What it does

This is a desktop browser app for the wai-il image generation model. You pick a character from a curated list of 5,150 entries, and the app builds the prompt, handles LoRA slots, ControlNet, IP Adapter, regional conditioning, and queues the job to either ComfyUI or A1111/Forge Neo via API. It also runs an ONNX image tagger locally (slowly, on CPU) and can upscale finished images through a companion ComfyUI node called MiraITU.

The interesting bit

The app is essentially a structured prompt generator wearing a full UI tuxedo. The real work—sampling, inference, model loading—still happens in your existing backend. What it adds is orchestration: a character database with thumbnails, semi-automatic tag completion, spell-check, and enough toggles to handle Flux/SDXL diffusion models, GGUF checkpoints, and tiled upscaling without writing JSON by hand. The MiraITU upscaler is particularly specific: it tiles images to keep VRAM usage sane on 8GB cards, with separate code paths for “reference latent” models like Flux.2 versus tagger-dependent models like SDXL.

Key highlights

• Ships with 5,150-character database for waiIllustriousSDXL_v160; swapable to v120 via HuggingFace
• One-click packages available, plus a browser-based variant (SAAC) and Python CLI agent
• Supports ComfyUI, A1111, and Forge Neo with feature parity matrix documented (e.g., ComfyUI gets image color transfer; A1111 gets BREAK syntax)
• Built-in ONNX tagger (WD, CL, Camie models) running in Node.js—no Python backend, but ~3× slower than Python CPU, ~12× slower than GPU
• MiraITU tiled upscaling: 1.5×–8×, with tile-size/VRAM tables provided for 8GB and 24GB cards
• Drag-and-drop image handling for ControlNet preprocessing and upscaling

Caveats

• The tagger is explicitly CPU-bound in Electron; GPU acceleration “seems not working” per the README
• ComfyUI users must install a companion custom node (ComfyUI_Mira) for full functionality
• Database updates require manual file deletion to trigger re-download; git pull alone won’t refresh character lists
• Forge Neo and A1111 have divergent ControlNet API quirks (“none” vs “None”) that the app can’t auto-detect

Verdict

Worth a look if you’re deep in the wai-il ecosystem and want a visual front-end for prompt assembly and batch queuing. Skip it if you already have a ComfyUI workflow you love, or if your setup is A1111-only and you don’t need the character database—this is glue code, admittedly very thorough glue code, not a model or sampler itself.