AI media harness for planning, generating, and editing locally
TypeScript
Scope release notes to 0.3.2
Bittib6780b7133fd
diff --git a/RELEASE_NOTES.md b/RELEASE_NOTES.mdindex 19ce0d9..744b91f 100644--- a/RELEASE_NOTES.md+++ b/RELEASE_NOTES.md@@ -1,47 +1,21 @@# Slopus 0.3.2-This alpha update adds SeedVR2 and Real-ESRGAN export upscaling, Long Shot and Backdrop scenes, latent upscaling, clip speed controls, and a Settings agent, alongside improvements to generation, playback and memory use.+This alpha update adds Prompt video scenes, agent image drafts, Sol attention and sigma shift controls, and improves Zoom image editing, agent chat and generation reliability.## What's new- **Prompt video scenes.** New ordinary video scenes default to Prompt, with text and reference guidance but no start- or last-frame selectors. First & last frame remains available for frame conditioning.- **Agent image drafts and generation.** Ask the agent to prepare multiple images in the image bar for later, edit a selected draft, or generate it with a chosen generator. Preparation is saved before rendering, and the agent resumes after generation with the saved result.- **Sol attention in Settings.** Choose Sol attention under a generator's Performance settings for generation on CUDA GPUs. The selection is saved per generator and included in shared generator files.-- **Zoom image template (formerly Extend).** Draw or resize the output box, choose a generation resolution, and optionally describe the new surroundings. Output follows the box's aspect ratio and selected pixel budget, resizing the source to fit. Boxes inside the source regenerate the selected crop with refined detail at the target resolution. Positioned source conditioning keeps the scene anchored. SlopFab 1.28 refines new surroundings with Langevin iterations and blends edges with a Gaussian mask.-- **SeedVR2 upscaling for images and videos.** Restore exports at the selected output size using SeedVR2 3B FP16 and its VAE. Run on the local NVIDIA CUDA GPU or a selected CUDA LAN worker. Workers download missing standard weights or receive your selected local weight files.-- **Real-ESRGAN upscaling for images and videos.** Use RealESRGAN x4plus during export, with the result resized to your chosen output dimensions. CUDA and Vulkan are supported. Choose None to export without an upscaler.-- **Other weights settings.** Settings → Generator now includes downloads and optional local file paths for Real-ESRGAN, SeedVR2 and the MiniMax H3 latent upscaler. Upscaler downloads share the generator and LoRA download queue, with progress and cancellation in Work Queue.-- **Queued image and video exports.** Exports share a work queue with progress, cancellation and retained results, including exports using SeedVR2 on a worker.-- **Long Shot scenes.** Generate alternating fresh clips and latent bridges in 1 → 3 → 2 → 5 → 4 order. Bridge controls set how much of the previous ending and following beginning may change. Preview, playback and export use joined video and audio, and later bridges preserve earlier changes. Local and LAN worker generation are supported.-- **Backdrop scenes.** Generate subjects and action against a selected green, blue, black or white background.-- **Expanded Chroma Key.** Matching backdrop modes and automatic border detection add screen gain/balance, matte levels, background color removal, despill, shadow removal, edge softness, choke/grow and small-speck/hole cleanup in GPU preview and export.-- **Latent upscale for scene generation.** Generate at half resolution and upscale to the selected output size, locally or on a LAN worker. The toggle appears below Seed in Scene Generation settings, and its model is available in Other weights.-- **Clip speed controls.** Timeline clips now offer Speed in Timing, from 25% to 400%. Changes preserve the source range, adjust duration and later clips on the same track, and apply to preview and export, including audio pitch.-- **Timeline preview zoom.** Zoom up to 400% using the zoom menu, buttons or Ctrl+wheel, and scroll to inspect details. Choose Fit to see the whole frame again.-- **Agent in Settings.** Open a resizable agent panel from the Settings top bar to configure generators, weight paths and LoRAs, even without an open project.-- **Video and audio sigma shifts.** Set shifts and a separate audio step count under Advanced options at the bottom of Generation in the right panel, define generator template defaults, and configure per-stream LoRA overrides. Local and LAN worker generation use the same settings.-- **Separate audio generation steps.** Audio follows the video step count by default, with optional overrides for local and worker generation.-- **Continuation overlap locking.** Continue scenes offer Lock Overlap for local and worker generation, constraining overlapping video and audio to the source scene.-- **Selectable character sheet views.