Diffusers Tech Preview

The Diffusers Tech Preview allows early access to the new features based on Diffusers.

This is under active development, and is missing a few features. It is experimental! Please report any bugs to the #beta channel in our Discord server!

New upcoming features in our new engine

• LORA support - Place LORA files in the models/lora folder.
• Compel Prompt Parser - New, more powerful parser. In short:
  • no limit to the length of prompts (i.e. long prompts are supported)
  • Use + and - to increase/decrease the weight. E.g. apple, apple+, apple++, apple+++, or apple-, apple-- for different weights.
  • Use exact weights - 0.0 to 1.0 reduces the weight, 1.0 to 2.0 increases the weight. Think of it like a multiplier, like 1.5x or 0.5x: E.g. (apple)0.8 falling from a tree, (apple)1.5 falling from a tree, (apple falling)1.4 from a tree
  • You can group tokens together using parentheses/round-brackets. E.g. (apple falling)++ from a tree. Nested parentheses are supported.
  This clarifies a few things:
  • colon (:) is NOT used for blending. Neither is it used for weights. It has no impact and will be considered a part of the prompt.
  • (()) and [] do not affect the prompt's weights.
• More choices for img2img samplers
• Support for official inpainting models
• Generate images that tile seamlessly
• Clip Skip support allows to skip the last CLIP layer (recommended by some LORA models)
• New samplers: DDPM and DEIS
• Memory optimizations that allow the use of 2GB GPUs

Known issues

• Some LoRA consistently fail to load in EasyDiffusion
• Some LoRA are far more sensitive to alpha (compared to a11)
• Hangs sometimes on "compel is ready", while making the token.
• Some custom inpainting models don't work
• These samplers don't work yet: Unipc SNR, Unipc TQ, Unipc SNR2, DPM++ 2s Ancestral, DPM++ SDE, DPM Fast, DPM Adaptive, DPM2
• Hypernetwork doesn't work
• The time remaining in browser differs from the one in the console

Download all images

Save Settings Configuration

Select which settings should be remembered when restarting the browser