Skip to content

v0.2.3

Choose a tag to compare

View all tags
@github-actions github-actions released this 14 Jul 14:49
v0.2.3
dd5f7a8
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Big release: a redesigned model picker, support for every engine image backend, local single-file checkpoints, automatic engine version management, and the low-VRAM SDXL performance fix.

⚡ SDXL performance on low-VRAM GPUs

Generations that could take 10+ minutes on 8 GB cards (SDXL 1024×1024) now run at normal speed.

  • Auto CPU-offload: on startup the app probes GPU VRAM (nvidia-smi) and automatically enables CPU-offload on low-memory cards (≤ 12 GiB), keeping peak VRAM from spilling into slow shared memory. Exposed as a tri-state Auto / On / Off control in Settings — Auto is the default and decides per GPU.
  • Ships pinned to imference-engine v0.3.1, which also fixes the root-cause VRAM leak in weighted prompt encoding.

🎴 New model picker

The model dropdown is replaced by a card-grid picker dialog:

  • Models grouped by type, with search/filter.
  • Works for both cloud and local models.
  • A dedicated "My models" tab for manually-loaded custom checkpoints.

🧩 All engine image backends + custom local checkpoints

  • Every backend the engine ships is now selectable: SDXL, SD 1.5, Z-Image, FLUX, Chroma, Qwen-Image, and Anima.
  • Load your own .safetensors: pick a local checkpoint, choose which backend should load it, and (for transformer-only backends) optionally override the base repo. All seven backends are available in the custom-model flow — including Anima single-file DiT checkpoints (Civitai / ComfyUI form).
  • Runtime settings unified into a single Image block that applies across every image backend (instead of per-backend duplication).

🔄 Engine version management

  • The app version is now shown under the logo.
  • On startup the app checks the installed engine against the pinned version and auto-syncs the venv (force reinstall) when it's outdated — no manual pip step.
  • The installed engine version is shown in Settings under the engine section.
  • Cloud-only setups are respected: if no engine is installed (e.g. no GPU), the app does not auto-install one.

🛠️ Fixes & internals

  • Install torchvision alongside torch — required by the Anima / Cosmos forward pass.
  • Fix the sidecar for the engine 0.3.x RuntimeConfig field rename (enable_cpu_offload → enable_offload).
  • Pin sd_embed to a fixed commit for reproducible installs.

Upgrade

Download the v0.2.3 build. On first launch it will sync the bundled engine to the pinned version automatically (if an engine is installed).

Assets 7
Loading