tried to do a similar thing with stable diffusion. for some reason, they give isometric view, and not a side-scrolling view. what is this, Super Mario RPG?!

Looks like Metal Slug scenarios. Beautiful, but not ideal island for gameplay. I may try to do an AI map one day. Good idea.

I recently ditched my 6.5 year old gaming laptop and switched to a new rig capable of running local AI models. As luck would have it, a new SOTA image GenAI model called Flux came out few weeks ago.

When I saw this map, I wondered how Flux would handle Worms map creation. Here's the result:

Looks very good but the number of colors exceeds W:A. Can you regulate it?

Quote from: MonkeyIsland on August 15, 2024, 07:37 PMLooks very good but the number of colors exceeds W:A. Can you regulate it?

Probably there are some ways to do it (you should be even able to train a model on a specific image set, such as BnG maps), but I am not experienced with running local AI models yet.

Some other generations:

Thats pretty nice already Triad.

I made the backgrounds for the Big TTRR mappack with Adobe Firefly. I could not set the color limitation in Firefly but just made that in Photoshop by just limit the colors and PS does a good job with it.

Guess these maps would look good after that too.

Man, you could make some really interesting forest BnG maps with this! If they were complex enough!

Quote from: h3oCharles on August 15, 2024, 12:55 PMtried to do a similar thing with stable diffusion. for some reason, they give isometric view, and not a side-scrolling view. what is this, Super Mario RPG?!

You just need to be lucky with AI that draws photos.

Quote from: FoxHound on August 15, 2024, 03:52 PMLooks like Metal Slug scenarios. Beautiful, but not ideal island for gameplay. I may try to do an AI map one day. Good idea.

Yeah, that was my first try to do something related to WA with AI.
Though, AI that I used obviously didn't knew about WA, so that's why gameplay on this map is... let's say - not enjoyable.

Quote from: Triad on August 15, 2024, 08:17 PM

Quote from: MonkeyIsland on August 15, 2024, 07:37 PMLooks very good but the number of colors exceeds W:A. Can you regulate it?

Probably there are some ways to do it (you should be even able to train a model on a specific image set, such as BnG maps), but I am not experienced with running local AI models yet.

Some other generations:

Damn, that's GREAT results for an AI

Quote from: Dantroil on August 16, 2024, 11:54 AMDamn, that's GREAT results for an AI

Thanks, I use text-based AIs like ChatGPT to refine my prompts. Here's a quick guide on how I generated these maps:

• Find a good text-based AI
• Some powerful text AIs with free quotas:
• Claude 3.5 Sonnet by Anthropic
• ChatGPT-4o by OpenAI
• Gemini Pro 1.5 by Google
• Llama 3.1 405B by Meta
• You can also use the direct chat feature of the Chatbot Arena (and you can check the leaderboard to get an idea of which models are stronger than others).
• Give the text AI the following prompt as your first message:
  
  Code Select Expand
  You are Dreamweaver, an AI assistant specialized in creating and refining prompts for image generation. Your primary function is to generate two distinct prompts for each image request: one optimized for T5 Encoder Text Linear Projection, and another for CLIP MLP.
Key responsibilities:
1. Analyze user requests for image generation.
2. Create two separate prompts for each request:
   a. T5-optimized prompt
   b. CLIP-optimized prompt
3. Refine existing prompts to improve image generation results.
4. Explain the differences between T5 and CLIP prompts when needed.
Guidelines for prompt creation:
1. T5-optimized prompts:
   - Focus on clear, descriptive language
   - Use complete sentences or phrases
   - Emphasize overall scene composition and context
   - Include specific details about objects, colors, and spatial relationships
2. CLIP-optimized prompts:
   - Use concise, keyword-rich descriptions
   - Separate concepts with commas
   - Prioritize visual attributes and style descriptors
   - Include artistic references or specific visual techniques when relevant
Always provide both T5 and CLIP prompts for each request, clearly labeled. Be prepared to explain your choices and refine prompts based on user feedback. Maintain a balanced approach between creativity and technical accuracy in your prompt generation.
• Describe your prompt.
• Use as much detail as you can.
• You can even use image references like I did:
  
  
  
• If you are not satisfied with the results, ask text AI to troubleshoot your prompt.
  
  
  
• Repeat the process till you're satisfied.
  
  

Today I did another test with AI.

While I was not able to create complex maps from scratch, I thought, how about remixing existing WA maps? So, I gave the AI a BnG map, and used it as the base layout for the image I am prompting. The workflow looks like this:



And the results are not too bad (you should be able to use them on WA without any issues):

Excellent work Triad! We should play on one of these!