svjack's picture
Upload 1392 files
43b7e92 verified
|
raw
history blame
28.7 kB
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Diffusion ๋ชจ๋ธ ํ‰๊ฐ€ํ•˜๊ธฐ[[evaluating-diffusion-models]]
<a target="_blank" href="https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/evaluation.ipynb">
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
</a>
[Stable Diffusion](https://huggingface.co/docs/diffusers/stable_diffusion)์™€ ๊ฐ™์€ ์ƒ์„ฑ ๋ชจ๋ธ์˜ ํ‰๊ฐ€๋Š” ์ฃผ๊ด€์ ์ธ ์„ฑ๊ฒฉ์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์‹ค๋ฌด์ž์™€ ์—ฐ๊ตฌ์ž๋กœ์„œ ์šฐ๋ฆฌ๋Š” ์ข…์ข… ๋‹ค์–‘ํ•œ ๊ฐ€๋Šฅ์„ฑ ์ค‘์—์„œ ์‹ ์ค‘ํ•œ ์„ ํƒ์„ ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ๋‹ค์–‘ํ•œ ์ƒ์„ฑ ๋ชจ๋ธ (GAN, Diffusion ๋“ฑ)์„ ์‚ฌ์šฉํ•  ๋•Œ ์–ด๋–ป๊ฒŒ ์„ ํƒํ•ด์•ผ ํ• ๊นŒ์š”?
์ •์„ฑ์ ์ธ ํ‰๊ฐ€๋Š” ๋ชจ๋ธ์˜ ์ด๋ฏธ์ง€ ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ฃผ๊ด€์ ์ธ ํ‰๊ฐ€์ด๋ฏ€๋กœ ์˜ค๋ฅ˜๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๊ณ  ๊ฒฐ์ •์— ์ž˜๋ชป๋œ ์˜ํ–ฅ์„ ๋ฏธ์น  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด, ์ •๋Ÿ‰์ ์ธ ํ‰๊ฐ€๋Š” ์ด๋ฏธ์ง€ ํ’ˆ์งˆ๊ณผ ์ง์ ‘์ ์ธ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๊ฐ–์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ผ๋ฐ˜์ ์œผ๋กœ ์ •์„ฑ์  ํ‰๊ฐ€์™€ ์ •๋Ÿ‰์  ํ‰๊ฐ€๋ฅผ ๋ชจ๋‘ ๊ณ ๋ คํ•˜๋Š” ๊ฒƒ์ด ๋” ๊ฐ•๋ ฅํ•œ ์‹ ํ˜ธ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ๋ชจ๋ธ ์„ ํƒ์— ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.
์ด ๋ฌธ์„œ์—์„œ๋Š” Diffusion ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ •์„ฑ์  ๋ฐ ์ •๋Ÿ‰์  ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์ƒ์„ธํžˆ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์ •๋Ÿ‰์  ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด์„œ๋Š” ํŠนํžˆ `diffusers`์™€ ํ•จ๊ป˜ ๊ตฌํ˜„ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์Šต๋‹ˆ๋‹ค.
์ด ๋ฌธ์„œ์—์„œ ๋ณด์—ฌ์ง„ ๋ฐฉ๋ฒ•๋“ค์€ ๊ธฐ๋ฐ˜ ์ƒ์„ฑ ๋ชจ๋ธ์„ ๊ณ ์ •์‹œํ‚ค๊ณ  ๋‹ค์–‘ํ•œ [๋…ธ์ด์ฆˆ ์Šค์ผ€์ค„๋Ÿฌ](https://huggingface.co/docs/diffusers/main/en/api/schedulers/overview)๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ์—๋„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
## ์‹œ๋‚˜๋ฆฌ์˜ค[[scenarios]]
๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŒŒ์ดํ”„๋ผ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ Diffusion ๋ชจ๋ธ์„ ๋‹ค๋ฃน๋‹ˆ๋‹ค:
- ํ…์ŠคํŠธ๋กœ ์•ˆ๋‚ด๋œ ์ด๋ฏธ์ง€ ์ƒ์„ฑ (์˜ˆ: [`StableDiffusionPipeline`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/text2img)).
- ์ž…๋ ฅ ์ด๋ฏธ์ง€์— ์ถ”๊ฐ€๋กœ ์กฐ๊ฑด์„ ๊ฑด ํ…์ŠคํŠธ๋กœ ์•ˆ๋‚ด๋œ ์ด๋ฏธ์ง€ ์ƒ์„ฑ (์˜ˆ: [`StableDiffusionImg2ImgPipeline`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/img2img) ๋ฐ [`StableDiffusionInstructPix2PixPipeline`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pix2pix)).
- ํด๋ž˜์Šค ์กฐ๊ฑดํ™”๋œ ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ (์˜ˆ: [`DiTPipeline`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/dit)).
