PotionUI 0.0.11: open-source studio adds per-model prompt syntax, @-references to your own images
0roborus_ · reddit · 2026-09-27
Self-hosted open-source generation studio PotionUI ships 0.0.11 with a major prompt editor overhaul:
- Per-model syntax highlighting: MiniMax-H3's subjectdefinitions:, Qwen-Image's quoted render text, and YuE2's [Verse]/[Chorus] are colored as you type; typing / lists everything the model understands with one-line explanations.
- @-references: in presets accepting reference images/video/audio, typing @ picks an upload which becomes a chip with thumbnail; the model receives it as a stable <Picture 2> regardless of upload order.
- New pickers for phrases, variables and references open at the cursor with thumbnails; chips support fixed vs auto-shuffle values and exclusion.
- Performance & library: long prompts no longer re-render the whole form per keystroke; library search takes wildcards and regex, admins can filter by index date/usage and grant access in bulk, and model downloads suggest GPU-appropriate quant (bf16/fp8/int8/nvfp4).
- Chat can @ any model or LoRA (auto-injecting trigger words and strength), generation failures now show plain messages with error IDs, and the admin UI is unified.
GitHub: PotionUI/PotionUI.
More from Multimodal
- Flovaai turns a random lunch break idea into a cinematic AI video — alifcoder · 2026-09-27
- MiniMax H3 ecosystem daily: 1920s horror LoRA, 8Mpx still-node tests, and new training adapter — optimisticalish · 2026-09-27
- Seedance US 2.5 Now Supports Native 1080p Generation on fal — gorkem · 2026-09-27
- Krea 2 Trick Keeps Character Identity Consistent Across Poses — Epictetito · 2026-09-27
- How to Pose Two Characters with Identity Preservation in Krea 2? Klein Does It with One LoRA — Epictetito · 2026-09-27
- Is 12GB of VRAM on an RTX 5070 Enough to Run Local Image Models Like Qwen-Image 2.1? — FilipeDM · 2026-09-27