TEKKET: to, na co odkazuju níže, jsou webové "playgroundy", tam nepotřebuješ žádný výkonný hardware, protože nic nepočítáš na své straně, klidně to můžeš používat na mobilu. Pokud chceš zprovoznit nějaký stable diffusion gui u sebe doma, tak záleží hlavně na velikosti RAM, co máš na grafické kartě. Zkus třeba
https://nmkd.itch.io/t2i-gui - ale je toho hromada. Vozka tu doporučoval možná trochu lepší GUI, ale zatím jsem se nepřinutil to zkusit ;)
NMKD
text2image-gui/README.md at main · n00mkrad/text2image-gui · GitHubhttps://github.com/n00mkrad/text2image-gui/blob/main/README.md#system-requirementsSystem Requirements
OS: Windows 10/11 64-bit
Minimum:
GPU: Nvidia GPU with 4 GB VRAM, Maxwell Architecture (2014) or newer
Alternatively, with limited feature support: Any DirectML-capable GPU with 8 GB of VRAM
RAM: 8 GB RAM (Note: Pagefile must be enabled as swapping will occur with only 8 GB!)
Disk: 10 GB (another free 5 GB for temporary files recommended)
Recommended:
GPU: Nvidia GPU with 8 GB VRAM, Pascal Architecture (2016) or newer
RAM: 16 GB RAM
Disk: 12 GB on SSD (another free 5 GB for temporary files recommended)
Professional/DreamBooth-capable:
GPU: Nvidia GPU with 24GB VRAM, Turing Architecture (2018) or newer
RAM: 32 GB RAM
Disk: 12 GB on NVME SSD (another free 25 GB for temporary files recommended), system-managed paging file enabled
Features and How to Use Them
Prompt Input
Multiple prompts at once: Enter each prompt on a new line (newline-separated). Word wrapping does not count towards this.
Negative Prompt: Put words or phrases into this box to tell the AI to exclude those things when generating images.
Alternatively, you can also put the negative prompt into the regular prompt box by wrapping it in [brackets].
Emphasis: Use + after a word/phrase to make it more impactful, or - to do the opposite. You can also use to increase the effect. Wrap your phrase in parentheses if you want to apply it to more than one word.
Each plus/minus applies a multiplier of 1.1. So two +++ would be 1.1^3 = 1.331, and so on.
You can also type the strength manually after parentheses, e.g. a (huge)1.33 dog instead of a huge+++ dog
Syntax Examples: a green++ tree, a (big green)+ tree with orange- leaves (in the woods)++
Wildcards: Fill in words or phrases from a list into the prompt.
Inline: photo of a ~car,tree,dog~.
From File: photo of a ~objects for loading texts from objects.txt in your Wildcards folder in the SD GUI root folder.
Order: Use ~ for random/shuffled, ~~ for unchanged order, or ~~~ for sorted (A-Z) mode.
Additional Inputs
Load Image(s): Load an initialization image that will be used together with your text prompt ("img2img")
Loading multiple images means that each image will be processed separately.
Load Concept: Load a Textual Inversion concept to apply a style or use a specific character
Stable Diffusion Settings
Steps: More steps can increase detail, but only to a certain extent. Depending on the sampler, 20-60 is a good range.
Has a linear performance impact: Doubling the step count means each image takes twice as long to generate.
Prompt Guidance (CFG Scale): Lower values are closer to the raw output of the AI, higher values try to respect your prompt more accurately.
Use low values if you are happy with the AI's representation of your prompt. Use higher values if not - but going too high will degrade quality.
No performance impact, no matter the value.
Seed: Starting value for the image generation. Allows you to create the exact same image again by using the same seed.
When using the same seed, the image will only be identical if you also use the same sampler and resolution (and other settings).
Lock Seed Option: Disable incrementing the seed by 1 for each image. Only useful in combination with wildcards.
Resolution: Adjust image size. Only values that are divisible by 64 are possible. Sizes above 512x512 can lead to repeated patterns.
Higher resolution images require more VRAM and are slower to generate.
High-Resolution Fix: Enable this to avoid getting repeated patterns at high resolutions (~768px+). Can reduce fidelity though.
Sampler: Changes the way images are sampled. Euler Ancestral is the default because it's fast and tends to look good even with few steps.
Generate Seamless Images: Generates seamless/tileable images, very useful for making game textures or repeating backgrounds.
Image Viewer
Review current images: Use the scroll wheel while hovering over the image to go to the previous/next image.
Slideshow: The image viewer always shows the newest generated image if you haven't manually changed it in the last 3 seconds.
Context Menu: Right-click into the image area to show more options.
Pop-Up Viewer: Click into the image area to open the current image in a floating window.
Use the mouse wheel to change the window's size (zoom), right-click for more options, double-click to toggle fullscreen.