๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

Generative AI6

[Paper Review] Prompt-to-Prompt Image Editing with Cross Attention Control https://arxiv.org/abs/2208.01626 Prompt-to-Prompt Image Editing with Cross Attention ControlRecent large-scale text-driven synthesis models have attracted much attention thanks to their remarkable capabilities of generating highly diverse images that follow given text prompts. Such text-based synthesis methods are particularly appealing to humansarxiv.org   ๊ธฐ์กด LLI (Large-scale language-image) mo.. 2025. 2. 13.
[Paper Review] Classifier-Free Diffusion Guidance https://arxiv.org/abs/2207.12598 Classifier-Free Diffusion GuidanceClassifier guidance is a recently introduced method to trade off mode coverage and sample fidelity in conditional diffusion models post training, in the same spirit as low temperature sampling or truncation in other types of generative models. Classifier garxiv.org  Introduce ์ด ๋…ผ๋ฌธ์€ classifier guidance ๋…ผ๋ฌธ์—์„œ classifier ์„ ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ ๋„ c.. 2025. 2. 5.
[Notable] Low Temperature Samples Low Temperature Samples๋Š” ์ƒ์„ฑ ๋ชจ๋ธ์—์„œ ์ƒ˜ํ”Œ ํ’ˆ์งˆ์„ ๋†’์ด๊ณ  ๋‹ค์–‘์„ฑ์„ ์ค„์ด๋Š” ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค.์ฃผ๋กœ ํ™•๋ฅ  ๋ถ„ํฌ์˜ "์ƒ˜ํ”Œ๋ง ์˜จ๋„(temperature)"๋ฅผ ์กฐ์ ˆํ•˜์—ฌ ์ƒ์„ฑ ๊ฒฐ๊ณผ์— ์˜ํ–ฅ์„ ์ค๋‹ˆ๋‹ค. 1. "Temperature"์˜ ์˜๋ฏธTemperature๋Š” ํ™•๋ฅ  ๋ถ„ํฌ์˜ "๋‚ ์นด๋กœ์›€(sharpness)" ๋˜๋Š” "๋ถˆํ™•์‹ค์„ฑ(uncertainty)"์„ ์กฐ์ ˆํ•˜๋Š” ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ์ž…๋‹ˆ๋‹ค.์ˆ˜ํ•™์ ์œผ๋กœ๋Š” ์†Œํ”„ํŠธ๋งฅ์Šค(softmax) ํ•จ์ˆ˜์—์„œ ์ž์ฃผ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค:์—ฌ๊ธฐ์„œ:T = temperature (์˜จ๋„)zi = ๋กœ์ง“(logit) ๊ฐ’ (๋ชจ๋ธ์ด ์˜ˆ์ธกํ•œ ์ ์ˆ˜)P(xi) = ์ตœ์ข… ํ™•๋ฅ  2. Temperature์˜ ์˜ํ–ฅ๋†’์€ ์˜จ๋„ (Tโ‰ซ1)ํ™•๋ฅ  ๋ถ„ํฌ๊ฐ€ **ํ‰ํ‰(flat)**ํ•ด์ง€๊ณ , ๋” ๋‹ค์–‘ํ•œ ์ƒ˜ํ”Œ์ด ์ƒ์„ฑ๋จ๋ชจ๋ธ์ด ๋ถˆํ™•์‹คํ•œ ์„ ํƒ์„ ๋” ๋งŽ์ด .. 2025. 2. 5.
[Notable] Evaluation Metrics ํ•ด๋‹น ๊ธ€์€ chatGPT ๋กœ ์ž‘์„ฑ๋œ ๊ธ€ ์ž…๋‹ˆ๋‹ค. 2025. 2. 4.
[Paper Review] High-Resolution Image Synthesis with Latent Diffusion Models (Aka. Stable Diffusion) https://arxiv.org/abs/2112.10752 High-Resolution Image Synthesis with Latent Diffusion ModelsBy decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond. Additionally, their formulation allows for a guiding mechanism tarxiv.org ์ด๋ฒˆ ์ฃผ์ œ๋Š” ์•„์ฃผ ์œ ๋ช…ํ•œ Stable Diffuion ๋…ผ๋ฌธ์„ ๋ฆฌ๋ทฐํ•ด๋ณด๋„.. 2025. 2. 4.
Diffusion(DDPM) diffusion ์ด๋ž€? diffusion process๋ฅผ ์ด์šฉํ•œ Generative ๋ชจ๋ธ์ด๋‹ค. Denoising diffusion ๋ชจ๋ธ์—๋Š” ๋‘๊ฐ€์ง€ ๊ณผ์ •์ด ์žˆ๋‹ค. Forward Diffusion process์—์„œ๋Š” ์ดˆ๊ธฐ ์กฐ๊ฑด๊ณผ ๋ณ€๋™์„ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฏธ๋ž˜ ๊ฐ’์„ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ˜๋ฉด, Reverse Diffusion process์—์„œ๋Š” ์ข…๋‹จ ๊ฐ’์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์ดˆ๊ธฐ ๊ฐ’์ด๋‚˜ ๊ฒฝ๋กœ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค. Forward Diffusion process Generative ๋ชจ๋ธ์—์„œ ์‚ฌ์šฉ๋˜๋Š” ํ™•๋ฅ ๋ก ์  ๋ชจ๋ธ๋ง ๊ธฐ๋ฒ•์ด๋‹ค. ์ดˆ๊ธฐ ์กฐ๊ฑด๊ณผ ๋ณ€๋™์„ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ํ™•๋ฅ  ๋ณ€์ˆ˜์˜ ๊ฐ’์„ ์—…๋ฐ์ดํŠธํ•˜์—ฌ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ๋‹ค. Forward Diffusion process๋Š” ๋ธŒ๋ผ์šด ์šด๋™(Brownian motion)์„ ๊ธฐ๋ฐ˜์œผ.. 2023. 7. 11.