Skip to content

FEATURES.md

Latest commit

 

History

History
147 lines (118 loc) · 8 KB

FEATURES.md

File metadata and controls

147 lines (118 loc) · 8 KB
Raw

AIWF Studio Feature List

This file is the repo-facing feature inventory for GitHub readers. It separates working paths from planned or gated paths so the README can stay honest as the project grows.

Application Surfaces

  • Studio UI / Gradio Lab: original Gradio workspace launched by AIWF Studio Gradio Lab.bat or python launch_gradio.py.
  • Modern UI: cleaner Gradio shell launched by python webui_modern.py.
  • Pro UI: FastAPI plus React/TypeScript/Vite shell launched by AIWF Studio Pro.bat or python launch_pro.py.
  • Shared backend: all UI shells use the same local model folders, outputs, settings, and services.
  • Pro API: /api/pro/runtime, /api/pro/bootstrap, /api/pro/generate, /api/pro/data, /api/pro/logs, and /api/pro/settings.
  • Compatibility API: native /api/v1 plus A1111-style /sdapi/v1 adapter.

Pro UI Rebuild

  • Left-rail app navigation for Create, Models, Data, Monitor, Logs, and Settings.
  • Menu bar actions for workspace navigation, layout reset, tool popups, and prompt copy.
  • Preset 1 visual direction: cream workspace, navy navigation, teal/coral accents.
  • Prompt panel with model, engine, aspect, sampler, seed, size, batch, steps, and CFG controls.
  • Prompt helper card using a lazy browser-side Transformers.js model for lightweight prompt structure suggestions.
  • Canvas preview with scrollable output stage, action chips, and bottom output dock.
  • Model inventory grouped by detected engine family.
  • Data workspace for recent output receipts, artifact buckets, observed resolutions, and output root status.
  • Logs workspace for runtime status, backend-discovered log files, event rows, and resource snapshots.
  • Settings workspace for backend paths, generation defaults, layout memory, and runtime flags.
  • Monitor workspace for runtime state, resource meters, queue status, and recent events.
  • Scroll-safe panels and modal bodies for long settings and dense controls.

Image Generation

  • txt2img, img2img, and inpaint modes through the Diffusers backend.
  • Stable Diffusion 1.5, SDXL, SD3.5, Flux, Flux.2 Klein, Z-Image, Qwen Image, SANA, and related local checkpoint families where wired.
  • sampler, scheduler, steps, CFG, seed, size, VAE, clip skip, and hires fix controls where supported by the selected family.
  • live preview, interrupt, continuous generation, and job history in Studio.
  • prompt styles, wildcards, prompt files, dynamic prompt syntax, and Compel support.
  • textual inversion embedding discovery and prompt insertion.
  • PNG metadata writing and PNG Info import back into generation controls.
  • prompt receipt summaries after generation.

LoRA And Extra Networks

  • local LoRA scanning from dedicated LoRA folders and configured extra roots.
  • LoRA metadata inspection, aliases, saved strengths, and trigger-word helpers.
  • Studio prompt tools for quick LoRA insertion and multi-slot LoRA stack composition.
  • Runtime SD/SDXL/SD3.5-style LoRA application through Diffusers adapter loading.
  • Runtime compatibility checks that block mismatched LoRA architecture families.
  • LoRA fuse worker for CPU-side Diffusers folder export with one or more selected LoRAs.
  • Wan stage LoRAs for supported 5B and high/low transformer routes.
  • Wan LoRA filename/runtime filtering so 5B and 14B/A14B LoRAs are not casually mixed.
  • Flux and newer transformer-image LoRA application is intentionally blocked until wired and tested.
  • ONNX LoRA support is not implemented; it requires model merge/export handling.

Inpaint, Masking, And Control

  • inpaint image/mask editor flow.
  • keep-original and last-result source handling.
  • outpaint canvas expansion.
  • SAM-assisted mask presets when local SAM models are installed.
  • single ControlNet unit in the Image advanced panel.
  • local ControlNet model discovery and selection.
  • lightweight preprocessors where available.
  • Control LoRA detection for supported SD1.5 ControlNet LoRA checkpoints.

