v0.3.0

Lora's support.
Options --lora-dir and --lora.
Prompt parsing of <lora:NAME:MULT>.

Inpainting (img2img modifying only a region).
Mask from option --in-mask or from input image alpha channel.

Improves VAE tiling: reduced artifacts in the borders between tiles.

Supports loading bf16.

v0.2.5

Additional sampling methods: euler_a, dpm++2m, dpm++2s, dpm++2s_a
New option --s-ancestral enables ancestral sampling for any method.
VAE tiling (--vae-tile INT) for decoding and encoding.
FIX: safetensors loading will ignore too-large metadata entries instead of failing.

v0.2.2

Fixed to compile with the latest ggml from llama.cpp.
Fixed a bug in the tokenizer (it aborted with "illustration").
Fixed non-square images (width != height).

2024-07-13

First binary release for Windows 64 bits supporting Vulkan and CPU inference.