Your First Image

This tutorial walks you through generating your first diffusion image. By the end, you’ll have a generated image viewable in the admin UI.

Note

Make sure you’ve completed Installation & Setup and have all services running via ./start.sh.

Overview

The image generation workflow has three steps:

  1. Create a Prompt — describe what you want to generate

  2. Create a DiffusionJob — select a model and link it to your prompt

  3. View the result — the job auto-queues to Celery and generates the image

Step 1: Create a Prompt

Navigate to Diffusion > Prompts > Add Prompt in the app.

Fill in the following:

Source Prompt (required)

The text description of what you want to generate. For example:

A serene mountain landscape at golden hour, dramatic lighting,
snow-capped peaks reflected in a still alpine lake

That’s the only required field. You can optionally configure prompt enhancement:

Enhancement Method (optional)

Choose how the prompt gets enhanced before generation:

  • No Enhancement (default) — uses your prompt as-is

  • Rule-based — adds quality tags and style descriptors automatically

  • HuggingFace Local Model — enhances via Qwen2.5-3B running locally

  • LLM API — highest quality, requires ANTHROPIC_API_KEY in .env

Enhancement Style (optional)

If using enhancement: Auto-detect, Photography, Artistic, Realistic, Cinematic, or Coloring Book.

Creativity (optional)

Controls enhancement intensity from 0.0 (conservative) to 1.0 (creative). Default: 0.7.

Click Save. If you chose an enhancement method, the enhanced prompt will be generated when the job runs.

Step 2: Create a DiffusionJob

Navigate to Diffusion > Diffusion Jobs > Add Diffusion Job.

Fill in the Configuration tab:

Diffusion Model (required)

Select the model to use. Good choices for a first run:

  • SDXL Turbo — fastest (4 steps), good for quick tests

  • Flux.1-dev — high quality (28 steps), slower but excellent results

  • Juggernaut XL v9 — photorealistic (30 steps), supports negative prompts

Prompt (required)

Select the prompt you just created from the dropdown.

Identifier (optional)

A friendly name like my-first-image. Used in the output filename.

LoRA Model (optional)

Apply a style modifier. The dropdown is filtered to show only LoRAs compatible with your selected model’s architecture.

Leave the Parameters tab at defaults for your first run — each model has sensible defaults configured in presets.json (steps, guidance scale, dimensions).

Click Save.

Step 3: Watch It Generate

When you save, the job is automatically queued to Celery. You don’t need to do anything else.

You can monitor progress in two ways:

  • Flower at http://localhost:5555 — shows the task in the active/completed list

  • Job detail page — refresh the page to see status updates (pending → queued → processing → completed)

The first run for a given model will be slower because it downloads the model weights from HuggingFace. Subsequent runs use the cached model.

Step 4: View the Result

Once the job status shows completed, the detail page displays:

Results tab

  • Thumbnail previews of all generated images (clickable to view full size)

  • Images are saved to media/diffusion/

Generation Metadata

A JSON panel showing the exact parameters used: resolved prompt, model, steps, guidance scale, seed, LoRA details, and timing.

Experimenting Further

Now that you have the basics, try:

  • Different models — compare SDXL Turbo (speed) vs. Flux.1-dev (quality) vs. Juggernaut XL (photorealism)

  • Prompt enhancement — enable rule-based or LLM enhancement and compare the enhanced prompt to your original

  • LoRAs — apply a style LoRA (anime, photorealistic, fantasy) to the same prompt and see how it changes

  • Multiple images — set num_images to 3 or 4 on the Parameters tab to generate variations with different seeds

  • Negative prompts — for models that support them (Juggernaut, Realistic Vision, Qwen), add negative prompts to exclude unwanted elements

  • Compel weighting — on SDXL/SD15 models, use (dramatic lighting:1.5) syntax to emphasize concepts

Next Steps