Model Management

  • local checkpoint, LoRA, VAE, ControlNet, SAM, enhancement, Wan, Flux, LTX, ReActor, and audio model scanning.
  • model sorter and header inspection for common checkpoint, LoRA, quantized, and video model formats.
  • curated download catalog entries for common local model folders.
  • import helpers for model folders from A1111, ComfyUI, and shared libraries.
  • model aliases and prompt trigger-word helpers.
  • checkpoint blend worker.
  • conversion worker for supported single-file to Diffusers-folder exports.
  • quantization receipt/export tooling for selected safetensors workflows.

Enhance And Segment

  • image upscale.
  • GFPGAN and CodeFormer style restoration when models are installed.
  • old-photo restore pipeline.
  • tiled upscale controls for local VRAM limits.
  • SAM mask generation when SAM weights are installed.
  • GroundingDINO text-guided boxes when the optional dependency is available.

Image Lab

  • image route maturity matrix tracking against AUTOMATIC1111 parity expectations.
  • XYZ plot runner.
  • batch img2img/inpaint runner.
  • loopback runner.
  • native GET /api/v1/image/maturity endpoint.

Video

  • Wan image-to-video through explicit local routes: 5B safetensors, 14B FP8/safetensors, or matched GGUF High Noise plus Low Noise transformer pairs.
  • Wan runtime filtering for high/low pair selection, VAE generation, text encoder kind, and LoRA compatibility.
  • Wan preflight and route diagnostics before launch.
  • optional LTX 2.3 text/image-to-video through an isolated worker engine.
  • optional RIFE post-processing to write 30 FPS or 60 FPS output after generation.
  • standalone RIFE frame interpolation tab for existing videos.
  • optional ReActor post-processing from first key frame, uploaded image, or saved face model.
  • optional NVIDIA RTX VSR / Video Effects SDK post-processing when the SDK is installed.
  • conservative route selection so incompatible Wan model families are not mixed by accident.

Audio And Video-Audio

  • standalone Audio tab for generating music or sound effects after a video.
  • optional video-conditioned audio post-processing through MMAudio in an isolated engine venv.
  • generated audio muxing after video when a supported local audio backend is installed.
  • audio lab architecture for future multitrack project, waveform, MIDI, and plugin workflows.

Monitoring, Logs, And Data

  • generated-output history and library search over saved outputs.
  • JSONL generation logging with --genlog for SD, SDXL, and Wan timings without prompt text.
  • Pro Data workspace for artifact receipts and output health.
  • Pro Logs workspace for runtime rows, event rows, discovered log files, and resource snapshots.
  • Pro Monitor workspace for runtime state, resource meters, queue status, and recent events.

Settings, Paths, And Remote Access

  • saved workspace settings.
  • launch settings for GPU, network, API, attention, and runtime behavior.
  • extra model-library roots and extra checkpoint roots.
  • external tool path settings for SDKs and optional engines.
  • Tailscale-friendly remote access guidance.
  • security warnings for network listening without auth.

Training And Local AI Development

  • training tabs are hidden/opt-in while generation paths are hardened.
  • Kohya LoRA training engine status and isolated engine design exist, but training is not a default user path.
  • ED2 full fine-tune worker scaffolding exists behind isolated engine setup.
  • LLM LoRA/QLoRA/full fine-tune config builders exist for future local assistant work.
  • Training must remain human-confirmed; no assistant or UI automation should auto-start training.

Experimental Or Gated

  • Flux LoRA, Flux ControlNet, Flux img2img, and Flux inpaint are blocked until wired and tested.
  • ONNX LoRA, textual inversion, and VAE swapping are not implemented.
  • TensorRT engine cache, shape ranges, and LoRA refit are Engine Lab only.
  • SageAttention and similar acceleration paths remain benchmark-gated.
  • Runtime attention diagram: CUDA 13 attention experiment.
  • Wan resident/streamed offload modes remain under performance validation.
  • VQA/image QA helper for Pro UI is planned but not wired yet.
  • Plugin ecosystem and workflow authoring are planned.