2 Stable Diffusion myths I tested on an 8 GB GPU

0 Просмотры

  • Спасибо! Поделитесь с друзьями!

    Вам не понравилось видео. Спасибо за то что поделились своим мнением!

aii
Издатель
Myth 1: smaller images are much faster.
Same prompt, same seed, 25 steps, low-VRAM mode: 512 px 45 s · 768 px 52 s · 1024 px 65 s.
❌ 4× fewer pixels, only 1.45× faster, and blurrier (SDXL is trained at 1024).

Myth 2: a negative prompt fixes everything.
Same seed, without vs with "blurry, deformed hands, extra fingers, low quality, cartoon, oversaturated":
⚠️ only 10 % of the pixels really changed on one image, 30 % on the other. Both versions looked fine: nothing was broken.

My take: stay at 1024 px, and use a negative prompt only for a problem you can actually see.
Times measured while a 3D render was also running on the PC: treat them as orders of magnitude.

Tools: Stable Diffusion XL base 1.0 by Stability AI (CreativeML Open RAIL++-M: commercial use allowed, with use-based restrictions), run locally with Hugging Face diffusers (Apache-2.0).
Narration voice: PocketTTS by Kyutai (CC-BY 4.0). Captions: Whisper large-v3-turbo via faster-whisper (MIT).

#AI #StableDiffusion #SDXL #AIArt #OpenSource
Категория
AI Картинки
Выполните вход или зарегистрируйтесь для отправки комментария.
Комментариев нет.