Text-To-Pokemon

Main Features & Characteristics: Fine-tuned on Stable Diffusion using BLIP captioned Pokémon images, capable of generating Pokémon-style images from text descriptions. No prompt engineering is required to generate characters.

Usage & Input Parameters:

• prompt (string): Input prompt, e.g., "Yoda".
• num_outputs (integer): Number of images to output, options range from 1 to 4, default is 1.
• num_inference_steps (integer): Number of denoising steps, range 1 to 50, default is 25.
• guidance_scale (number): Scale for classifier-free guidance, range 1 to 20, default is 7.5.
• seed (integer): Random seed, leave blank to randomize.

Integration Methods:

1. Node.js / Python API: Install the Replicate client library, set the REPLICATE_API_TOKEN environment variable, and run the model using the version hash ff6cc781634191dd3c49097a615d2fc01b0a8aae31c448e55039a04dcbf36bba to get image URLs or write to disk.
2. HTTP API: Send a POST request to https://api.replicate.com/v1/predictions using curl.
3. Cog: Install Cog and use cog predict r8.im/lambdal/text-to-pokemon@sha256:... to download and run the model locally.
4. Docker: Run the model in a local GPU environment using the docker run command and call the API.

Target Users & Core Advantages: Aimed at AI art enthusiasts, game developers, and Pokémon fans. Core advantages include specialized fine-tuning for Pokémon style, highly thematic generation results, and support for open-source local deployment.

Pricing & Cost: Runs on Replicate, costing approximately $0.047 per run (around 21 runs per $1). Runs on Nvidia T4 GPU hardware, with predictions typically completing within 4 minutes. The model is open source and can be run for free on local hardware.

Typical Use Cases: Input "Yoda" to generate a Pokémon-style Yoda master image; input "Girl with a pearl earring", "Donald Trump", etc., to generate fun Pokémon-style characters.

Model Background: Trained by Justin Pinkney at Lambda Labs using 2xA6000 GPUs on Lambda GPU Cloud for around 15,000 steps (about 6 hours, at a cost of about $10). Datasets and weights are open-sourced on Hugging Face.