🛡️ ComfyUI Binyuan · Ultimate Sampler V6.5 EN (Bilingual)

Bilingual build of the all-in-one Binyuan sampler: English UI for English users, Chinese UI for Chinese users. Source & issues: https://github.com/yjitiu/ComfyUI-Binyuan-Sampler-EN · the Chinese-only edition stays separate (node key BinyuanUltimateSampler), so both can be installed side by side.

Node: 🛡️ Binyuan Sampler V6.5 EN｜中英双语版 — category Binyuan
Node key: BinyuanUltimateSamplerEN
Widget labels: English ASCII keys, translated by locales/en|zh/nodeDefs.json (official ComfyUI i18n)
Combo values are bilingual, e.g. Inherit upstream ｜继承上游模型
All error messages are bilingual, e.g. Please select a Checkpoint / 请选择 Checkpoint

Install

Option A — drop-in zip (recommended)

Download binyuan_sampler_plugin_english_v6.5.zip from this repo (or from the GitHub release).
Unzip it into ComfyUI/custom_nodes/ (the folder name does not matter).
Restart ComfyUI → right-click → Binyuan → 🛡️ Binyuan Sampler V6.5 EN｜中英双语版.

Option B — git clone straight into custom_nodes

cd ComfyUI/custom_nodes
git clone https://huggingface.co/yjitiu/ComfyUI-Binyuan-Sampler-EN binyuan_sampler_plugin_english

What it does

One node does load model → encode prompt → sample → decode:

Load Mode = Checkpoint (whole file) — pick a .safetensors checkpoint.
Load Mode = Separate — pick Diffusion Model + CLIP 1/2 + VAE and set CLIP Type to match the architecture (Flux / SD3 / Wan / Qwen-Image / Krea2 / Z-Image / LTXV / Hunyuan / Lumina2 / …).
Weight Precision: default / fp8 / nvfp4 / bf16 / fp16 …
LoRA: use the built-in LoRA manager panel (+ Add LoRA), or write JSON into LoRA JSON.
img2img: wire an image into Upstream Image 1 and lower Denoise.
Outputs IMAGE, MODEL, CLIP, VAE, LATENT, positive/negative CONDITIONING and the denoised latent, so it can be chained with other nodes.

Chaining with the Chinese edition

Works both ways: connect MODEL / CLIP / VAE (outputs 2 / 3 / 4) to External Model / External CLIP / External VAE on the other node and set Chaining Mode = Inherit upstream. The inheriting node reuses the already loaded model (no second load, saves VRAM).

Verified

Tested end-to-end on a real Krea2 stack (krea2_turbo_fp8 + qwen3vl_4b_fp8_scaled, CLIP Type = krea2, krea2RealVae_v10, RTX 5070 Ti, 768×1024, 8 steps, CFG 1.0):

txt2img — 20.5 s
img2img (upstream image, denoise 0.6)
LoRA via LoRA JSON
chaining English → Chinese node and Chinese → English node

Troubleshooting

Symptom	Cause / fix
`No operator found for memory_efficient_attention_forward` … `capability (12, 0)`	xformers has no kernels for your GPU (RTX 50 series / sm_120). The stock `KSampler` fails the same way — it is an environment issue. Launch ComfyUI with `--use-pytorch-cross-attention --disable-xformers`. RTX 40 series and older are unaffected.
`ComfyUI-GGUF custom node not found`	You picked a `.gguf` model. Install `ComfyUI-GGUF` into `custom_nodes/`.
`CLIP Type` seems ignored / weird encode errors	Your ComfyUI core is too old. `krea2`/`boogu` need v0.26.2+; on v0.24.1 / v0.22.2 / v0.20.3 some types are missing and silently fall back to `stable_diffusion`. Update ComfyUI.

中文说明

这是 中英双语版：英文界面用户看到英文标签，中文界面用户看到中文标签（通过 ComfyUI 官方 locales/ 翻译机制）。节点键名与中文版不同（BinyuanUltimateSamplerEN vs BinyuanUltimateSampler），两个版本可同时安装、互不覆盖。下拉选项值与报错信息均为中英双语，即使翻译文件未生效也能看懂。

安装：把 binyuan_sampler_plugin_english_v6.5.zip 解压到 ComfyUI/custom_nodes/，重启，右键 → Binyuan → 🛡️ Binyuan Sampler V6.5 EN｜中英双语版。

注意：.gguf 需要另装 ComfyUI-GGUF；内核版本过老时部分 CLIP Type 不可用（见上表）。

Downloads last month: -; Downloads are not tracked for this model. How to track