## ์ •์„ฑ์  ํ‰๊ฐ€[[qualitative-evaluation]]
์ •์„ฑ์  ํ‰๊ฐ€๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์˜ ์ธ๊ฐ„ ํ‰๊ฐ€๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ํ’ˆ์งˆ์€ ๊ตฌ์„ฑ์„ฑ, ์ด๋ฏธ์ง€-ํ…์ŠคํŠธ ์ผ์น˜, ๊ณต๊ฐ„ ๊ด€๊ณ„ ๋“ฑ๊ณผ ๊ฐ™์€ ์ธก๋ฉด์—์„œ ์ธก์ •๋ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ ํ”„๋กฌํ”„ํŠธ๋Š” ์ฃผ๊ด€์ ์ธ ์ง€ํ‘œ์— ๋Œ€ํ•œ ์ผ์ •ํ•œ ๊ธฐ์ค€์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
DrawBench์™€ PartiPrompts๋Š” ์ •์„ฑ์ ์ธ ๋ฒค์น˜๋งˆํ‚น์— ์‚ฌ์šฉ๋˜๋Š” ํ”„๋กฌํ”„ํŠธ ๋ฐ์ดํ„ฐ์…‹์ž…๋‹ˆ๋‹ค. DrawBench์™€ PartiPrompts๋Š” ๊ฐ๊ฐ [Imagen](https://imagen.research.google/)๊ณผ [Parti](https://parti.research.google/)์—์„œ ์†Œ๊ฐœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
[Parti ๊ณต์‹ ์›น์‚ฌ์ดํŠธ](https://parti.research.google/)์—์„œ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค:
> PartiPrompts (P2)๋Š” ์ด ์ž‘์—…์˜ ์ผ๋ถ€๋กœ ๊ณต๊ฐœ๋˜๋Š” ์˜์–ด๋กœ ๋œ 1600๊ฐœ ์ด์ƒ์˜ ๋‹ค์–‘ํ•œ ํ”„๋กฌํ”„ํŠธ ์„ธํŠธ์ž…๋‹ˆ๋‹ค. P2๋Š” ๋‹ค์–‘ํ•œ ๋ฒ”์ฃผ์™€ ๋„์ „ ์ธก๋ฉด์—์„œ ๋ชจ๋ธ์˜ ๋Šฅ๋ ฅ์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
![parti-prompts](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/parti-prompts.png)
PartiPrompts๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์—ด์„ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค:
- ํ”„๋กฌํ”„ํŠธ (Prompt)
- ํ”„๋กฌํ”„ํŠธ์˜ ์นดํ…Œ๊ณ ๋ฆฌ (์˜ˆ: "Abstract", "World Knowledge" ๋“ฑ)
- ๋‚œ์ด๋„๋ฅผ ๋ฐ˜์˜ํ•œ ์ฑŒ๋ฆฐ์ง€ (์˜ˆ: "Basic", "Complex", "Writing & Symbols" ๋“ฑ)
์ด๋Ÿฌํ•œ ๋ฒค์น˜๋งˆํฌ๋Š” ์„œ๋กœ ๋‹ค๋ฅธ ์ด๋ฏธ์ง€ ์ƒ์„ฑ ๋ชจ๋ธ์„ ์ธ๊ฐ„ ํ‰๊ฐ€๋กœ ๋น„๊ตํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
์ด๋ฅผ ์œ„ํ•ด ๐Ÿงจ Diffusers ํŒ€์€ **Open Parti Prompts**๋ฅผ ๊ตฌ์ถ•ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” Parti Prompts๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์ปค๋ฎค๋‹ˆํ‹ฐ ๊ธฐ๋ฐ˜์˜ ์งˆ์  ๋ฒค์น˜๋งˆํฌ๋กœ, ์ตœ์ฒจ๋‹จ ์˜คํ”ˆ ์†Œ์Šค ํ™•์‚ฐ ๋ชจ๋ธ์„ ๋น„๊ตํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค:
- [Open Parti Prompts ๊ฒŒ์ž„](https://huggingface.co/spaces/OpenGenAI/open-parti-prompts): 10๊ฐœ์˜ parti prompt์— ๋Œ€ํ•ด 4๊ฐœ์˜ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€๊ฐ€ ์ œ์‹œ๋˜๋ฉฐ, ์‚ฌ์šฉ์ž๋Š” ํ”„๋กฌํ”„ํŠธ์— ๊ฐ€์žฅ ์ ํ•ฉํ•œ ์ด๋ฏธ์ง€๋ฅผ ์„ ํƒํ•ฉ๋‹ˆ๋‹ค.
- [Open Parti Prompts ๋ฆฌ๋”๋ณด๋“œ](https://huggingface.co/spaces/OpenGenAI/parti-prompts-leaderboard): ํ˜„์žฌ ์ตœ๊ณ ์˜ ์˜คํ”ˆ ์†Œ์Šค diffusion ๋ชจ๋ธ๋“ค์„ ์„œ๋กœ ๋น„๊ตํ•˜๋Š” ๋ฆฌ๋”๋ณด๋“œ์ž…๋‹ˆ๋‹ค.
์ด๋ฏธ์ง€๋ฅผ ์ˆ˜๋™์œผ๋กœ ๋น„๊ตํ•˜๋ ค๋ฉด, `diffusers`๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ช‡๊ฐ€์ง€ PartiPrompts๋ฅผ ์–ด๋–ป๊ฒŒ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ์•Œ์•„๋ด…์‹œ๋‹ค.
๋‹ค์Œ์€ ๋ช‡ ๊ฐ€์ง€ ๋‹ค๋ฅธ ๋„์ „์—์„œ ์ƒ˜ํ”Œ๋งํ•œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค: Basic, Complex, Linguistic Structures, Imagination, Writing & Symbols. ์—ฌ๊ธฐ์„œ๋Š” PartiPrompts๋ฅผ [๋ฐ์ดํ„ฐ์…‹](https://huggingface.co/datasets/nateraw/parti-prompts)์œผ๋กœ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
```python
from datasets import load_dataset
# prompts = load_dataset("nateraw/parti-prompts", split="train")
# prompts = prompts.shuffle()
# sample_prompts = [prompts[i]["Prompt"] for i in range(5)]
# Fixing these sample prompts in the interest of reproducibility.
sample_prompts = [
"a corgi",
"a hot air balloon with a yin-yang symbol, with the moon visible in the daytime sky",
"a car with no windows",
"a cube made of porcupine",
'The saying "BE EXCELLENT TO EACH OTHER" written on a red brick wall with a graffiti image of a green alien wearing a tuxedo. A yellow fire hydrant is on a sidewalk in the foreground.',
]
```
์ด์ œ ์ด๋Ÿฐ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Stable Diffusion ([v1-4 checkpoint](https://huggingface.co/CompVis/stable-diffusion-v1-4))๋ฅผ ์‚ฌ์šฉํ•œ ์ด๋ฏธ์ง€ ์ƒ์„ฑ์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค :
```python
import torch
seed = 0
generator = torch.manual_seed(seed)
images = sd_pipeline(sample_prompts, num_images_per_prompt=1, generator=generator).images
```
![parti-prompts-14](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/parti-prompts-14.png)
`num_images_per_prompt`๋ฅผ ์„ค์ •ํ•˜์—ฌ ๋™์ผํ•œ ํ”„๋กฌํ”„ํŠธ์— ๋Œ€ํ•ด ๋‹ค๋ฅธ ์ด๋ฏธ์ง€๋ฅผ ๋น„๊ตํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์ฒดํฌํฌ์ธํŠธ([v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5))๋กœ ๋™์ผํ•œ ํŒŒ์ดํ”„๋ผ์ธ์„ ์‹คํ–‰ํ•˜๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜ต๋‹ˆ๋‹ค:
![parti-prompts-15](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/parti-prompts-15.png)
๋‹ค์–‘ํ•œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋“  ํ”„๋กฌํ”„ํŠธ์—์„œ ์ƒ์„ฑ๋œ ์—ฌ๋Ÿฌ ์ด๋ฏธ์ง€๋“ค์ด ์ƒ์„ฑ๋˜๋ฉด (ํ‰๊ฐ€ ๊ณผ์ •์—์„œ) ์ด๋Ÿฌํ•œ ๊ฒฐ๊ณผ๋ฌผ๋“ค์€ ์‚ฌ๋žŒ ํ‰๊ฐ€์ž๋“ค์—๊ฒŒ ์ ์ˆ˜๋ฅผ ๋งค๊ธฐ๊ธฐ ์œ„ํ•ด ์ œ์‹œ๋ฉ๋‹ˆ๋‹ค. DrawBench์™€ PartiPrompts ๋ฒค์น˜๋งˆํฌ์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๊ฐ๊ฐ์˜ ๋…ผ๋ฌธ์„ ์ฐธ์กฐํ•˜์‹ญ์‹œ์˜ค.
