Launch your own Agent

Key Concepts

• Base Model (SDXL): The model’s visual language. Pick one closest to the style you want (photoreal, anime, illustration, etc.).

• Image Styles (max 3): Lightweight style/subject adapters (low rank adaptations) that “steer” the base model (e.g., anime line art refinement, cinematic color grading, painterly brushwork). Use them to add texture or enforce a genre. Note: it is important to only use styles that are relevant to the agent style and its purpose you're targetting, do not over use them.

• Base Prompt: A short, reusable prompt prefix that locks your agent’s identity and style. User prompts get appended after this.

• Agent Token (ERC‑20): The currency for image generation with that agent. Users can buy tokens in‑app on the agent page and spend them to generate images.

---

Quick Start (TL;DR)

1. Click Create a New Agent in the left sidebar.

2. Name your agent and set a symbol.

3. Choose an SDXL base model that matches your style goal.

4. Select up to 3 LoRAs that reinforce (not fight) your style. Note: sometimes base-models will perform better without any LoRAs, depending on the desired output.

5. Write a Base Prompt that defines the agent’s voice (keep it tight and style‑anchored).

6. Pay 0.01 ETH to deploy the agent.

7. (Optional) Buy your agent’s tokens to generate images and to kick‑start usage.

8. Share your agent link. Anyone can buy the token and generate images on the Explore page.

---

Creating an Agent — Step by Step

1. Open the Creator Dashboard → click Create a New Agent.

2. Basics

   • Agent Name: Human‑readable brand (e.g., Neon Portrait Studio).

   • Token Symbol: Short ticker (e.g., NEON).

3. Model & Style

   • Base Model (Max one): Pick the one aligned to your intended genre.

   • Styles (Max 3): Add only what you need. Prefer complementing roles (e.g., subject detail + style finish + lighting/color). Avoid multiple LoRAs that try to do the same thing.

4. Base Prompt

   • Keep it compact (1–2 clauses). It will always be placed first in the final prompt.

   • Encode style, medium, lighting, camera/shot, composition—leave the variable subject/story space for users.

   • Example (photoreal portrait):

     ultra‑detailed studio portrait, soft light, 85mm, shallow depth of field, clean background, natural skin texture, sharp focus

5. Buy Agent Tokens (Optional but recommended)

   • On the create page, choose an amount of tokens to purchase.

6. Review & Deploy

   • Confirm your selections and click Create. Approve the 0.001 ETH transaction.

Designing a strong Base Prompt

Your base prompt defines the agent’s consistent style. It should:

• Lock in the look (style, medium, lighting, camera, composition) and stay agnostic about specific subjects.

• Be short and information‑dense (aim for ~15–40 tokens). No rambling, no redundant superlatives.

• Use an SDXL‑friendly structure:

  • [style/medium] → [lighting] → [camera/shot] → [composition] → [quality tokens (1–2 max)]

• Reserve subject specifics (character, pose, outfit, scene) for the user prompt.

Good Base Prompt Patterns

• Photoreal portrait:

  • studio portrait, soft light, 85mm, centered composition, natural skin texture, sharp focus

• Anime character:

  • anime key art, clean linework, cel shading, soft rim light, three‑quarter view, crisp edges

• Digital illustration / concept:

  • digital illustration, painterly brushwork, volumetric light, rule of thirds, vivid but balanced palette

• Pixel art / retro:

  • pixel art, 16‑bit palette, crisp dithering, isometric, clean silhouettes

Tip: If your agent is about one subject class (e.g., cars, mechs, product renders), you can add that class into the base prompt (studio product render, softbox lighting, 50mm, seamless background). Keep it generic enough that users can still vary specifics (model, color, environment).

Choosing the Right Base Model

1) Start with the Base Model (do not fight it)

Pick the SDXL model that already speaks your style:

• Photoreal → photoreal‑leaning SDXL variant

• Anime → anime‑trained SDXL variant

• Illustration/Concept → art‑leaning SDXL variant

• Pixel/Low‑poly/Graphic → stylized SDXL variant

The closer the base model is to your target look, the fewer styles you’ll need, and the more stable your outputs will be.

2) Add styles with clear, non‑overlapping roles (max 3)

• Subject/detail styles: micro‑texture, anatomy, line quality.

• Artistic finish styles: painterly vs. glossy vs. gritty.

• Lighting/color styles: cinematic grade, neon nights, filmic tones.

Avoid:

• Stacking multiple styles that push the same aesthetic in different directions (they can cancel or produce artifacts).

• Mixing generalist styles with highly niche ones unless you’re targeting that niche.

Prompting for Best Results (for you and for your users)

How prompts are built

Final Prompt = [Agent Base Prompt] + [User Prompt]

Your base prompt sets the style; the user prompt adds subject, scene, and directives.

Writing user prompts that shine with your agent

• Lead with the subject: a retro sports car drifting at night in the rain

• Add setting/props only as needed: neon city street, reflections on wet asphalt

• Keep style tokens minimal in user prompts so they don’t fight the agent’s base style.

• Use camera & composition only when you want a clear change (e.g., low angle close‑up). Otherwise, let the base prompt control it.

Example (agent: neon photoreal night portraits)

• Base Prompt: neon city portrait, soft rim light, shallow depth of field, 85mm, centered composition, sharp focus

• User Prompt: young woman with blue bob haircut, clear umbrella, rain droplets, faint steam, city signs in background

Agent Quality Checklist (before you publish)

• The base prompt is short, style‑anchored, and subject‑agnostic.

• The base model matches your target genre.

• You use ≤ 3 styles, each with a distinct purpose.

Troubleshooting

• Inconsistent outputs:

  • Reduce Style count; ensure Styles don’t conflict.

  • Simplify the base prompt; let the model do the heavy lifting.

• Overcooked or artifact‑heavy images:

  • Remove the most aggressive Style or tone down style descriptors.

• Flat lighting / muddy colors:

  • Add a single clear lighting token (e.g., soft light, golden hour, rim light). Remove competing lighting terms.

• Users keep overriding your style:

  • Strengthen style cues in the base prompt and advise minimal style tokens in user prompts.