the photo you're restyling.
• resolution - default 1024; the long edge is resized to this. -1 keeps the original size.
• seed / steps / cfg - 0, 35, 5.5 defaults, all sane.
• denoise - the one knob you'll actually fiddle with. Push toward 0.9 for a stronger restyle, drop to ~0.6 to keep the original closer.
• watermark - defaults to "ComfyUI\nGenerated by AI". Clear it if you don't want a stamp on every output. Most people hit this by surprise.
• control_net (optional) - the output of the pack's other node, MZ_StylizePhotoControlNetApply, if you want tile/canny guardrails on the composition.

Output is a single IMAGE, ready for a Save Image node.

Installing it

ComfyUI Manager → search "ComfyUI-StylizePhoto-MZ", or:

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-StylizePhoto-MZ

then restart ComfyUI. There's no requirements.txt, which is the one real trap: the code imports OpenCV (cv2) at load time, and ComfyUI doesn't ship it. If the nodes don't appear after install, add it to your ComfyUI Python environment:

pip install opencv-python

Where people get burned

• First run downloads from ModelScope, the Chinese Hugging Face mirror - the clay LoRA plus the watermark font. You need internet to that host, and it can be slow or flaky outside China.
• It's an SDXL-era pack. Last commit May 2024, repo now archived, and style_type has exactly one entry: clay.v1. Use an SDXL checkpoint, not Flux, and don't hold your breath for new styles.
• The watermark is on by default. It's a deliberate feature ("自带水印功能" is literally in the README), but if your output suddenly has a caption you didn't write, that's this node.

Fair warning: the models live on the author's personal ModelScope account (wailovet), so if they ever get pulled, the node breaks with a download error and there's no offline fallback. It's a hobbyist pack - tiny, archived, and fine with that. For the claymation look with zero assembly it does the job; just know exactly what you're plugging in.