Skip to content

v0.45.0

Choose a tag to compare

View all tags
@vulogov vulogov released this 07 Jun 11:28
· 44 commits to main since this release
v0.45.0
4b5554b

plakat style train — learn a style from a folder of images into a LoRA you drop onto any generation. Creation, not just detection.

plakat style train --from-dir ./imgs --base sd35 --trigger "wcstyle watercolour…" --out style.safetensors
plakat generate "a harbour, wcstyle watercolour…" --model sd35-medium --lora style.safetensors
  • Phase 1: SD 3.5. Mixed-precision (BF16 base + F32 LoRA on attention), trains on a 24 GB Mac at 256² with periodic checkpoints. Output is a standard diffusers-PEFT .safetensors. Training and generation are separated (train once, render forever).
  • Verified end-to-end: 9 watercolour references → LoRA → fresh subjects render in the trained style (191/191 targets merged).
  • Docs: tutorial + reference.

Also fixed

  • SD 3 --lora on diffusers checkpoints (sd35-medium) was silently 0/N merged — the merge now remaps diffusers→SAI first (191/191). Fixes any SD3 LoRA.
  • SD 2.1 404 (gated repo) → repointed to an ungated v-prediction mirror.

SDXL + SD 1.5 training bases land in 0.46.0.

Full Changelog: v0.44.0...v0.45.0

Assets 8
Loading