Instructions to use lodestones/Kromoss with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use lodestones/Kromoss with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("lodestones/Kromoss", 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
- Local Apps Settings
- Draw Things
- DiffusionBee
BF16 + DMD/DMD2 gets crispy for me, but 50-step BF16 looks way better (Even longer)
#8
by uzumymwka - opened
hey, been testing Kromoss a lot and wanted to ask about something with the BF16 version, im on ComfyUI, 8GB VRAM, 32GB RAM
what i'm seeing is this:
- BF16 + DMD / DMD2 at around 10 steps, CFG 1, strength 1.0 gives me a weird crispy / overprocessed look
- skin gets too harsh, too textured, sometimes kinda dirty or crunchy
- faces can also start looking a bit off in a way that doesn't feel like natural camera detail
but when i do: - BF16 without turbo, 50 steps
- or even BF16 with only a realism / amateur snapshot style lora
the results look a lot better and more natural
so at this point it doesn't seem like BF16 itself is the issue
it feels more like the problem happens specifically when i use DMD / DMD2 on top of BF16
the strange part is that the INT8 version with turbo often looks cleaner for me, even if it's a bit smoother
so i wanted to ask:
- is this expected with the current DMD / DMD2 loras on Kromoss BF16?
- is there any sampler / scheduler / setting combo you recommend specifically for BF16 + turbo?
- does BF16 tend to react more aggressively to DMD / DMD2 than the INT8 convrot version.??
really liking the model overall, just trying to understand if i'm doing something wrong or if this is just how the current turbo loras behave on BF16
thanks