<Tip>
๋ชจ๋ธ์ด ํ›ˆ๋ จ ์ค‘์ผ ๋•Œ ์ถ”๋ก  ์ƒ˜ํ”Œ์„ ์‚ดํŽด๋ณด๋Š” ๊ฒƒ์€ ํ›ˆ๋ จ ์ง„ํ–‰ ์ƒํ™ฉ์„ ์ธก์ •ํ•˜๋Š” ๋ฐ ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค. [ํ›ˆ๋ จ ์Šคํฌ๋ฆฝํŠธ](https://github.com/huggingface/diffusers/tree/main/examples/)์—์„œ๋Š” TensorBoard์™€ Weights & Biases์— ๋Œ€ํ•œ ์ถ”๊ฐ€ ์ง€์›๊ณผ ํ•จ๊ป˜ ์ด ์œ ํ‹ธ๋ฆฌํ‹ฐ๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.
</Tip>
## ์ •๋Ÿ‰์  ํ‰๊ฐ€[[quantitative-evaluation]]
์ด ์„น์…˜์—์„œ๋Š” ์„ธ ๊ฐ€์ง€ ๋‹ค๋ฅธ ํ™•์‚ฐ ํŒŒ์ดํ”„๋ผ์ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์•ˆ๋‚ดํ•ฉ๋‹ˆ๋‹ค:
- CLIP ์ ์ˆ˜
- CLIP ๋ฐฉํ–ฅ์„ฑ ์œ ์‚ฌ๋„
- FID
### ํ…์ŠคํŠธ ์•ˆ๋‚ด ์ด๋ฏธ์ง€ ์ƒ์„ฑ[[text-guided-image-generation]]
[CLIP ์ ์ˆ˜](https://arxiv.org/abs/2104.08718)๋Š” ์ด๋ฏธ์ง€-์บก์…˜ ์Œ์˜ ํ˜ธํ™˜์„ฑ์„ ์ธก์ •ํ•ฉ๋‹ˆ๋‹ค. ๋†’์€ CLIP ์ ์ˆ˜๋Š” ๋†’์€ ํ˜ธํ™˜์„ฑ๐Ÿ”ผ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค. CLIP ์ ์ˆ˜๋Š” ์ด๋ฏธ์ง€์™€ ์บก์…˜ ์‚ฌ์ด์˜ ์˜๋ฏธ์  ์œ ์‚ฌ์„ฑ์œผ๋กœ ์ƒ๊ฐํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. CLIP ์ ์ˆ˜๋Š” ์ธ๊ฐ„ ํŒ๋‹จ๊ณผ ๋†’์€ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
[`StableDiffusionPipeline`]์„ ์ผ๋‹จ ๋กœ๋“œํ•ด๋ด…์‹œ๋‹ค:
```python
from diffusers import StableDiffusionPipeline
import torch
model_ckpt = "CompVis/stable-diffusion-v1-4"
sd_pipeline = StableDiffusionPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16).to("cuda")
```
์—ฌ๋Ÿฌ ๊ฐœ์˜ ํ”„๋กฌํ”„ํŠธ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค:
```python
prompts = [
"a photo of an astronaut riding a horse on mars",
"A high tech solarpunk utopia in the Amazon rainforest",
"A pikachu fine dining with a view to the Eiffel Tower",
"A mecha robot in a favela in expressionist style",
"an insect robot preparing a delicious meal",
"A small cabin on top of a snowy mountain in the style of Disney, artstation",
]
images = sd_pipeline(prompts, num_images_per_prompt=1, output_type="np").images
print(images.shape)
# (6, 512, 512, 3)
```
๊ทธ๋Ÿฌ๊ณ  ๋‚˜์„œ CLIP ์ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
```python
from torchmetrics.functional.multimodal import clip_score
from functools import partial
clip_score_fn = partial(clip_score, model_name_or_path="openai/clip-vit-base-patch16")
def calculate_clip_score(images, prompts):
images_int = (images * 255).astype("uint8")
clip_score = clip_score_fn(torch.from_numpy(images_int).permute(0, 3, 1, 2), prompts).detach()
return round(float(clip_score), 4)
sd_clip_score = calculate_clip_score(images, prompts)
print(f"CLIP score: {sd_clip_score}")
# CLIP score: 35.7038
```
์œ„์˜ ์˜ˆ์ œ์—์„œ๋Š” ๊ฐ ํ”„๋กฌํ”„ํŠธ ๋‹น ํ•˜๋‚˜์˜ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค. ๋งŒ์•ฝ ํ”„๋กฌํ”„ํŠธ ๋‹น ์—ฌ๋Ÿฌ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•œ๋‹ค๋ฉด, ํ”„๋กฌํ”„ํŠธ ๋‹น ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€์˜ ํ‰๊ท  ์ ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
์ด์ œ [`StableDiffusionPipeline`]๊ณผ ํ˜ธํ™˜๋˜๋Š” ๋‘ ๊ฐœ์˜ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋น„๊ตํ•˜๋ ค๋ฉด, ํŒŒ์ดํ”„๋ผ์ธ์„ ํ˜ธ์ถœํ•  ๋•Œ generator๋ฅผ ์ „๋‹ฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋จผ์ €, ๊ณ ์ •๋œ ์‹œ๋“œ๋กœ [v1-4 Stable Diffusion ์ฒดํฌํฌ์ธํŠธ](https://huggingface.co/CompVis/stable-diffusion-v1-4)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค:
```python
seed = 0
generator = torch.manual_seed(seed)
images = sd_pipeline(prompts, num_images_per_prompt=1, generator=generator, output_type="np").images
```
๊ทธ๋Ÿฐ ๋‹ค์Œ [v1-5 checkpoint](https://huggingface.co/runwayml/stable-diffusion-v1-5)๋ฅผ ๋กœ๋“œํ•˜์—ฌ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค:
```python
model_ckpt_1_5 = "runwayml/stable-diffusion-v1-5"
sd_pipeline_1_5 = StableDiffusionPipeline.from_pretrained(model_ckpt_1_5, torch_dtype=weight_dtype).to(device)
