Spaces:

adamelliotfields

/

diffusion

like 9

Running on Zero

After dogfooding this for a while:

• I love the speed (loading and inference) of SD 1.5 still; combined with good prompts, TI embeddings, and super resolution it's not trash
• IP-Adapter and ControlNet isn't great on SD 1.5, more proof of concept, good for experimentation (I never use it)
• Image-to-image is great for style transfer
• Fast negative TI embeddings are convenient
• I use the same styles and settings with the same models; should just create simple presets
• I either use seed 42 or random
• I don't use LoRAs because I'd rather just pay 4 cents to use a commercial model that looks better
• I always use PNG format
• I always use Karras sigmas
• I never use custom dimensions
• I never use clip skip, freeU, or taesd; DeepCache is still useful for localhost
• I don't like the random prompts from Stable Studio; I just use "cat" with a style
• I'd like to have SD 1.5 and XL models in one app

To make a simple image generation app:

• Remove settings, image-to-image, and controlnet tabs; remove LoRA dropdowns and all the other unused settings
• Remove negative prompt; styles will have negatives and fast negative embedding
• Model selection replaced with "Quality" selection: SD 1.5 (good) or XL (best)
• Always use optimal inference steps; DeepCache can be used for speed
• Replace styles with presets (combination of styles and parameters)
• Single "reproduceable" checkbox to use seed 42; otherwise random
• Provide a hint for constructing a prompt rather than generate random ones