Qwen Image 2.1: Text to Image
Generate images from text prompts with Qwen Image 2.1, the latest open-weight Qwen-Image model, producing a single 1024×1024 image over 25 sampling steps. Generation and instruction-based editing share one model with native 2K output, professional typography, and alpha-channel support. Ideal for poster and layout design, transparent asset creation, and rapid concept visualization.
Try an example
About this workflow
This workflow generates high-quality images from text prompts using Qwen Image 2.1, an open-weight diffusion model known for strong layout and typography rendering and native 2K (2048×2048) output. At its core is a custom Qwen generator node (UUID: c291ceec-b98f-4751-9d0b-7bc288f27b30) that loads either qwen_image_2.1_bf16.safetensors or the memory-saving qwen_image_2.1_int8_convrot.safetensors.
By default, it produces a single 1024×1024 image in about 25 sampling steps, with support for transparent (alpha) backgrounds when saved as PNG. ResolutionSelector sets the canvas size, and SaveImageAdvanced handles file output and metadata.
Technically, the subgraph encapsulates prompt conditioning and a built-in sampler inside the Qwen node, so you don’t need separate sampler nodes. You provide a prompt (and optionally a negative prompt if exposed), choose a resolution, and the node performs the full diffusion pass. The workflow’s MarkdownNote/Note nodes document on-canvas instructions, while SaveImageAdvanced ensures your output is written with correct color and alpha handling. This makes it ideal for fast concept visualizations, posters, and assets that benefit from clean text and layout—without leaving the ComfyUI canvas.
Questions & answers
Use qwen_image_2.1_bf16.safetensors for the best image fidelity if your GPU has ample VRAM. Use qwen_image_2.1_int8_convrot.safetensors to reduce memory requirements at a small quality trade-off. If you encounter out-of-memory errors—especially at 2K—switch to int8 or lower the resolution.
Be explicit in your prompt about hierarchy and placement (e.g., “bold sans-serif headline, centered title, small subheading, clean margins”). Use 1024–2048 px resolution, try 30–40 steps for more refinement, and keep the seed fixed while iterating so you can make controlled adjustments.
Qwen Image 2.1 supports instruction-based editing with the same underlying model, but this specific workflow is set up for text-to-image generation. To do edits, add or switch to a compatible Qwen editing node/subgraph that accepts an input image and edit instructions.
Ensure you save as PNG in SaveImageAdvanced. If the model outputs an alpha channel (transparent background), PNG will keep it; JPG will discard transparency. You can verify by opening the saved image in an editor that shows the checkerboard background.