🐋 千问图像集成采样器——Github:@luguoli

If you've ever built a Qwen-Image or Z-Image workflow the "official" way, you know the feeling: a prompt encoder, an AuraFlow shift node, a CFG normalization node, a latent creator, a KSampler, a VAE decode, plus scaling and cleanup nodes, all draped in a dozen wires. This node is that entire pipeline collapsed into one box. Load your model, clip and VAE, type a prompt, and it does prompt encoding, sampling, decoding, optional auto-save, even VRAM cleanup - and it's exactly how people actually run Qwen and Z-Image these days. The community recommendation for a Z-Image depth workflow was literally "use this, two or three nodes is enough."

It's the flagship of luguoli's pack (the whale 🐋, "千问图像集成采样器"), and it does double duty: text-to-image and image-to-image in a single node, including Qwen-Image-Edit-style instruction editing with up to five reference images.

How it works

Pick generation_mode. In text-to-image it just tokenizes your prompt and makes an empty latent at the width/height you set. In image-to-image the interesting stuff happens: each reference image is encoded twice at once - once into the vision-language encoder as 384×384 image tokens, and once into the VAE at the image's strict original resolution (rounded to a multiple of 8). That dual-encode is the whole Qwen-Edit design. And the "strict original resolution" part is the point of this node: the official TextEncodeQwenImageEditPlus used to upscale your reference, which is where the infamous offset/shrink artifacts came from. This keeps the latent true to the input, and the pack's own code comment says exactly that - it exists to solve the offset problem while keeping the official node's instruction-following.

Before sampling it can also apply two tweaks. auraflow_shift (default 3.0) patches a flow-sampling shift onto the model - the AuraFlow technique for balancing schedule shape, speed and quality. cfg_norm_strength (default 1.0) installs a post-CFG normalization that stops high guidance from oversaturating. Both only engage when set above zero, so 0 turns them off. If you've wired in controlnet_data from the companion loader, each condition gets applied here, with repaint masks scaled to match your target size.

Inputs that matter

• generation_mode - text-to-image or image-to-image.
• width / height - required for text-to-image (it throws otherwise). For image-to-image, setting them letterboxes your references to that size while keeping aspect ratio; 0 means no scaling.
• image1–image5 - the references, up to five, image1 being the main one. No image1 in image-to-image mode and it refuses to run.
• denoise - lower it in image-to-image to keep more of the original.
• steps (default 4), cfg (default 1.0), sampler_name, scheduler - the usual suspects. Qwen is LLM-encoded and prompt it in sentences, not tags; it's happy in the 4–20 step range, and pushing CFG much past ~7 burns the image.
• batch_size - 1–10. Start at 1 and watch your VRAM.
• auto_save_output_folder - set a path (absolute, or relative to ComfyUI's output dir) and every result saves as prefix_seed_00000.png. Leave blank to disable.

Three outputs: the generated Image, the Latent (handy if you want to chain a refiner without a decode/encode round trip), and the Scaled Image - a preview of what the sampler actually saw, which saves you an "oh no, it upscaled my reference" moment.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/luguoli/ComfyUI-Qwen-Image-Integrated-KSampler.git
pip install -r requirements.txt

Then restart ComfyUI. Or ComfyUI Manager → search "ComfyUI-Qwen-Image-Integrated-KSampler" → Install. The requirements are all standard (torch, transformers, einops, psutil) and the repo ships no weights. The real dependency is the model: Qwen-Image is a 20B MMDiT and a VRAM hog, so quantization is basically deployment, not an option - GGUF, fp8, or Nunchaku SVDQuant on consumer cards. This node also handles Z-Image fine, and that one runs on a 6GB card.

Gotchas

Text-to-image without width/height errors out on purpose, and image-to-image without image1 does too - both messages are bilingual and clear. The sound notification is Windows-only (winsound), so don't expect a chime on Linux. The "clean CPU memory after finish" option is aggressive: it flushes file cache, processes and dynamic libraries, so enable it only when a long session is leaking. And if you want the Chinese UI, the post-ComfyUI-0.3.68 localization break means you'll need to run the pack's 自动汉化节点.bat and install ComfyUI-DD-Translation. That's a lot of Chinese characters for one node - but then, the node itself just saved you about seven wires.