Build a Text‑to‑Image Workflow with Dify in 5 Minutes

In this tutorial the author shows how to use Dify's workflow feature to build a text‑to‑image pipeline, comparing it to traditional workflow engines like Activiti and Flowable and emphasizing low‑code visual composition.

Workflow basics : Dify workflows connect nodes via a drag‑and‑drop canvas, support global variables, API calls (e.g., weather data), error handling, scheduled triggers, and event‑driven actions such as WeChat message pushes. They can integrate 300+ large models including GPT‑4, Claude, and Chinese models.

Key concepts :

Nodes are the building blocks that execute specific functions.

Variables (system, environment, session) link inputs and outputs across nodes.

Dify offers two workflow types: chaflow for memory‑aware multi‑turn conversations, and workflow for automation and batch processing such as translation, data analysis, content generation, and email automation.

Step‑by‑step construction :

Create a new workflow in Dify.

Understand the Dify DSL, a YML‑based file format describing the application, model parameters, and orchestration.

Deploy ComfyUI as the image generation engine. Download the repository from github.com/comfyanonymous/ComfyUI, extract it to a path without Chinese characters, and configure the Windows firewall to allow port 8188 (Advanced Settings → Inbound Rules → New Rule → Port → 8188 → Allow).

Download the Stable Diffusion checkpoint (e.g., stable-diffusion-v1-5-archive) from Hugging Face and place it in ComfyUI\models\checkpoints.

Start ComfyUI by running run_cpu.bat, which opens http://127.0.0.1:8188/ in a browser.

Authorize the ComfyUI service in Dify: search for ComfyUI in the Dify tool panel and set the server URL to http://host.docker.internal:8188 (Docker for Windows/Mac resolves this to the host IP).

Workflow node configuration :

Start node: define a variable that holds the textual description for image generation.

LLM model node: use a prompt that converts the user’s text into an English image prompt without returning the reasoning process. Example prompt: "Based on the user input, generate an English text‑to‑image prompt. Respond only with the prompt."

Optional code execution node: perform string manipulation on the LLM output (e.g., using the cursor tool for non‑programmers).

Txt2Img node: select the model v1-5-pruned-emaonly-fp16.safetensors, set ModelType to SD3, STEP to 20, Scheduler to normal, Width/Height to 512, and add a negative prompt such as "bad art, ugly, deformed, watermark, duplicated, discontinuous lines".

End node: output the generated image.

Tips : Reduce STEPS or resolution if generation is slow; some prompts may be filtered, so describe them indirectly.

The author demonstrates the final result with screenshots of the generated images.

Overall, the tutorial illustrates how Dify’s modular workflow can turn a creative process into reusable AI‑powered components, enabling anyone to build their own “AI digital employee.”