Repository navigation
Image Studio v0.1.1
Image Studio v0.1.1
Image Studio is a local-first image creation desktop app built with Rust,
Tauri, and React. This release contains unsigned builds for Windows x64 and
Apple Silicon macOS. 0.1.1 is the first update after the public
v0.1.0: it turns 8 GB VRAM support, the low-memory install path, and the
Windows desktop reliability work into downloadable builds for both platforms.
Image Studio 是使用 Rust、Tauri 与 React 构建的本地优先桌面图片创作工作台。
本次发布包含 Windows x64 与 Apple Silicon macOS 构建,当前产物均未签名。
0.1.1 是 v0.1.0 公开后的第一个更新:把 8 GB 显存支持、安装期低内存
量化路径与 Windows 桌面可靠性修复变成两平台可直接下载的构建。
What's New in 0.1.1 / 0.1.1 更新内容
8 GB VRAM compact tier / 8 GB 显存紧凑档
- Five more ways to generate on an 8 GB card. Qwen-Image 2.1,
Qwen-Image 2.1 Turbo, MOSS-Image-SFT, and FLUX.2 Klein 4B (text-to-image and
edit) now run on 8 GB Turing cards through a frozen compact recipe:
768×768,float16, measured CUDA reserved peaks 5.44–5.96 GiB on an
RTX 2060 SUPER 8 GB (sm75, Windows 11). BiRefNet and Real-ESRGAN were never
VRAM-bound. - Automatic tier selection. The host reads the device compute capability;
the model page and composer state the compact tier, and a device below both
tiers keeps the model disabled with the unmet requirement as its reason. - Z-Image Turbo stays at 12 GB. The model overflows in fp16 on BF16-less
Turing architectures; the measurement trail is in
doc/16-zimage-quantization-investigation.md. - 8 GB 显卡可用的生成能力增至五个。 Qwen-Image 2.1、Qwen-Image 2.1 Turbo、
MOSS-Image-SFT 与 FLUX.2 Klein 4B(文生图与编辑)通过冻结的紧凑档配方
768×768、float16在 8 GB Turing 卡上运行;RTX 2060 SUPER 8 GB(sm75,
Windows 11)实测 CUDA reserved 峰值 5.44–5.96 GiB。BiRefNet 与
Real-ESRGAN 本来就不受显存限制。 - 自动选择档位。 宿主读取设备计算能力,模型页与输入区标注紧凑档;低于
两档门槛的设备保持禁用,并显示未满足的最底要求。 - Z-Image Turbo 仍要求 12 GB。 该模型在无 BF16 的 Turing 架构上存在 fp16
溢出,完整测量过程见
doc/16-zimage-quantization-investigation.md。
Low-memory install-time quantization / 安装期低内存量化
- Install-time NF4 preparation now runs in a shard-streaming child process.
Each ~1 GiB tensor group is prepared by a short-lived child, so the worker
commit peak is 1.95–3.6 GiB instead of 16.4–23.1 GiB. The installer gate
drops from available physical RAM ≥ 28 GiB or commit headroom ≥ 26 GiB to
available physical RAM ≥ 8 GiB and commit headroom ≥ 6 GiB: a 16 GiB
Windows machine can install Qwen-Image 2.1 and MOSS-Image-SFT, and the
prepared caches stay byte-identical. - Standard-tier generation fixed after the new install path. The Qwen text
encoder is now loaded withdevice_map="cpu", low_cpu_mem_usage=Truein the
standard tier as well, so the staged CUDA cleanup actually releases reserved
memory. - 安装期 NF4 量化改为分片流式。 每个约 1 GiB 的 tensor 组由短生命周期子进程
量化,worker commit 峰值从 16.4–23.1 GiB 降到 1.95–3.6 GiB。安装门槛从
可用物理内存 ≥ 28 GiB 或 commit 余量 ≥ 26 GiB 降到
可用物理内存 ≥ 8 GiB 且 commit 余量 ≥ 6 GiB:16 GiB Windows 机器也能安装
Qwen-Image 2.1 与 MOSS-Image-SFT,且准备好的缓存逐字节一致。 - 新安装路径下的标准档生成已修复。 标准档 Qwen 文本编码器同样改为
device_map="cpu", low_cpu_mem_usage=True装载,分段 CUDA 清理能够真正
释放 reserved 显存。
Windows desktop reliability / Windows 桌面可靠性
- No more re-hashing the install on every send. The 25–35 GB install is
verified once; send preflight dropped from 60 s+ stalls to ~46 ms. - Stranded turns are bounded. A 60 s turn watchdog fails a wedged send with
a retry entry instead of leaving the composer spinning. - Cancellation is fast. The process group is force-reaped, so cancel settles
in 3–7 s instead of ~120 s, with no orphan processes. - Honest install state. Model rows project the real installer phase
(includingpreparing_model) and no longer claim "not installed". - Downloads recover. Every artifact retries up to five times with backoff,
and a part that fails checksum validation is deleted and re-downloaded instead
of blocking the install forever. - No console windows. Supervisor, worker and installer children are created
withCREATE_NO_WINDOW, and the flag now survives suspended job spawns. - A real log file.
