ad-llama

Create everything you need to start advertising your product ideas!!

Example Product Ideas::

• Organic Dog Food
• Sugar-free Soda
• Coffee
• Electric Car
• Toothpaste

Preamble

Originally built in a weekend for the Streamlit LLM Hackathon 2023 (https://streamlit.io/community/llm-hackathon-2023), the idea originated from some Langchain documentation about using models from Replicate (https://replicate.com/explore) in a SimpleSequentialChain (https://api.python.langchain.com/en/latest/chains/langchain.chains.sequential.SimpleSequentialChain.html), and a desire to move out of the OpenAI ecosystem and start using Llama 2 (https://ai.meta.com/llama/).

• The app combines several language and text to image models and uses Streamlit for the User Interface.
• It uses a SequentialChain from LangChain to generate the product name, slogan and logo description in sequence with calls to Llama 2.
• Then this logo description is fed into 3 different text to image models to generate the logos.

How this works

Enter an API key for Replicate (create an account, and follow the standard steps), enter a short description of a product, and sit back and let the app give you the following::

• A name for the product
• A slogan for the product
• 3 possible logos you can use for the product

Models Used

Large Language Models:

• meta/llama-2-13b-chat: https://replicate.com/meta/llama-2-13b-chat

Text To Image Models:

• galleri5/icons: https://replicate.com/galleri5/icons
• mejiabrayan/logoai: https://replicate.com/mejiabrayan/logoai
• cjwbw/magifactory-t-shirt-diffusion: https://replicate.com/cjwbw/magifactory-t-shirt-diffusion

To-Do and Notes

• Parametrize inputs for all the different Replicate models
• Llama 2 temperature, top_k, etc need to be experimented with to standardize text output format
• Understand how to use negative prompting better for the text2image models
• Try non-logo image models, Stable diffusion based
• Sometimes the logo description (which is fed into the image models) triggers an NSFW filter warning, the way to fix this is to usually just re-run, but is there a better solution?
• Some of the slogans are repetitive and a bit.. cheesy?, changing Llama 2 temperature does not seem to affect this too much.
• Sometimes the logo models don't understand that they need to generate a product logo, but generate images instead. Needs to be addressed.
• Try llama-70b
• The code needs to be modularized
• Image resolutions need to be optimized
• Add in another step in the pipeline:: placing ad logos on products (rucksacks, bottles, coffee mugs etc.)
• Perhaps also try to generate an ad jingle