images_1_5 = sd_pipeline_1_5(prompts, num_images_per_prompt=1, generator=generator, output_type="np").images
```
๊ทธ๋ฆฌ๊ณ  ๋งˆ์ง€๋ง‰์œผ๋กœ CLIP ์ ์ˆ˜๋ฅผ ๋น„๊ตํ•ฉ๋‹ˆ๋‹ค:
```python
sd_clip_score_1_4 = calculate_clip_score(images, prompts)
print(f"CLIP Score with v-1-4: {sd_clip_score_1_4}")
# CLIP Score with v-1-4: 34.9102
sd_clip_score_1_5 = calculate_clip_score(images_1_5, prompts)
print(f"CLIP Score with v-1-5: {sd_clip_score_1_5}")
# CLIP Score with v-1-5: 36.2137
```
[v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5) ์ฒดํฌํฌ์ธํŠธ๊ฐ€ ์ด์ „ ๋ฒ„์ „๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ CLIP ์ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉํ•œ ํ”„๋กฌํ”„ํŠธ์˜ ์ˆ˜๊ฐ€ ์ƒ๋‹นํžˆ ์ ์Šต๋‹ˆ๋‹ค. ๋ณด๋‹ค ์‹ค์šฉ์ ์ธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด์„œ๋Š” ์ด ์ˆ˜๋ฅผ ํ›จ์”ฌ ๋†’๊ฒŒ ์„ค์ •ํ•˜๊ณ , ํ”„๋กฌํ”„ํŠธ๋ฅผ ๋‹ค์–‘ํ•˜๊ฒŒ ์‚ฌ์šฉํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
<Tip warning={true}>
์ด ์ ์ˆ˜์—๋Š” ๋ช‡ ๊ฐ€์ง€ ์ œํ•œ ์‚ฌํ•ญ์ด ์žˆ์Šต๋‹ˆ๋‹ค. ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์˜ ์บก์…˜์€ ์›น์—์„œ ํฌ๋กค๋ง๋˜์–ด ์ด๋ฏธ์ง€์™€ ๊ด€๋ จ๋œ `alt` ๋ฐ ์œ ์‚ฌํ•œ ํƒœ๊ทธ์—์„œ ์ถ”์ถœ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ์ด๋“ค์€ ์ธ๊ฐ„์ด ์ด๋ฏธ์ง€๋ฅผ ์„ค๋ช…ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ๊ฒƒ๊ณผ ์ผ์น˜ํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์—ฌ๊ธฐ์„œ๋Š” ๋ช‡ ๊ฐ€์ง€ ํ”„๋กฌํ”„ํŠธ๋ฅผ "์—”์ง€๋‹ˆ์–ด๋ง"ํ•ด์•ผ ํ–ˆ์Šต๋‹ˆ๋‹ค.
</Tip>
### ์ด๋ฏธ์ง€ ์กฐ๊ฑดํ™”๋œ ํ…์ŠคํŠธ-์ด๋ฏธ์ง€ ์ƒ์„ฑ[[image-conditioned-text-to-image-generation]]
์ด ๊ฒฝ์šฐ, ์ƒ์„ฑ ํŒŒ์ดํ”„๋ผ์ธ์„ ์ž…๋ ฅ ์ด๋ฏธ์ง€์™€ ํ…์ŠคํŠธ ํ”„๋กฌํ”„ํŠธ๋กœ ์กฐ๊ฑดํ™”ํ•ฉ๋‹ˆ๋‹ค. [`StableDiffusionInstructPix2PixPipeline`]์„ ์˜ˆ๋กœ ๋“ค์–ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํŽธ์ง‘ ์ง€์‹œ๋ฌธ์„ ์ž…๋ ฅ ํ”„๋กฌํ”„ํŠธ๋กœ ์‚ฌ์šฉํ•˜๊ณ  ํŽธ์ง‘ํ•  ์ž…๋ ฅ ์ด๋ฏธ์ง€๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์Œ์€ ํ•˜๋‚˜์˜ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค:
![edit-instruction](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/edit-instruction.png)
๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ํ•œ ๊ฐ€์ง€ ์ „๋žต์€ ๋‘ ์ด๋ฏธ์ง€ ์บก์…˜ ๊ฐ„์˜ ๋ณ€๊ฒฝ๊ณผ([CLIP-Guided Domain Adaptation of Image Generators](https://arxiv.org/abs/2108.00946)์—์„œ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค) ํ•จ๊ป˜ ๋‘ ์ด๋ฏธ์ง€ ์‚ฌ์ด์˜ ๋ณ€๊ฒฝ์˜ ์ผ๊ด€์„ฑ์„ ์ธก์ •ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค ([CLIP](https://huggingface.co/docs/transformers/model_doc/clip) ๊ณต๊ฐ„์—์„œ). ์ด๋ฅผ "**CLIP ๋ฐฉํ–ฅ์„ฑ ์œ ์‚ฌ์„ฑ**"์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
- ์บก์…˜ 1์€ ํŽธ์ง‘ํ•  ์ด๋ฏธ์ง€ (์ด๋ฏธ์ง€ 1)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
- ์บก์…˜ 2๋Š” ํŽธ์ง‘๋œ ์ด๋ฏธ์ง€ (์ด๋ฏธ์ง€ 2)์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค. ํŽธ์ง‘ ์ง€์‹œ๋ฅผ ๋ฐ˜์˜ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
๋‹ค์Œ์€ ๊ทธ๋ฆผ์œผ๋กœ ๋œ ๊ฐœ์š”์ž…๋‹ˆ๋‹ค:
![edit-consistency](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/edit-consistency.png)
์šฐ๋ฆฌ๋Š” ์ด ์ธก์ • ํ•ญ๋ชฉ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ๋ฏธ๋‹ˆ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค€๋น„ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋จผ์ € ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋กœ๋“œํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
```python
from datasets import load_dataset
dataset = load_dataset("sayakpaul/instructpix2pix-demo", split="train")
dataset.features
```
```bash
{'input': Value(dtype='string', id=None),
'edit': Value(dtype='string', id=None),
'output': Value(dtype='string', id=None),
'image': Image(decode=True, id=None)}
```
์—ฌ๊ธฐ์—๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ•ญ๋ชฉ์ด ์žˆ์Šต๋‹ˆ๋‹ค:
- `input`์€ `image`์— ํ•ด๋‹นํ•˜๋Š” ์บก์…˜์ž…๋‹ˆ๋‹ค.
