AI media harness for planning, generating, and editing locally
TypeScript
Tighten README features and setup
Bitti7991af703b15
diff --git a/README.md b/README.mdindex a13e933..512a06d 100644--- a/README.md+++ b/README.md@@ -26,43 +26,47 @@ Slopus brings an AI assistant, a scene generator, reusable references, a timelin## Features-- **Two project types.** Video projects have **Timeline**, **Generator**, **References** and **Export** tabs; Image projects have **Editor**, **References** and **Export**.-- **Agent.** Claude Code, Codex, OpenRouter, or a local OpenAI-compatible model in a docked pane (Ctrl+Shift+A). It can plan scenes, build references, and edit video and image projects.-- **Scene generator.** A board of scenes and shot tiles. Scene types: First & last frame, Animate, Pose, Character replace, Extend and Bridge. Attach references and start frames, and set steps, seeds and generator presets. Runs in the **Work queue**, where jobs can be cancelled.-- **Shot controls.** Set shot timing, visual action, spoken lines and language, with camera and performance settings plus scene-wide look, sound and music. Insert reusable references as smart chips in shot descriptions.-- **References.** Characters, animals, products, locations and styles with descriptions, images and generated icons. Start from searchable presets, browse as Icons or Details, and see which scenes use each one. Attach video references, preview and trim their clip ranges, and choose whether to include their soundtracks.-- **RefMods.** Attach pre-encoded H3 reference `.safetensors` files to reusable references, with independent **Strength** and **Copies** controls. Use them with Ref2VA generators for video, still images, image edits and reference icons.-- **LoRAs.** Add local adapters or download built-in options such as TaoMate 3-Step, Turbo and LightX2V Turbo. Combine multiple LoRAs in a generator template, reorder them, adjust their strengths, and set optional step-count overrides.-- **LAN workers.** Run `slopus-worker` on another computer and choose it in **Settings → Workers** to generate there. Workers are found on the local network automatically (mDNS) or added by address. Weights with download links download on the worker; local-only weights, references and source images are sent from this computer. Results are saved to the project as if they were generated locally.-- **Generator templates.** Download model weights from Settings with progress, cancellation and retry, or use your own local files. Save model and attention settings per template, and optionally enable **MotionCache** to reuse denoising work in supported generation modes.+- **One workspace.** Every project has **Timeline**, **Video**, **Image**, **References** and **Export** tabs.+- **Agent.** Plan scenes, create references and edit projects with Claude Code, Codex, OpenRouter or a local OpenAI-compatible model. Open the agent pane with Ctrl+Shift+A.+- **Scene generator.** Generate scenes with First & last frame, Animate, Pose, Character replace, Extend or Bridge. Set steps, seeds and presets; manage jobs in the **Work queue**.+- **Shot controls.** Set timing, action, dialogue, language, camera and performance. Add scene-wide look, sound and music, and cite references with smart chips.+- **References.** Reuse descriptions, images and video clips across scenes. Start from searchable presets, see where references are used, and trim clips with optional sound.+- **RefMods.** Use pre-encoded H3 `.safetensors` references with Ref2VA generators for video, images, edits and icons. Set **Strength** and **Copies** per attachment.+- **LoRAs.** Combine local or downloadable adapters in generator templates, with ordering, strength and step-count controls.+- **LAN workers.** Generate on another Windows or Linux computer. Workers download available weights, receive local files as needed, and return results to your project.+- **Generator templates.** Download weights or use local files. Save model and attention settings per template, with optional **MotionCache** for supported modes.- **Timeline.** Video and audio tracks with thumbnails, waveforms, snapping, trimming, splitting, zoom and keyboard shortcuts. Transform, opacity, chroma key, fades and wipes, plus GPU effects (sharpen, blur, color correction, vignette, `.cube` LUTs) that can be bypassed.-- **Image editor.** Build a hierarchy of objects, text and groups with placement boxes, then generate with MiniMax H3. Use **Image root** to open a PNG or JPEG (or **Edit** a result) and apply edits in hierarchy order with bounding-box inpainting. Results and drafts stay in the image bar.+- **Image editor.** Generate images from prompts, objects and text, or open an image and describe whole-image or local edits. Keep originals, drafts and versions in the image bar.- **Export.** Video: MP4 in H.264, VP9 or AV1 (as your encoders allow) at a chosen resolution, frame rate and quality; exports keep running in the background. Images: pick one from the image bar and save it as JPG or PNG at a chosen size and quality.