- Added [DINOv2](https://github.com/facebookresearch/dinov2) for high-performance visual features (contributed by [@Laurent2916](https://github.com/Laurent2916))
- Added [FreeU](https://github.com/ChenyangSi/FreeU) for improved quality at no cost (contributed by [@isamu-isozaki](https://github.com/isamu-isozaki))
- Added [Restart Sampling](https://github.com/Newbeeer/diffusion_restart_sampling) for improved image generation ([example](https://github.com/Newbeeer/diffusion_restart_sampling/issues/4))
- Added [Self-Attention Guidance](https://github.com/KU-CVLAB/Self-Attention-Guidance/) to avoid e.g. too smooth images ([example](https://github.com/SusungHong/Self-Attention-Guidance/issues/4))
- Added [T2I-Adapter](https://github.com/TencentARC/T2I-Adapter) for extra guidance ([example](https://github.com/TencentARC/T2I-Adapter/discussions/93))
- Added [MultiDiffusion](https://github.com/omerbt/MultiDiffusion) for e.g. panorama images
Goal: turn Refiners' mascot into a [Dragon Quest Slime](https://en.wikipedia.org/wiki/Slime_(Dragon_Quest)) plush in a one-shot manner thanks to a powerful combo of adapters:
- IP-Adapter: to capture the Slime plush visual appearance into an image prompt (no prompt engineering needed)
- T2I-Adapter: to guide the generation with the mascot's geometry
- Self-Attention-Guidance (SAG): to increase the sharpness
> Note: this will download the original weights from https://huggingface.co/ which takes some time. If you already have this repo cloned locally, use the `--from /path/to/stabilityai/stable-diffusion-xl-base-1.0` option instead.
And then convert IP-Adapter and T2I-Adapter weights (note: SAG is parameter-free):
For now, given [finegrain](https://finegrain.ai)'s mission, we are focusing on image edition tasks. We support:
| Adapter | Foundation Model |
| ----------------- | ------- |
| LoRA | `SD15``SDXL` |
| ControlNets | `SD15` |
| Ref Only Control | `SD15` |
| IP-Adapter | `SD15``SDXL` |
| T2I-Adapter | `SD15``SDXL` |
## Motivation
At [Finegrain](https://finegrain.ai), we're on a mission to automate product photography. Given our "no human in the loop approach", nailing the quality of the outputs we generate is paramount to our success.
It's a framework to easily bridge the last mile quality gap of foundation models like Stable Diffusion or Segment Anything Model (SAM), by adapting them to specific tasks with lightweight trainable and composable patches.
It's because model adaptation is a new paradigm that goes beyond our specific use cases. Our hope is to help people looking at creating their own adapters save time, whatever the foundation model they're using.
We are huge fans of PyTorch (we actually were core committers to [Torch](http://torch.ch/) in another life), but we felt it's too low level for the specific model adaptation task: PyTorch models are generally hard to understand, and their adaptation requires intricate ad hoc code.
Instead, we needed:
- A model structure that's human readable so that you know what models do and how they work right here, right now
- A mechanism to easily inject parameters in some target layers, or between them
- A way to easily pass data (like a conditioning input) between layers even when deeply nested
- Native support for iconic adapter types like LoRAs and their community trained incarnations (hosted on [Civitai](http://civitai.com/) and the likes)
Refiners is designed to tackle all these challenges while remaining just one abstraction away from our beloved PyTorch.
## Key Concepts
### The Chain class
The `Chain` class is at the core of Refiners. It basically lets you express models as a composition of basic layers (linear, convolution, attention, etc) in a **declarative way**.
E.g.: this is how a Vision Transformer (ViT) looks like with Refiners:
```python
import torch
import refiners.fluxion.layers as fl
class ViT(fl.Chain):
# The Vision Transformer model structure is entirely defined in the constructor. It is
# ready-to-use right after i.e. no need to implement any forward function or add extra logic
The `Adapter` API lets you **easily patch models** by injecting parameters in targeted layers. It comes with built-in support for canonical adapter types like LoRA, but you can also implement your custom adapters with it.
E.g. to inject LoRA layers in all attention's linear layers:
If you're interested in understanding the diversity of use cases for foundation model adaptation (potentially beyond the specific adapters supported by Refiners), we suggest you take a look at these outstanding papers:
### SAM
- [Medical SAM Adapter](https://arxiv.org/abs/2304.12620)
