ComfyUI Text-to-Image Tutorial: Generate Your First AI Image

This lesson uses ComfyUI's most basic text-to-image workflow. First, follow the previous lesson to download a model (such as an SDXL model) and place it in models/checkpoints.

Load the default workflow

The default workflow in newer ComfyUI versions changes to use newer models (for example, ComfyUI 0.39 defaults to Z-Image Turbo, with the model split across two nodes: Load Diffusion Model and Load CLIP). This lesson uses the classic, easy-to-understand SDXL text-to-image workflow. You can:

• Search for SDXL in Templates and choose the basic text-to-image template, or
• Add the nodes and connect them yourself as shown below

The workflow looks like this:

[Load Checkpoint] ─MODEL──────────────────────────────▶ [KSampler] ─LATENT─▶ [VAE Decode] ─IMAGE─▶ [Save Image]
       │ CLIP ─▶ [CLIP Text Encode (positive prompt)] ─────▶     ▲  ▲                 ▲
       │ CLIP ─▶ [CLIP Text Encode (negative prompt)] ─────▶ ────┘  │                 │
       │                       [Empty Latent Image] ──────────┘                 │
       └ VAE ──────────────────────────────────────────────────────────────────┘

Configure each node

1. Load Checkpoint: load the model

Click ckpt_name and select the model you downloaded. This node outputs three things:

• MODEL: The diffusion model; connect it to KSampler
• CLIP: The text encoder; connect it to both CLIP Text Encode nodes
• VAE: Connect it to VAE Decode

2. CLIP Text Encode: prompts

There are two:

• Positive prompt (connected to KSampler's positive): what you want to see
• Negative prompt (connected to negative): what you don't want to see

Most image models understand English prompts best:

Positive: a cozy coffee shop on a rainy evening in Taipei, warm lights, people with umbrellas outside the window, cinematic lighting, highly detailed
Negative: blurry, low quality, watermark, text, deformed

See Prompting tips for how to write prompts.

3. Empty Latent Image: canvas size

Set the image width and height, and the number of images to generate at once (batch_size).

Using dimensions far below or above the model's training resolution can cause duplicated people or distorted compositions. For a larger image, generate it at the recommended size first, then upscale it.

4. KSampler: generate

This is the node that actually generates the image. The main parameters are:

See KSampler parameters explained for details on each parameter.

5. VAE Decode: decode

Converts the latent output from KSampler into an actual image. Connect the vae input to the VAE output from Load Checkpoint.

6. Save Image: save

Displays the image and saves it to the ComfyUI/output folder. You can set the filename prefix with filename_prefix. It can also include a subfolder, such as coffee/rainy.

To preview without saving, use the Preview Image node instead.

Run

Click Run (Ctrl + Enter). The first run takes longer because the model needs to load. Later runs are much faster when you only change the prompt.

After the image appears:

• If you don't like the composition, run it again (the seed will change randomly).
• If you like an image and want to refine it, set control_after_generate to fixed to keep the seed, then change the prompt or parameters to compare the differences.

Generate in batches

• Set batch_size to 4 to generate four images at once (this requires more GPU memory).
• Or click Run multiple times in a row to queue one image per run.

Generated images include the workflow

PNG images in the output folder record the complete workflow and parameters. Drag one back onto the ComfyUI canvas to restore its settings.

Common issues

Next, take a closer look at each KSampler parameter.

Updated October 12, 2026