koharu-rs/koharu

Description On Linux the WebGPU canvas is unavailable or falls back to software rendering on machines where the GPU stack is fine. crates/koharu-app/src/app.rs sets the WebGPU switches through Cef::command_line_args: .enable_features(["Vulkan", "VulkanFromANGLE"]) .command_line_args([ ("--enable-unsafe-webgpu", None), ("use-angle", Some("vulkan")), ("--ozone-platform", Some("x11")), ]) but in tauri-runtime-cef those switches are applied to the browser process only. Chromium builds each child process's command line from a fixed allowlist (ChromeContentBrowserClientCef::AppendExtraCommandLineSwitches) and drops everything else, so the GPU process never sees use-angle=vulkan or the Vulkan/VulkanFromANGLE feature names. The GPU process then initializes ANGLE on its default backend instead of Vulkan, and navigator.gpu reports no usable adapter ("webgpu found no adapters") — the canvas shows as unavailable or runs on SwiftShader. --ozone-platform=x11 (#1110) reached the GPU pro

Description Would be great to have selected text box to move by 1 pixel with directional key input. Thank you. Additional context No response

Description Suggestions for the local model catalog and defaults, from about a week of measuring models for manga translation into Spanish on a 16 GB VRAM GPU (RX 9060 XT) with 16 GB of system RAM. The numbers are small benchmarks graded against human translations, not a formal evaluation, but the gaps are large enough to be useful. Translation: Gemma 4 12B it QAT (already in the catalog as unsloth/gemma-4-12B-it-qat-GGUF; I run the lmstudio-community Q4_0 through LM Studio). It is the best local translator I found for this hardware. On 74 ambiguous lines that only their scene resolves, with a work note (see #339 and #1013): Source Gemma 4 12B DeepL English 7/8 3/8 Japanese 4/5 2/5 Chinese 4.7/5 3/5 Korean 88% 47% Qwen3 8B scored 71% on Korean, and EXAONE 3.5 7.8B (abliterated) 47%. On a real 23-page Japanese chapter (231 balloons), the full pipeline reached about 94% of balloons without a serious error. Two settings mattered a lot for Gemma: Thinking off. With reasoni

Description When you launch the application, nothing can be typed anywhere, including for the name of a new project. In the settings menus, nothing can be typed there either. Essentially, the application is completely unusable. Reverting back to version 0.83.0 fixes the problem. I would suspect that this issue is caused by the upgrade to Tauri v3 alpha (4a3c7b0), since it seems to be the only plausible culprit between 0.83.0 and 0.83.1. Steps to reproduce Open the application Click in the textbox for a new project name Attempt to type anything Your environment Koharu: 0.83.1 and later are affected OS: Fedora 44 KDE GPU: GeForce RTX 2060 Mobile Additional context No response

Issue Translation model pickers are empty with no error despite local hardcoded models being there. get_translation_models never resolves, so the list never updates. Cause openai-compatible and lm-studio ship default base URLs (http://localhost:11434/v1, http://localhost:1234), so they are probed even when unconfigured. The HTTP client has no read timeout and the providers are awaited together, so a probed localhost endpoint that never responds hangs the whole listing, local models are withheld too. Proposed Fixes Opt-in enabled switch (off by default) for openai-compatible and lm-studio. Make the base URLs blank by default so no provider is ever probed unless a URL is explicitly configured. Add a per-provider timeout (30s deadline) so one stalled endpoint no longer hangs the whole listing. Steps to reproduce Open the translation model dropdown. Your environment OS: Windows 11 Koharu: 0.82.1 Additional context No response