<data>/logs/app.log(8 MiB rotation) records panics,
generation-time memory, and structured install/generation events. - Clearer failures. Failure copy was clarified and localized, the
hardware-blocked message is scoped to the model it applies to, the ncnn probe
log decodes non-ASCII paths lossily, and native save dialogs are populated
with the current file name. - 不再重复哈希整个安装。 25–35 GB 安装内容只在安装时校验一次;发送预检从
60 s+ 卡顿降到约 46 ms。 - 滞留回合有界。 60 s 回合看门狗会把卡死的发送判为失败并提供重试入口,
输入区不再无限转圈。 - 取消更快。 超期时强制回收整棵进程组,取消从约 120 s 降到 3–7 s,
且没有残留进程。 - 安装状态可信。 模型行投影真实安装阶段(含
preparing_model),
不再错误显示"尚未安装"。 - 下载可恢复。 每个文件最多 5 次退避重试;校验失败的分片会被删除并重新
下载,而不是让安装永远失败。 - 不再弹终端窗口。 supervisor、worker 与安装器子进程统一使用
CREATE_NO_WINDOW,并且该标志在挂起作业 spawn 后仍然生效。 - 有真正的日志。
<data>/logs/app.log(8 MiB 轮转)记录 panic、生成期
内存,以及结构化的安装/生成事件。 - 错误文案更清楚。 失败提示已清理并中文化,硬件门槛提示只针对对应型号,
ncnn 探测日志对非 ASCII 路径做有损解码,原生保存对话框会填入当前文件名。
Verification / 验证
- CI quality gate passed on both platforms. Rust formatting, Clippy with
warnings denied, workspace tests (199 desktop + integration, 46 protocol),
60 frontend tests, frontend typecheck/build, andopenspec validate --all --strictran on Windows x64 and macOS arm64. - Real-machine acceptance. Windows 11 / RTX 2060 SUPER 8 GB covers the
compact tier, installs, generation, cancellation, watchdog and console
behavior; Windows 10 22H2 / RTX 3080 Ti 12 GB covers the standard tier
(Klein 50 s, Z-Image 55 s, BiRefNet 26 s, Real-ESRGAN 9 s, Klein edit 47 s,
MOSS install 364 s / generation 143 s, Qwen install 415 s / generation 84 s),
native export, restart persistence, non-ASCII data root, download retry and
bad-checksum recovery. - 双平台 CI 质量门通过。 Rust 格式、Clippy
-D warnings、workspace 测试
(桌面 199 + 集成、协议 46)、前端 60 测试与类型检查/构建、
openspec validate --all --strict均在 Windows x64 与 macOS arm64 上运行。 - 真机验收。 Windows 11 / RTX 2060 SUPER 8 GB 覆盖紧凑档、安装、生成、
取消、看门狗与控制台窗口;Windows 10 22H2 / RTX 3080 Ti 12 GB 覆盖标准档
(Klein 50 s、Z-Image 55 s、BiRefNet 26 s、Real-ESRGAN 9 s、Klein 编辑 47 s、
MOSS 安装 364 s / 生成 143 s、Qwen 安装 415 s / 生成 84 s)、原生导出、
重启持久化、非 ASCII 数据根、下载重试与坏校验恢复。
Known Limitations / 已知限制
- Z-Image Turbo requires a 12 GB card. The compact recipe identity exists in
the protocol, but the catalog deliberately does not offer it on 8 GB. - The 8 GB compact tier is certified on Turing sm75 (RTX 2060 SUPER). Other
8 GB architectures are expected to work through the same recipes but are not
certified in this release. - Windows virtual memory still matters. The new install gate is RAM/commit
based; a machine whose auto-managed pagefile leaves very little commit
headroom still needs a larger virtual-memory setting before installing
Qwen-Image 2.1 or MOSS-Image-SFT. - Unsigned builds. Windows SmartScreen and macOS Gatekeeper can block the