- `edit`์€ ํŽธ์ง‘ ์ง€์‹œ์‚ฌํ•ญ์„ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
- `output`์€ `edit` ์ง€์‹œ์‚ฌํ•ญ์„ ๋ฐ˜์˜ํ•œ ์ˆ˜์ •๋œ ์บก์…˜์ž…๋‹ˆ๋‹ค.
์ƒ˜ํ”Œ์„ ์‚ดํŽด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
```python
idx = 0
print(f"Original caption: {dataset[idx]['input']}")
print(f"Edit instruction: {dataset[idx]['edit']}")
print(f"Modified caption: {dataset[idx]['output']}")
```
```bash
Original caption: 2. FAROE ISLANDS: An archipelago of 18 mountainous isles in the North Atlantic Ocean between Norway and Iceland, the Faroe Islands has 'everything you could hope for', according to Big 7 Travel. It boasts 'crystal clear waterfalls, rocky cliffs that seem to jut out of nowhere and velvety green hills'
Edit instruction: make the isles all white marble
Modified caption: 2. WHITE MARBLE ISLANDS: An archipelago of 18 mountainous white marble isles in the North Atlantic Ocean between Norway and Iceland, the White Marble Islands has 'everything you could hope for', according to Big 7 Travel. It boasts 'crystal clear waterfalls, rocky cliffs that seem to jut out of nowhere and velvety green hills'
```
๋‹ค์Œ์€ ์ด๋ฏธ์ง€์ž…๋‹ˆ๋‹ค:
```python
dataset[idx]["image"]
```
![edit-dataset](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/edit-dataset.png)
๋จผ์ € ํŽธ์ง‘ ์ง€์‹œ์‚ฌํ•ญ์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ์ด๋ฏธ์ง€๋ฅผ ํŽธ์ง‘ํ•˜๊ณ  ๋ฐฉํ–ฅ ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
[`StableDiffusionInstructPix2PixPipeline`]๋ฅผ ๋จผ์ € ๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค:
```python
from diffusers import StableDiffusionInstructPix2PixPipeline
instruct_pix2pix_pipeline = StableDiffusionInstructPix2PixPipeline.from_pretrained(
"timbrooks/instruct-pix2pix", torch_dtype=torch.float16
).to(device)
```
์ด์ œ ํŽธ์ง‘์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค:
```python
import numpy as np
def edit_image(input_image, instruction):
image = instruct_pix2pix_pipeline(
instruction,
image=input_image,
output_type="np",
generator=generator,
).images[0]
return image
input_images = []
original_captions = []
modified_captions = []
edited_images = []
for idx in range(len(dataset)):
input_image = dataset[idx]["image"]
edit_instruction = dataset[idx]["edit"]
edited_image = edit_image(input_image, edit_instruction)
input_images.append(np.array(input_image))
original_captions.append(dataset[idx]["input"])
modified_captions.append(dataset[idx]["output"])
edited_images.append(edited_image)
```
๋ฐฉํ–ฅ ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋จผ์ € CLIP์˜ ์ด๋ฏธ์ง€์™€ ํ…์ŠคํŠธ ์ธ์ฝ”๋”๋ฅผ ๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค:
```python
from transformers import (
CLIPTokenizer,
CLIPTextModelWithProjection,
CLIPVisionModelWithProjection,
CLIPImageProcessor,
)
clip_id = "openai/clip-vit-large-patch14"
tokenizer = CLIPTokenizer.from_pretrained(clip_id)
text_encoder = CLIPTextModelWithProjection.from_pretrained(clip_id).to(device)
image_processor = CLIPImageProcessor.from_pretrained(clip_id)
image_encoder = CLIPVisionModelWithProjection.from_pretrained(clip_id).to(device)
```
์ฃผ๋ชฉํ•  ์ ์€ ํŠน์ •ํ•œ CLIP ์ฒดํฌํฌ์ธํŠธ์ธ `openai/clip-vit-large-patch14`๋ฅผ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ด๋Š” Stable Diffusion ์‚ฌ์ „ ํ›ˆ๋ จ์ด ์ด CLIP ๋ณ€ํ˜•์ฒด์™€ ํ•จ๊ป˜ ์ˆ˜ํ–‰๋˜์—ˆ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ [๋ฌธ์„œ](https://huggingface.co/docs/transformers/model_doc/clip)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
๋‹ค์Œ์œผ๋กœ, ๋ฐฉํ–ฅ์„ฑ ์œ ์‚ฌ๋„๋ฅผ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด PyTorch์˜ `nn.Module`์„ ์ค€๋น„ํ•ฉ๋‹ˆ๋‹ค:
```python
import torch.nn as nn
import torch.nn.functional as F
class DirectionalSimilarity(nn.Module):
def __init__(self, tokenizer, text_encoder, image_processor, image_encoder):
super().__init__()
self.tokenizer = tokenizer
self.text_encoder = text_encoder
self.image_processor = image_processor
self.image_encoder = image_encoder
def preprocess_image(self, image):
image = self.image_processor(image, return_tensors="pt")["pixel_values"]
return {"pixel_values": image.to(device)}
def tokenize_text(self, text):
inputs = self.tokenizer(
text,
max_length=self.tokenizer.model_max_length,
