This repository contains a project that utilizes DeepSeek Janus Pro for text-to-image generation and Real-ESRGAN for super-resolution upscaling. The goal is to generate and enhance realistic cat-friendly environments based on descriptive prompts.
- Image Generation: Uses DeepSeek Janus Pro to generate cat-friendly room environments based on detailed prompts.
- Image Upscaling: Leverages Real-ESRGAN to enhance image quality for better visual fidelity.
- Customizable Inputs: Users can define room attributes like lighting, furniture, and interactive cat elements.
git clone https://github.com/UlianaDzhumok/deepseek_janus_pro_experiments.git
cd deepseek_janus_pro_experimentsAll dependencies installation place in the beggining of the notebook
Define image and question in the script and run:
# Load the image
image = Image.open("./frame_0.jpg")
# Define inputs
question = "How many cats do you see in the photo? Describe their appearance and behaviour."
seed = 42
top_p = 0.8
temperature = 0.5
# Call the function with the correct image format
multimodal_understanding([image], question, seed, top_p, temperature)
"I see two cats in the photo. One cat is white and is standing near the wooden shelf on the left side of the image. The other cat is mostly white with some black markings and is sitting on the floor near the center of the image. Both cats appear to be calmly observing their surroundings."
Modify refined_prompt in the script and run:
from generate_image import generate_image
generated_images = generate_image(
prompt="Describe the cat-friendly room",
seed=4937,
guidance=11,
t2i_temperature=0.8
)After generating an image, apply Real-ESRGAN upscaling:
from upsample import image_upsample
upscaled_image = image_upsample(generated_images[0])
upscaled_image.show()"A cozy and inviting room for cats with soft furniture, natural lighting, interactive toys, and climbing structures."
This project is open-source and available under the MIT License.