** Enable Close up, Front view, Side view and Back view independently. Side view is off by default; generation, combined images and debug prompts include only enabled views. When enabled, the front view provides a shared clothing reference for the other views.-- **1280 export presets.** Image and video export now offer 1280 × 1280, 2276 × 1280, 1280 × 2276 and 1280 × 1600 presets for the supported aspect ratios.+- **Zoom image template (formerly Extend).** Restored image extension with an optional prompt and output sized to the drawn box's aspect ratio and selected pixel budget. Boxes inside the source regenerate the selected crop with refined detail at the target resolution. Positioned source conditioning, native refinement and edge blending improve the generated surroundings.+- **Video and audio sigma shifts.** Set separate shifts for video and audio, define generator template defaults, and configure per-stream LoRA overrides. Local and LAN worker generation use the same settings. Sigma shifts and the existing audio step override are grouped under Advanced options at the bottom of Generation in the right panel.## Fixes and improvements-- Fixed CUDA transformer-loading memory failures caused by pinning entire large checkpoints in system memory. Uploads now use bounded buffers, account for weight-conversion memory and retry allocation failures with more CPU offloading while preserving precision and output settings.-- Agent chat appends replies and retains earlier messages, corrections and errors throughout the conversation. Raw JSON and command payloads stay hidden; readable replies and summaries are shown.+- Fixed CUDA transformer-loading memory failures with bounded upload buffers, more accurate memory budgeting and automatic retries with more CPU offloading, while preserving precision and output settings.+- Agent chat retains earlier replies, corrections and errors throughout the conversation. Raw JSON and command payloads stay hidden; readable replies and summaries are shown.+- Zoom reduces visible extension seams by preventing grey canvas padding from affecting source encoding and widening the edge blend. Small source areas retain an unblended interior, and source transparency is retained.- Backdrop prompts describe a flat, constant color field in every shot, limit reference guidance to subject appearance, and use neutral studio styling when no look is specified.-- Zoom reduces visible extension seams by keeping grey canvas padding out of source VAE encoding and widening the native edge blend. The full source remains the composition reference, small source areas retain an unblended interior, and source transparency is retained.-- Generate all includes the final scene in an even-length Long Shot sequence, continuing from the preceding scene and preserving earlier joined video and audio. Four scenes generate in 1 → 3 → 2 → 4 order; adding another scene turns the final continuation into a bridge on its next generation.-- Long Shot timeline playback keeps the same video player running across adjacent scenes in a joined generation, avoiding a playback restart at bridge boundaries.-- Reduced memory retained between generations by releasing consumed SlopFab samples and handles, completed queue snapshots, encoder buffers on errors, and abandoned LAN generation results.-- Still-image resolution and aspect ratio are saved separately from video project settings. Editing or selecting an image no longer changes the canvas used for video generation and continuation.-- Image export offers sizes closest to the selected image's actual aspect ratio, even when project or saved generation settings differ.-- New image scenes default to None style mode, which adds no style instructions to the generated prompt.-- Video and image generation headers show the effective step count from the selected generator's active LoRAs, including fixed sampling schedules.-- Scene generation prompts mark shot times as `[Shot 2] At 00:03.500,`.-- Generation diagnostics separate VAE model opening, weight loading, decoding and cleanup times to help identify intermittent image-generation slowdowns.-- Inspector instructions appear as tooltips on field labels, and Generator and References toolbars align with their inspector headers.-- Renamed the continuation source edge Beginning to Start.-- The agent input clears after sending a message.-- Work queue activity indicators keep moving when Windows interface animations are disabled, and download labels are shorter.-- Updated the bundled Windows and Linux SlopFab runtimes for the new generation and upscaling features and decode diagnostics. Update LAN workers alongside the app.+- Scene type label tooltips now explain each scene type's behavior.+- Renamed Add reference to Reference browser and centered the agent send and stop buttons.+- Updated the bundled Windows and Linux SlopFab runtimes to 1.28.2 for Zoom refinement, seam fixes and transformer-loading memory recovery. Update LAN workers alongside the app.