- **Undo everywhere.** Save, Undo and Redo sit in the title bar; Ctrl+Z and Ctrl+Y cover edits from any view, the agent and project settings.- **Your data.** Projects and media stay on disk. API keys stay in this computer's settings, outside project files.## Setup-Local video generation requires separately installed model weights and GPU support. In **Settings → Generator**, configure a generator's model paths and check that its status reads **Ready**. Configure OpenRouter or a local model endpoint in **Settings → Agents** to make it available in the assistant's agent list.+In **Settings → Generator**, download or select model weights and check that the generator reads **Ready**. Local generation requires a supported GPU. Configure OpenRouter or a local model endpoint in **Settings → Agents**.### Linux-Slopus runs on x86_64 Linux. Install `Slopus-<version>-linux-x64.deb` (Debian, Ubuntu) or `Slopus-<version>-linux-x64.rpm` (Fedora, openSUSE), or make `Slopus-<version>-linux-x64.AppImage` executable and run it; the AppImage updates itself. The app needs WebKitGTK 2.42 or later for video decoding, export and GPU effects. Local generation has the same requirements as the Linux worker below. See [docs/linux.md](docs/linux.md).+Install the x86_64 `.deb` or `.rpm` package, or make the AppImage executable and run it. The AppImage supports in-app updates.++Requires WebKitGTK 2.42+ for video and GPU features. The bundled generation runtime requires NVIDIA driver libraries and CUDA 13 cuBLAS, even when using Vulkan. See the [Linux guide](docs/linux.md) for details.### Generating on another computer-Workers run on Windows or Linux. On a computer with a capable GPU on the same network, unpack the worker package and start it:+Unpack a worker package on a GPU-equipped computer on your network:| Platform | Package | Start || --- | --- | --- |-| Windows x64 | `Slopus-<version>-windows-x64-worker.zip` (`slopus-worker.exe`, `slopfab.dll`) | `slopus-worker.exe --weights D:\Models` |-| Linux x86_64 | `Slopus-<version>-linux-x64-worker.tar.gz` (`slopus-worker`, `libslopfab.so`) | `./slopus-worker --weights ~/models` |+| Windows x64 | `Slopus-<version>-windows-x64-worker.zip` | `slopus-worker.exe --weights D:\Models` |+| Linux x86_64 | `Slopus-<version>-linux-x64-worker.tar.gz` | `./slopus-worker --weights ~/models` |-The Linux worker needs no desktop or UI libraries, but it does need an NVIDIA GPU with the NVIDIA driver and CUDA 13 cuBLAS (`libcublas.so.13`); it looks for cuBLAS in `/usr/local/cuda*` if the system linker does not find it. Allow the worker through the firewall (TCP 47321, and UDP 5353 for discovery).+1. Start the worker and allow TCP 47321 and UDP 5353 through the firewall.+2. In **Settings → Workers**, click **Use** beside the discovered worker, or add it by address. Missing weights download on first use when download links are available.+3. Select **This computer** to switch back to local generation.-In Slopus, open **Settings → Workers** and click **Use** next to the worker. Generators whose weights have download links become usable even if they are not downloaded on this computer; the worker downloads missing weights into its weights folders on first use. Run `slopus-worker --help` for options: `--port`, `--name`, `--token` (require an access token), `--data`, `--weights` (repeatable), `--backend` and `--no-mdns`. The worker uses CUDA when it is available and otherwise falls back to Vulkan, printing a warning if an NVIDIA GPU is present without a usable CUDA installation. On Windows it uses CUDA 13 if installed, otherwise CUDA 12; the Linux runtime uses CUDA 13. `--backend cuda13`, `--backend cuda12`, `--backend cuda` or `--backend vulkan` overrides the choice. Choose **This computer** to generate locally again.+Windows workers prefer CUDA 13, then CUDA 12, with Vulkan as a fallback. Linux workers also support `--backend vulkan`, but the bundled runtime still needs the libraries above. No desktop libraries are required. Run `slopus-worker --help` for backend, access-token and port options.-The project's GPU setup recommendations are:+Windows GPU setup:| GPU | Generation backend || --- | --- |diff --git a/docs/linux.md b/docs/linux.mdindex 546d16d..9439012 100644--- a/docs/linux.md+++ b/docs/linux.md@@ -41,9 +41,12 @@ On Linux `libslopfab.so` is installed in Tauri's resource folder,executable first, as on Windows and in development builds, and then in`../lib/Slopus/`.-Generation has the same requirements as the Linux worker: an NVIDIA GPU with its-driver and CUDA 13 cuBLAS (`libcublas.so.13`), found in `/usr/local/cuda*` when-the system linker does not.+The bundled `libslopfab.so` directly links NVIDIA's driver library+(`libcuda.so.1`) and CUDA 13 cuBLAS (`libcublas.so.13`). Both must be installed+for the runtime to load, even when generation uses Vulkan. The Linux worker+supports `--backend vulkan`; this selects Vulkan for computation but does not+remove those library dependencies. Slopus also searches `/usr/local/cuda*` for+cuBLAS when the system linker cannot find it.## WebKitGTK