Instructions to use Hippotes/NVIDIA-PiD-v1.5-repackages with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use Hippotes/NVIDIA-PiD-v1.5-repackages with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("Hippotes/NVIDIA-PiD-v1.5-repackages", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
2048 to 8192 workflow
Hey! Made a small repo, hasn't gotten much attention (relatively new). For Krea workflows at 2048 to output to 8192 native with PID. Including the generation time you're looking at ~100s for the final output 8192x8192 image. Exponentially faster than a native gen (also Krea's sweet spot is 2k).
If you want to check it out or perhaps give a star, I'm very much looking for the community's awareness of the project.
https://github.com/CommanderTurtle/seamingly-epic - (AGPL)
Basically a seam-merger plus a small few bit of nodes that perform PID upscales on four quadrants. /tree/main/workflows Then, solving the seam programmatically with some advanced math with the quick swipe of a terminal, using laplacians and what's best described as an "upscale cloud" to determine where structure can be merged efficiently. Working extremely well. At this rate, I'd say 2048 to 8192 is a pretty reliable gen, especially if you preview prior by just doing typical 1024 images (until it's a decent gen), since it sucks waiting a long time for 2048 anyways.
Biggest downside to 2048 gens is you can never PID upscale. This changes that, by assuming four 1024x1024 quadrants. Appreciate your experiments and repacks/workflows. Community is king.