🛡️ Binyuan Sampler V2.3 EN｜中英双语版

If you've stared at a 30-node ComfyUI graph and thought "this is four wires and a tin of noodles," the Binyuan sampler is aimed at you. It's one node - class BinyuanUltimateSamplerEN, display name 🛡️ Binyuan Sampler V2.2 EN｜中英双语版, category Binyuan - that loads a checkpoint or a split stack, encodes both prompts, builds the latent, samples, decodes, and hands you a finished IMAGE. It also exposes model, CLIP, VAE, conditioning and latents as outputs so you can chain it into a second stage.

This is the bilingual build; the Chinese edition uses a different node key, so both can live side by side. It's MIT and declares zero dependencies, which is rare for a pack with this much surface area.

Why the all-in-one shape exists

This is the context-bus pattern: instead of dragging model, clip, vae, conditioning, latent, size, steps and seed across the graph as a dozen parallel wires, you bundle them. Efficiency Nodes popularized it, ComfyUI-Easy-Use threads a pipe dict - same idea, bigger surface. The payoff is legibility. The failure mode is stale contents: when one node holds everything, it's easy to run base weights while believing a LoRA is applied. Binyuan's loader lives inside the sampler, so nothing gets unpacked out of order - but the LoRA panel, chaining mode and latent source decide what's inside the bundle, and you have to read them.

How it works

Two load modes. Checkpoint (whole file) picks a baked checkpoint. Separate (Flux/SD3/Diffusion) takes a diffusion model, clip_1, optional clip_2, a vae (default baked_vae) and a clip_type from a 30-entry list - flux, flux2, krea2, hidream, hunyuan_image, chroma and friends. .gguf files show up in those dropdowns because the node dynamically imports City96's ComfyUI-GGUF rather than vendoring it.

Sampling runs on ComfyUI's public API: 44 samplers (euler default), 9 schedulers (simple default), and a weight_precision cast (fp8_e4m3fn, nvfp4, bf16…). Prompt boxes support {a|b|c} dynamic prompts drawn from the seed.

One genuinely unusual knob: attention_backend, defaulting to Auto safe. For a Krea2 or Qwen-Image-2.1 model loaded by this node it can pick Comfy Kitchen when available, otherwise ComfyUI's sub-quadratic backend - a model-local patch, not a startup flag, so models wired in from elsewhere keep their own setup.

The controls you'll actually touch

• load_mode / checkpoint / diffusion_model / clip_type / vae - your model. clip_type must match the architecture; a mismatch is the classic "why is my conditioning garbage" cause.
• positive_prompt / negative_prompt - plain text boxes, {a|b|c} supported.
• steps, cfg, sampler, scheduler, denoise, seed, width, height, batch_size - the usual. cfg defaults to 1 and flux_guidance to 3.5, right for distilled models: at exactly 1 ComfyUI skips the negative pass rather than weighting it to zero.
• chaining_mode - Own model, Inherit upstream, Auto detect. Inherit upstream reuses external_model / external_clip / external_vae, so a second sampler doesn't load the weights twice.
• latent_source - Empty latent, External latent first, Upstream image first, Stitch upstream images. The older upstream_image_handling widget is kept only for workflow compatibility; latent_source drives behavior.
• lora_json - the + Add LoRA panel writes here, but it also parses plain text (name.safetensors:0.8, one per line, # comments) or a JSON array with per-model and per-CLIP strengths.

Outputs, in order: IMAGE, MODEL, CLIP, VAE, LATENT, POSITIVE, NEGATIVE, DENOISED. IMAGE into SaveImage is the whole beginner workflow; feed MODEL/CLIP/VAE/LATENT into a second Binyuan node to skip a reload. Two newer clusters: advanced_sampling turns steps into a total schedule and makes start_at_step / end_at_step / add_noise / return_with_leftover_noise authoritative (denoise is ignored there), and noise_strength (0–2) scales initial noise in both modes. For Qwen-Image-2.1, qwen_edit_mode plus qwen_reference_resolution: native reference editing starts from an empty latent so the node pins denoise to 1.0, while VAE redraw keeps the source canvas and takes 0.25–0.55 happily.

Install

ComfyUI Manager → search "Binyuan Sampler". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/yjitiu/ComfyUI-Binyuan-Sampler-EN
# restart ComfyUI, then hard-refresh the browser (Ctrl+F5)

The folder name doesn't matter; the folder's contents do. Keep __init__.py, js/, locales/, LICENSE and pyproject.toml - copying just the Python file loses the LoRA panel and the labels. After updating, restart the backend and refresh, then re-add the node if the new widgets still aren't there.

No models ship with it. .gguf support needs ComfyUI-GGUF in custom_nodes/ComfyUI-GGUF. Qwen-Image-2.1 needs a core with TextEncodeQwenImage21, the matching text encoder, and the 64-channel Qwen VAE.

Where people get burned

The attention error that isn't this node's fault. If sampling dies with No operator found for memory_efficient_attention_forward ... requires device with capability <= (9, 0) but your GPU has capability (12, 0), you have xformers installed and its prebuilt kernels don't cover your GPU. Every sampling node fails the same way, stock KSampler included. Fix it at launch:

python main.py --use-pytorch-cross-attention --disable-xformers

Cross-family latents. Flux/Krea and Qwen latents differ in channel count and scaling. Don't pass LATENT between them just because the sockets are the same color - pass IMAGE and let the second model re-encode it.

Knobs that look dead. Low denoise on an empty latent gets protected back to 1; denoise does nothing in advanced mode; noise_strength does nothing there when add_noise is off; a connected external_sigmas overrides steps, scheduler and denoise; end_at_step <= start_at_step skips sampling entirely. And if the step count arrives on a wire from a shared integer, editing the node's own box changes nothing.

Finally: don't install two copies of the same bilingual node ID, and back up custom_nodes before updating. This is a young single-author pack with no cross-version compatibility guarantee.