🛡️ Binyuan采样器 V5.5

This node ("🛡️ Binyuan采样器 V5.5") wants to eat your whole graph. Where most workflows are a checkpoint loader, a dual CLIP loader, a VAE loader, two text-encoders, an empty-latent node, a KSampler and a decoder wired together across half the canvas, this is one box that does all of it: load the model, encode both prompts, make the latent, sample, decode, and hand you the finished image. It's the same "pipe" family as Efficiency Nodes' Efficient Loader - one fat node instead of a noodle plate - just taken further.

Here's the thing you should know up front: this is a genuinely obscure pack. Zero search impressions on comfy.icu, basically no English Reddit discussion, and the entire interface is in Chinese. It's the personal project of one GitHub user (yjitiu), not a Comfy Org staple. That doesn't make it bad - the code is thoughtful, and the README is well-written in both languages - but you're a bit of a pioneer if you use it, and the next section of this page is basically your map.

How it works

Under the hood it's a single BinyuanUltimateSamplerV9 class that leans entirely on ComfyUI's stable public API: load_checkpoint_guess_config, DualCLIPLoader, common_ksampler. No exotic dependencies (the pyproject declares none), which is refreshing for a node this ambitious.

The two modes in 加载模式 (load mode) matter. 整包Checkpoint loads a single .safetensors checkpoint and reads the VAE from it. 分离式 (split) lets you pick the diffusion model, CLIP_1/CLIP_2 and VAE separately - which is how you run Flux, SD3, Wan, Qwen-Image, Krea2, Z-Image, LTXV and friends. The CLIP_类型 dropdown must match your model's real architecture, or the load/encode blows up; the code even canonicalizes quirky names (pid/pixl → pixeldit) so you don't have to.

The cleverest bits are invisible. It reads the loaded model's latent format to build a correct empty latent (right channel count, right downscale ratio, even 3D latents for video models like Wan). It validates that your VAE's latent channels match the model before sampling, which saves you from the classic "sampled fine, decode failed" disaster. And if you feed it a video VAE it flattens the 5D output to standard 4D IMAGE so SaveImage doesn't choke.

The inputs that actually matter

• 串联模式 (cascade mode) - set to 继承上游模型 (inherit) and wire your own loaded stack into 外部模型 / 外部CLIP / 外部VAE, and this node becomes a pure sampler again. That's the safest way to use it: you keep your normal loaders, it just does the encode→sample→decode.
• 权重精度 - fp8_e4m3fn / fp8_e5m2 / bf16 / fp16 / nvfp4 for the diffusion model, on top of the load. nvfp4 is Blackwell-only; on a 30-series card it won't help you. Note that picking a .gguf file sidesteps this entirely - GGUF is self-quantized.
• Flux引导 (guidance) - Flux-family models don't do classic CFG, so this injects a guidance scale into the conditioning instead. 3.5 is the default and most people quietly run less; the KB notes the community argues 3.5 is too high for a lot of prompts.
• 上游图像_1/2/3 + 重绘强度 - feed images in for img2img; there's even a 上游图像拼接 mode that stitches multiple images into one canvas.

Outputs: 图像 goes to SaveImage. The rest - 模型, CLIP, VAE, Latent, 正面条件, 负面条件 - are all exposed so you can wire this monster into downstream nodes (ControlNet, hires, whatever).

Install

Easiest via ComfyUI Manager (search "Binyuan Sampler"), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/yjitiu/ComfyUI-Binyuan-Sampler.git binyuan_sampler_plugin_v5.5

Then restart ComfyUI. There are no pip dependencies. One real dependency: if you pick a .gguf model, the code dynamically loads ComfyUI-GGUF (city96's pack) from custom_nodes/ComfyUI-GGUF - install that too or GGUF files will raise an error.

Where people get burned

The CLIP_类型 mismatch is the number one failure - always match it to the model family, and keep the VAE from the same family (Wan/Qwen need the wan2.1 VAE). The node tries to help: on any error it prints [错误] to the console and returns a blank 64×64 image instead of crashing, so a black output you didn't generate means go read the console. And remember everything is labeled in Chinese - if you're swapping between this and English packs, you'll be translating widget names in your head for a while. Fair warning, not a bug.