Image generation engine upgraded
We've upgraded our image generation engine for better quality and more faithful prompts.
Prompt weights work like in traditional image generation tools
(word:1.3)to emphasize,[word]to soften, stacked((…)), andBREAKare all supported- Long prompts (over 75 tokens) are applied all the way to the end
- Prompts you find elsewhere now come much closer to their original examples
Better inpainting
- Areas you did not paint stay exactly as they were
- New Only masked option under Inpaint area: it redraws only around the painted area at full resolution. Best for small fixes like eyes or hands
Truer colors, faster finishing
- Each model's VAE is now applied correctly, so colors come out as intended
- The final image processing is faster, so you wait less
New sampler
- DPM++ 2M SDE Karras is now available
Coming soon
- Hires fix: re-draws your image at a larger size with finer detail
- Anime upscaler: enlarges anime and illustration images while keeping lines crisp
Please note
<lora:…>tags inside a prompt are ignored. Pick LoRAs with the LoRA button- Because of the engine changes, the same seed and settings may give a slightly different image than before
