Instructions to use ByteDance/SDXL-Lightning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use ByteDance/SDXL-Lightning with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("ByteDance/SDXL-Lightning", 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
- Draw Things
- DiffusionBee
Is there something wrong with the 8-step model?
For all speed models especially lora need
At u case connect sdxl_lightning_8step.safetensors + sdxl_lightning_8step_lora.safetensors
sdxl_lightning_8step_lora.safetensors
https://huggingface.co/ByteDance/SDXL-Lightning/blob/main/sdxl_lightning_8step_lora.safetensors
For all speed models especially lora need
At u case connect sdxl_lightning_8step.safetensors + sdxl_lightning_8step_lora.safetensors
Never listen to random people like this who have no idea what they are talking about. This lora is designed to allow a full SDXL base model to create 8 step images. This is NOT for lighting models. A little research would easily tell you that.
The only thing I'm getting are abstract images.
The reason you are getting this is because you are trying to run high cfg on a lighting model. You run these no higher than 2 cfg. Recommended CFG on any lighting/dmd/hyper/super, etc models is 1-2 nothing over 2 ever. 1 is ideal using no negative prompts as SDXL doesn't like negative prompts anyways. If you need a negative prompt you should still run the CFG low. I never higher than 1.2 or 1.3. 5 is way too high and that is your issue.
I know this is an old post from 2024 but it doesn't matter. Anyone who finds this post having the same issues may fight this response useful so I'm replying anyways.