padding="max_length",
truncation=True,
return_tensors="pt",
)
return {"input_ids": inputs.input_ids.to(device)}
def encode_image(self, image):
preprocessed_image = self.preprocess_image(image)
image_features = self.image_encoder(**preprocessed_image).image_embeds
image_features = image_features / image_features.norm(dim=1, keepdim=True)
return image_features
def encode_text(self, text):
tokenized_text = self.tokenize_text(text)
text_features = self.text_encoder(**tokenized_text).text_embeds
text_features = text_features / text_features.norm(dim=1, keepdim=True)
return text_features
def compute_directional_similarity(self, img_feat_one, img_feat_two, text_feat_one, text_feat_two):
sim_direction = F.cosine_similarity(img_feat_two - img_feat_one, text_feat_two - text_feat_one)
return sim_direction
def forward(self, image_one, image_two, caption_one, caption_two):
img_feat_one = self.encode_image(image_one)
img_feat_two = self.encode_image(image_two)
text_feat_one = self.encode_text(caption_one)
text_feat_two = self.encode_text(caption_two)
directional_similarity = self.compute_directional_similarity(
img_feat_one, img_feat_two, text_feat_one, text_feat_two
)
return directional_similarity
```
์ด์ œย `DirectionalSimilarity`๋ฅผ ์‚ฌ์šฉํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
```python
dir_similarity = DirectionalSimilarity(tokenizer, text_encoder, image_processor, image_encoder)
scores = []
for i in range(len(input_images)):
original_image = input_images[i]
original_caption = original_captions[i]
edited_image = edited_images[i]
modified_caption = modified_captions[i]
similarity_score = dir_similarity(original_image, edited_image, original_caption, modified_caption)
scores.append(float(similarity_score.detach().cpu()))
print(f"CLIP directional similarity: {np.mean(scores)}")
# CLIP directional similarity: 0.0797976553440094
```
CLIP ์ ์ˆ˜์™€ ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ, CLIP ๋ฐฉํ–ฅ ์œ ์‚ฌ์„ฑ์ด ๋†’์„์ˆ˜๋ก ์ข‹์Šต๋‹ˆ๋‹ค.
`StableDiffusionInstructPix2PixPipeline`์€ `image_guidance_scale`๊ณผ `guidance_scale`์ด๋ผ๋Š” ๋‘ ๊ฐ€์ง€ ์ธ์ž๋ฅผ ๋…ธ์ถœ์‹œํ‚ต๋‹ˆ๋‹ค. ์ด ๋‘ ์ธ์ž๋ฅผ ์กฐ์ •ํ•˜์—ฌ ์ตœ์ข… ํŽธ์ง‘๋œ ์ด๋ฏธ์ง€์˜ ํ’ˆ์งˆ์„ ์ œ์–ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋‘ ์ธ์ž์˜ ์˜ํ–ฅ์„ ์‹คํ—˜ํ•ด๋ณด๊ณ  ๋ฐฉํ–ฅ ์œ ์‚ฌ์„ฑ์— ๋ฏธ์น˜๋Š” ์˜ํ–ฅ์„ ํ™•์ธํ•ด๋ณด๊ธฐ๋ฅผ ๊ถŒ์žฅํ•ฉ๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์˜ ๊ฐœ๋…์„ ํ™•์žฅํ•˜์—ฌ ์›๋ณธ ์ด๋ฏธ์ง€์™€ ํŽธ์ง‘๋œ ๋ฒ„์ „์˜ ์œ ์‚ฌ์„ฑ์„ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด `F.cosine_similarity(img_feat_two, img_feat_one)`์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ข…๋ฅ˜์˜ ํŽธ์ง‘์—์„œ๋Š” ์ด๋ฏธ์ง€์˜ ์ฃผ์š” ์˜๋ฏธ๊ฐ€ ์ตœ๋Œ€ํ•œ ๋ณด์กด๋˜์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ฆ‰, ๋†’์€ ์œ ์‚ฌ์„ฑ ์ ์ˆ˜๋ฅผ ์–ป์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
[`StableDiffusionPix2PixZeroPipeline`](https://huggingface.co/docs/diffusers/main/en/api/pipelines/pix2pix_zero#diffusers.StableDiffusionPix2PixZeroPipeline)์™€ ๊ฐ™์€ ์œ ์‚ฌํ•œ ํŒŒ์ดํ”„๋ผ์ธ์—๋„ ์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
<Tip>
CLIP ์ ์ˆ˜์™€ CLIP ๋ฐฉํ–ฅ ์œ ์‚ฌ์„ฑ ๋ชจ๋‘ CLIP ๋ชจ๋ธ์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ‰๊ฐ€๊ฐ€ ํŽธํ–ฅ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค
</Tip>
***IS, FID (๋‚˜์ค‘์— ์„ค๋ช…ํ•  ์˜ˆ์ •), ๋˜๋Š” KID์™€ ๊ฐ™์€ ๋ฉ”ํŠธ๋ฆญ์„ ํ™•์žฅํ•˜๋Š” ๊ฒƒ์€ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค***. ํ‰๊ฐ€ ์ค‘์ธ ๋ชจ๋ธ์ด ๋Œ€๊ทœ๋ชจ ์ด๋ฏธ์ง€ ์บก์…”๋‹ ๋ฐ์ดํ„ฐ์…‹ (์˜ˆ: [LAION-5B ๋ฐ์ดํ„ฐ์…‹](https://laion.ai/blog/laion-5b/))์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ๋˜์—ˆ์„ ๋•Œ ์ด๋Š” ๋ฌธ์ œ๊ฐ€ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์™œ๋ƒํ•˜๋ฉด ์ด๋Ÿฌํ•œ ๋ฉ”ํŠธ๋ฆญ์˜ ๊ธฐ๋ฐ˜์—๋Š” ์ค‘๊ฐ„ ์ด๋ฏธ์ง€ ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•ด ImageNet-1k ๋ฐ์ดํ„ฐ์…‹์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ๋œ InceptionNet์ด ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. Stable Diffusion์˜ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹์€ InceptionNet์˜ ์‚ฌ์ „ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์…‹๊ณผ ๊ฒน์น˜๋Š” ๋ถ€๋ถ„์ด ์ œํ•œ์ ์ผ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ ๋”ฐ๋ผ์„œ ์—ฌ๊ธฐ์—๋Š” ์ข‹์€ ํ›„๋ณด๊ฐ€ ์•„๋‹™๋‹ˆ๋‹ค.
