ArtificialGuyBR

Home / Blog / Fixing things

Generate iOS App Icons with Stable Diffusion: IconsMI

IconsMI is a Stable Diffusion checkpoint for iOS-style app icons. Trigger word, prompt recipe, negative prompt, and the 2k vs 5500 checkpoint tradeoff.

1 sources cited Fixing things

On this page

  1. What the model does
  2. Trigger word and prompt recipe
  3. Two checkpoints: 2k vs 5500
  4. How to use it
  5. Common errors and fixes
  6. FAQ
  7. Sources

Generate iOS App Icons with Stable Diffusion: IconsMI

IconsMI is a Stable Diffusion text-to-image checkpoint by ArtificialGuyBR fine-tuned specifically to generate iOS-style app icons. Instead of fighting generic prompts for a square icon that reads like an app tile, you activate the model with the trigger word IconsMi, describe the interface you want, and get icon-style output in a single pass. The Hugging Face model card reports 144 likes and 90 downloads.

What the model does

IconsMI is a fine-tuned Stable Diffusion checkpoint — not a control network and not a LoRA — that has specialized toward app-icon visuals. The technique the author recommends in the model card is to describe the kind of app you want rather than abstract icon vocabulary. Describe the interface, e.g. a news app, music app, or sports app, and then describe what should appear in the image, e.g. "a reporter microphone". The author notes that Stable Diffusion does not yet understand the abstract "icon" concept well, so concrete subject descriptions produce better results.

Trigger word and prompt recipe

Usage is a single condition, stated in the model card: you have to use the word IconsMi in the prompt. The author's best positive prompt from testing:

highly detailed, trending on artstation, ios icon app, IconsMi

And the negative prompt that performed best in the author's tests:

out of frame, duplicate, watermark, signature, text, ugly, sketch, deformed, mutated, blurry, mutilated, ugly sketch

The working pattern is: IconsMi + a concrete description of the app subject, e.g. IconsMi, a reporter microphone for a news app icon.

Two checkpoints: 2k vs 5500

The model was trained to 7,200 steps with checkpoints saved every 700 steps after 2,000. Two are shipped in the repository, both triggered with the same IconsMi prompt:

For icon concepts that need recognizable subjects and variation, the 2k checkpoint is the better default; for polish, try the 5500. Nothing prevents sampling both on the same prompt and picking favorites.

How to use it

The repository ships a complete diffusers StableDiffusionPipeline (feature extractor, safety checker, scheduler, text encoder, tokenizer, UNet, VAE) plus the two .ckpt checkpoints, and it is tagged endpoints_compatible, so it can run through Hugging Face Inference Endpoints. In diffusers, load it with the standard Stable Diffusion pipeline class and pass the trigger-word prompt above. In Automatic1111/ComfyUI-style workflows, load the .ckpt files directly.

Common errors and fixes

Error | Cause | Fix | Source

|---|---|---|---

Generic output, no icon look | IconsMi trigger word missing | Keep IconsMi in the prompt | Model card Abstract "icon" request fails | SD doesn't understand icon abstraction | Describe a concrete app subject, e.g. "a reporter microphone" | Model card Bad edges, artifacts | Missing negative prompt | Use the negative prompt from the card | Model card Output lacks variety | 5500 checkpoint overfitted to quality | Use the 2k checkpoint | Model card

FAQ

What is the trigger word? IconsMi — the model card states it must be in the prompt.

Which base model was it trained on? The card does not state a specific base; it is distributed as a diffusers StableDiffusionPipeline checkpoint, created November 2022.

Which checkpoint should I use? 2k for creativity and variety, 5500 for cleaner image quality — the two trade off against each other.

What license applies? CreativeML OpenRAIL-M.

Sources