OdysseyXL-2.0!

Yesterday I was messing around with OdysseyXL and I noticed that it was really good at generating himan faces but when it tried to generate kandscapes, it did not work as well as I had hoped. As much as I wanted to just forgrt it and move on, but I couldn't so I decided to fine-tune the modle to generate landscapes and scenery with more realism.

I trained OdysseyXL-2.0 on 4 A100 GPUs from Lambda Cloud because in my opinion Lambda provides the cheapest rates for the best GPUs. I used a mixture of public datasets and hand-collected some images. Unlike OdysseyXL 1.0, the new dataset had more scenery and landscapes than pictures of humans.

As you can see, the overall quality of OdysseyXL jumped so much that we can compare it to the SOTA FLUX models, specifically FLUX.1 [dev]. While FLUX does have some better generations e.g the clouds. We can see that OdysseyXL 2.0 generated the mountain better. The weights of OdysseyXL is now public on my HuggingFace account! You can also use the model using this code snippet:

from diffusers import DiffusionPipeline
 
pipe = DiffusionPipeline.from_pretrained("Spestly/OdysseyXL-2.0")
 
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

~ Aayan Mishra (NotSpestly)