|
--- |
|
license: mit |
|
pipeline_tag: image-to-image |
|
tags: |
|
- ControlNetModel |
|
- StableDiffusionXLControlNetInpaintPipeline |
|
- 'ControlNetModel' |
|
- stable-diffusion-xl |
|
- inpainting |
|
- light-estimation |
|
- relighting |
|
base_model: stabilityai/stable-diffusion-xl-base-1.0 |
|
instance_prompt: >- |
|
a perfect mirrored reflective chrome ball sphere |
|
inference: false |
|
--- |
|
# Chromeball SDXL EV0 Wild 1500 |
|
|
|
**THIS IS NOT A PART OF DIFFUSIONLIGHT PAPER** |
|
|
|
Unlike [Chromeball-sdxl-ev0](https://huggingface.co/pureexe/chromeball-sdxl-ev0) repo, this controlnet is fine-tuned use 1500 in the wild image (Taken from Unsplash) instead of fully synthetic dataset. |
|
|
|
We use a ControlNet to output the chrome ball from SDXL instead of using LoRA which enables the SDXL Turbo to output lighting in one denosing step |
|
|
|
If the speed is less of a concern for you and want to focus on the quality. Please check out the LoRA model at [DiffusionLight](https://huggingface.co/DiffusionLight/DiffusionLight) |
|
|
|
|
|
|
|
## ControlNet vs LoRA: The SDXL Turbo |
|
|
|
TODO: I will explained later that SDXL Turbo makes LoRA impossible to give an input signal in one step. So, we use a control to guide output instead |
|
|
|
|
|
## Download model |
|
|
|
Weights for this model are available in Safetensors format. |
|
|
|
[Download](/DiffusionLight/DiffusionLight/tree/main) them in the Files & versions tab. |
|
|
|
## Trigger words |
|
|
|
| Chromeball | Prompt | |
|
| ----------- | ----------- | |
|
| EV0 | a perfect mirrored reflective chrome ball sphere | |
|
|
|
## Chromeball generation |
|
TODO: I will update how to generate chrome ball from controlnet in one Unet pass. |
|
|
|
|
|
PS. This repo is currently supported outing the EV0. We might support different EV in the future. |