first launch; verify the SHA-256 first (see Installation). - 模型与运行时单独下载。 每个生成模型首次安装需要下载约 25–33 GB,并准备
NF4 缓存;这些内容不随安装包分发。 - Z-Image Turbo 要求 12 GB 显卡; 紧凑档配方身份保留在协议中,但目录在
8 GB 设备上刻意不提供该型号。 - 8 GB 紧凑档仅在 Turing sm75(RTX 2060 SUPER)上认证; 其他 8 GB 架构
预期可通过同一配方运行,但本版本未认证。 - 16 GiB 级机器的 Windows 虚拟内存设置会影响安装。 新的安装门槛基于
可用内存/commit;自动管理页面文件导致 commit 余量过小的机器仍需调大
虚拟内存。 - 产物未签名。 Windows SmartScreen 与 macOS Gatekeeper 可能拦截首次
启动;请先核对 SHA-256(见安装说明)。
Downloads / 下载
- Windows x64
Image.Studio_0.1.1_x64-setup.exe - macOS 11+ Apple Silicon
Image-Studio-0.1.1-macos-arm64.zip - Verify with
SHA256SUMS-windows.txtorSHA256SUMS-macos.txt.
请使用对应平台的SHA256SUMS-*.txt校验文件。
sha256sum -c SHA256SUMS-windows.txt
sha256sum -c SHA256SUMS-macos.txtInstallation / 安装
Windows
The installer uses a current-user NSIS package. SmartScreen may show
“Windows protected your PC” because the build is unsigned. Verify the SHA-256
first, then use More info → Run anyway.
安装包采用 current-user NSIS。由于未签名,SmartScreen 可能提示
“Windows 已保护你的电脑”;请先核对 SHA-256,再选择“更多信息 → 仍要运行”。
macOS
The app is unsigned and not notarized. Gatekeeper may block the first launch.
Verify the SHA-256, then Control-click the app and choose Open, or run:
应用未签名、未公证,Gatekeeper 可能阻止首次启动。请先核对 SHA-256,然后
Control-click 应用并选择“打开”,或运行:
xattr -dr com.apple.quarantine "Image Studio.app"What Is Included / 包含内容
- Seven local image models covering text-to-image, single-image editing,
background removal with RGBA/mask output, and exact 4× upscaling, with the
compact 8 GB tier on top of the standard 12 GB tier. - A conversation-based creation workspace, works library, task center, model
manager, and persistent settings. - Verified Rust/Tauri desktop hosts for Windows x64 and macOS arm64.
- 七个本地图片模型,覆盖文生图、单图编辑、去背景 RGBA/mask 与精确 4× 放大;
在标准 12 GB 档之外新增 8 GB 紧凑档。 - 包含会话式创作、作品库、任务中心、模型管理与持久设置。
Model weights and private Python runtimes are downloaded separately from fixed
upstream sources and are not bundled in these archives.
模型权重和私有 Python 运行时不随安装包分发,需要在应用内从固定上游源安装。
Validation / 验证
The release was built and checked by GitHub Actions on Windows x64 and macOS
arm64. Workflow checks cover Rust formatting, Clippy with warnings denied,
workspace tests, frontend tests/build, OpenSpec strict validation, package
architecture, checksums, version consistency, and tag provenance.
本版本由 GitHub Actions 在 Windows x64 与 macOS arm64 runner 上构建和检查;
覆盖 Rust 格式、Clippy、workspace 测试、前端测试/构建、OpenSpec strict、
包架构、checksums、版本一致性及 tag 来源。CI 不下载模型权重,也不执行 GPU
推理;真实模型验收证据保存在仓库 verification/。
License / 许可证
Image Studio is licensed under Apache-2.0. See LICENSE, NOTICE, and
THIRD_PARTY_NOTICES.md. Model and runtime materials remain subject to their
upstream licenses.
Image Studio 采用 Apache-2.0;模型与运行时材料继续遵循各自上游许可证。