|
|
--- |
|
|
license: apache-2.0 |
|
|
pipeline_tag: text-to-image |
|
|
tags: |
|
|
- text-to-image |
|
|
- image-generation |
|
|
- yandex |
|
|
--- |
|
|
Alice AI ART dev |
|
|
--- |
|
|
by Yandex |
|
|
|
|
|
 |
|
|
|
|
|
Alice AI ART dev is 4.8B parameter diffusion UNet model capable of generating images from text prompts. |
|
|
|
|
|
Key features |
|
|
--- |
|
|
|
|
|
* **Relevance** A considerable amount of work was done to improve text-to-image alignment. According to the Side-by-Side evaluation, our model is competitive with Qwen-Image, despite being significantly smaller (4.8B parameters vs 20B parameters). |
|
|
* **Aesthetics** Our model is capable of generating high-quality images with a wide range of styles and themes. |
|
|
* **Accessibility** Alice AI ART dev is runnable on consumer-grade[^1] GPUs (for instance, NVIDIA RTX 3090) making it accessible to a wider audience. |
|
|
|
|
|
[^1] with weight offloading |
|
|
|
|
|
Usage |
|
|
--- |
|
|
The image generation pipeline can be loaded a follows |
|
|
```python |
|
|
pipe = YandexArtOSPipeline.from_pretrained( |
|
|
"yandex_art_os", |
|
|
cpu_offload=True |
|
|
) |
|
|
``` |
|
|
For memory-constrained GPUs we recommend to turn on `cpu_offload` flag: |
|
|
|
|
|
By default we use following sampling parameters: |
|
|
```python |
|
|
{ |
|
|
"num_inference_steps": 32, |
|
|
"cond_scale": 2.75, |
|
|
"unet_switch_timestep": 8, |
|
|
"karras_rho": 6.0, |
|
|
"method_name": "dpm-multistep", |
|
|
"sampler_kwargs": { |
|
|
"num_train_timesteps": 1000, |
|
|
"beta_start": 0.00001013, |
|
|
"beta_end": 0.019771934, |
|
|
"use_karras_sigmas": True, |
|
|
"algorithm_type": "sde-dpmsolver++" |
|
|
} |
|
|
} |
|
|
``` |
|
|
|
|
|
|
