import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Bl4ckSpaces/Project-Cosmos-V1.1", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]Project-Cosmos-V1.1 (Stabilized Edition)
Cosmos V1.1 fixes the tensor explosion issue of V1.0. It moves away from theoretical LLM DARE scaling and uses Sparse Transfer Learning.
The Fix
- Removed Aggressive Scaling: We no longer multiply weights by 10x. This eliminates the "deep-fried" noise artifacts.
- Sparse Cleaning: We still Drop (Prune) 40-80% of the weak noise data from NoobAI, ensuring only the clean "bone structure" and "core style" are transferred to IllumiYume.
- Result: The sharpness of DARE, without the destruction.
Recommended Settings
- Sampler: Euler a / DPM++ 2M SDE
- CFG: 5.0 - 7.0
- Steps: 28+
- Downloads last month
- 18