***์œ„์˜ ๋ฉ”ํŠธ๋ฆญ์„ ์‚ฌ์šฉํ•˜๋ฉด ํด๋ž˜์Šค ์กฐ๊ฑด์ด ์žˆ๋Š” ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [DiT](https://huggingface.co/docs/diffusers/main/en/api/pipelines/dit). ์ด๋Š” ImageNet-1k ํด๋ž˜์Šค์— ์กฐ๊ฑด์„ ๊ฑธ๊ณ  ์‚ฌ์ „ ํ›ˆ๋ จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.***
### ํด๋ž˜์Šค ์กฐ๊ฑดํ™” ์ด๋ฏธ์ง€ ์ƒ์„ฑ[[class-conditioned-image-generation]]
ํด๋ž˜์Šค ์กฐ๊ฑดํ™” ์ƒ์„ฑ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ [ImageNet-1k](https://huggingface.co/datasets/imagenet-1k)์™€ ๊ฐ™์€ ํด๋ž˜์Šค ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ๋ฐ์ดํ„ฐ์…‹์—์„œ ์‚ฌ์ „ ํ›ˆ๋ จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ชจ๋ธ์„ ํ‰๊ฐ€ํ•˜๋Š” ์ธ๊ธฐ์žˆ๋Š” ์ง€ํ‘œ์—๋Š” Frรฉchet Inception Distance (FID), Kernel Inception Distance (KID) ๋ฐ Inception Score (IS)๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ฌธ์„œ์—์„œ๋Š” FID ([Heusel et al.](https://arxiv.org/abs/1706.08500))์— ์ดˆ์ ์„ ๋งž์ถ”๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. [`DiTPipeline`](https://huggingface.co/docs/diffusers/api/pipelines/dit)์„ ์‚ฌ์šฉํ•˜์—ฌ FID๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์ด๋Š” ๋‚ด๋ถ€์ ์œผ๋กœ [DiT ๋ชจ๋ธ](https://arxiv.org/abs/2212.09748)์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
FID๋Š” ๋‘ ๊ฐœ์˜ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹์ด ์–ผ๋งˆ๋‚˜ ์œ ์‚ฌํ•œ์ง€๋ฅผ ์ธก์ •ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•ฉ๋‹ˆ๋‹ค. [์ด ์ž๋ฃŒ](https://mmgeneration.readthedocs.io/en/latest/quick_run.html#fid)์— ๋”ฐ๋ฅด๋ฉด:
> Frรฉchet Inception Distance๋Š” ๋‘ ๊ฐœ์˜ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ์ธก์ •ํ•˜๋Š” ์ง€ํ‘œ์ž…๋‹ˆ๋‹ค. ์‹œ๊ฐ์  ํ’ˆ์งˆ์— ๋Œ€ํ•œ ์ธ๊ฐ„ ํŒ๋‹จ๊ณผ ์ž˜ ์ƒ๊ด€๋˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ์œผ๋ฉฐ, ์ฃผ๋กœ ์ƒ์„ฑ์  ์ ๋Œ€ ์‹ ๊ฒฝ๋ง์˜ ์ƒ˜ํ”Œ ํ’ˆ์งˆ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. FID๋Š” Inception ๋„คํŠธ์›Œํฌ์˜ ํŠน์ง• ํ‘œํ˜„์— ๋งž๊ฒŒ ์ ํ•ฉํ•œ ๋‘ ๊ฐœ์˜ ๊ฐ€์šฐ์‹œ์•ˆ ์‚ฌ์ด์˜ Frรฉchet ๊ฑฐ๋ฆฌ๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ๊ตฌํ•ฉ๋‹ˆ๋‹ค.
์ด ๋‘ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์€ ์‹ค์ œ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹๊ณผ ๊ฐ€์งœ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹(์šฐ๋ฆฌ์˜ ๊ฒฝ์šฐ ์ƒ์„ฑ๋œ ์ด๋ฏธ์ง€)์ž…๋‹ˆ๋‹ค. FID๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋‘ ๊ฐœ์˜ ํฐ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ๊ณ„์‚ฐ๋ฉ๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด ๋ฌธ์„œ์—์„œ๋Š” ๋‘ ๊ฐœ์˜ ๋ฏธ๋‹ˆ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์ž‘์—…ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
๋จผ์ € ImageNet-1k ํ›ˆ๋ จ ์„ธํŠธ์—์„œ ๋ช‡ ๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ ๋‹ค์šด๋กœ๋“œํ•ด ๋ด…์‹œ๋‹ค:
```python
from zipfile import ZipFile
import requests
def download(url, local_filepath):
r = requests.get(url)
with open(local_filepath, "wb") as f:
f.write(r.content)
return local_filepath
dummy_dataset_url = "https://hf.co/datasets/sayakpaul/sample-datasets/resolve/main/sample-imagenet-images.zip"
local_filepath = download(dummy_dataset_url, dummy_dataset_url.split("/")[-1])
with ZipFile(local_filepath, "r") as zipper:
zipper.extractall(".")
```
```python
from PIL import Image
import os
dataset_path = "sample-imagenet-images"
image_paths = sorted([os.path.join(dataset_path, x) for x in os.listdir(dataset_path)])
real_images = [np.array(Image.open(path).convert("RGB")) for path in image_paths]
```
๋‹ค์Œ์€ ImageNet-1k classes์˜ ์ด๋ฏธ์ง€ 10๊ฐœ์ž…๋‹ˆ๋‹ค : "cassette_player", "chain_saw" (x2), "church", "gas_pump" (x3), "parachute" (x2), ๊ทธ๋ฆฌ๊ณ  "tench".
<p align="center">
<img src="https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/real-images.png" alt="real-images"><br>
<em>Real images.</em>
</p>
์ด์ œ ์ด๋ฏธ์ง€๊ฐ€ ๋กœ๋“œ๋˜์—ˆ์œผ๋ฏ€๋กœ ์ด๋ฏธ์ง€์— ๊ฐ€๋ฒผ์šด ์ „์ฒ˜๋ฆฌ๋ฅผ ์ ์šฉํ•˜์—ฌ FID ๊ณ„์‚ฐ์— ์‚ฌ์šฉํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
```python
from torchvision.transforms import functional as F
def preprocess_image(image):
image = torch.tensor(image).unsqueeze(0)
image = image.permute(0, 3, 1, 2) / 255.0
return F.center_crop(image, (256, 256))
real_images = torch.cat([preprocess_image(image) for image in real_images])
print(real_images.shape)
# torch.Size([10, 3, 256, 256])
```
์ด์ œ ์œ„์—์„œ ์–ธ๊ธ‰ํ•œ ํด๋ž˜์Šค์— ๋”ฐ๋ผ ์กฐ๊ฑดํ™” ๋œ ์ด๋ฏธ์ง€๋ฅผ ์ƒ์„ฑํ•˜๊ธฐ ์œ„ํ•ด [`DiTPipeline`](https://huggingface.co/docs/diffusers/api/pipelines/dit)๋ฅผ ๋กœ๋“œํ•ฉ๋‹ˆ๋‹ค.
```python
from diffusers import DiTPipeline, DPMSolverMultistepScheduler
dit_pipeline = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", torch_dtype=torch.float16)
dit_pipeline.scheduler = DPMSolverMultistepScheduler.from_config(dit_pipeline.scheduler.config)
dit_pipeline = dit_pipeline.to("cuda")
words = [
"cassette player",
"chainsaw",
"chainsaw",
"church",
"gas pump",
"gas pump",
"gas pump",
"parachute",
"parachute",
"tench",
]
class_ids = dit_pipeline.get_label_ids(words)
output = dit_pipeline(class_labels=class_ids, generator=generator, output_type="np")
fake_images = output.images
fake_images = torch.tensor(fake_images)
fake_images = fake_images.permute(0, 3, 1, 2)
print(fake_images.shape)
# torch.Size([10, 3, 256, 256])
```
์ด์ œ [`torchmetrics`](https://torchmetrics.readthedocs.io/)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ FID๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
```python
from torchmetrics.image.fid import FrechetInceptionDistance
fid = FrechetInceptionDistance(normalize=True)
fid.update(real_images, real=True)
fid.update(fake_images, real=False)
print(f"FID: {float(fid.compute())}")
# FID: 177.7147216796875
```
FID๋Š” ๋‚ฎ์„์ˆ˜๋ก ์ข‹์Šต๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ ๊ฐ€์ง€ ์š”์†Œ๊ฐ€ FID์— ์˜ํ–ฅ์„ ์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
- ์ด๋ฏธ์ง€์˜ ์ˆ˜ (์‹ค์ œ ์ด๋ฏธ์ง€์™€ ๊ฐ€์งœ ์ด๋ฏธ์ง€ ๋ชจ๋‘)
- diffusion ๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๋ฌด์ž‘์œ„์„ฑ
- diffusion ๊ณผ์ •์—์„œ์˜ ์ถ”๋ก  ๋‹จ๊ณ„ ์ˆ˜
- diffusion ๊ณผ์ •์—์„œ ์‚ฌ์šฉ๋˜๋Š” ์Šค์ผ€์ค„๋Ÿฌ
๋งˆ์ง€๋ง‰ ๋‘ ๊ฐ€์ง€ ์š”์†Œ์— ๋Œ€ํ•ด์„œ๋Š”, ๋‹ค๋ฅธ ์‹œ๋“œ์™€ ์ถ”๋ก  ๋‹จ๊ณ„์—์„œ ํ‰๊ฐ€๋ฅผ ์‹คํ–‰ํ•˜๊ณ  ํ‰๊ท  ๊ฒฐ๊ณผ๋ฅผ ๋ณด๊ณ ํ•˜๋Š” ๊ฒƒ์€ ์ข‹์€ ์‹ค์ฒœ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค
<Tip warning={true}>
FID ๊ฒฐ๊ณผ๋Š” ๋งŽ์€ ์š”์†Œ์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ทจ์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
* ๊ณ„์‚ฐ ์ค‘ ์‚ฌ์šฉ๋˜๋Š” ํŠน์ • Inception ๋ชจ๋ธ.
* ๊ณ„์‚ฐ์˜ ๊ตฌํ˜„ ์ •ํ™•๋„.
* ์ด๋ฏธ์ง€ ํ˜•์‹ (PNG ๋˜๋Š” JPG์—์„œ ์‹œ์ž‘ํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋‹ค๋ฆ…๋‹ˆ๋‹ค).
์ด๋Ÿฌํ•œ ์‚ฌํ•ญ์„ ์—ผ๋‘์— ๋‘๋ฉด, FID๋Š” ์œ ์‚ฌํ•œ ์‹คํ–‰์„ ๋น„๊ตํ•  ๋•Œ ๊ฐ€์žฅ ์œ ์šฉํ•˜์ง€๋งŒ, ์ €์ž๊ฐ€ FID ์ธก์ • ์ฝ”๋“œ๋ฅผ ์ฃผ์˜ ๊นŠ๊ฒŒ ๊ณต๊ฐœํ•˜์ง€ ์•Š๋Š” ํ•œ ๋…ผ๋ฌธ ๊ฒฐ๊ณผ๋ฅผ ์žฌํ˜„ํ•˜๊ธฐ๋Š” ์–ด๋ ต์Šต๋‹ˆ๋‹ค.
์ด๋Ÿฌํ•œ ์‚ฌํ•ญ์€ KID ๋ฐ IS์™€ ๊ฐ™์€ ๋‹ค๋ฅธ ๊ด€๋ จ ๋ฉ”ํŠธ๋ฆญ์—๋„ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค.
</Tip>
๋งˆ์ง€๋ง‰ ๋‹จ๊ณ„๋กœ, `fake_images`๋ฅผ ์‹œ๊ฐ์ ์œผ๋กœ ๊ฒ€์‚ฌํ•ด ๋ด…์‹œ๋‹ค.
<p align="center">
<img src="https://huggingface.co/datasets/diffusers/docs-images/resolve/main/evaluation_diffusion_models/fake-images.png" alt="fake-images"><br>
<em>Fake images.</em>
</p>