Skip to content

Image Studio v0.2.0

Choose a tag to compare

@github-actions github-actions released this 12 Nov 15:25
· 48 commits to main since this release

Image Studio v0.2.0

Image Studio 0.2.0 is a local-first image studio release for Apple
Silicon macOS and Windows x64. It focuses the workspace around conversations
and works, adds real conversation/work management, and closes the remaining
desktop reliability gaps found by macOS and Windows acceptance.

Image Studio 0.2.0 是面向 Apple Silicon macOS 与 Windows x64 的本地优先
图片工作台版本。本次更新把界面重新聚焦到会话与作品,补齐真正的会话/作品管理,
并修复了 macOS 与 Windows 验收中剩余的桌面可靠性问题。

Highlights / 更新内容

A focused workspace / 更专注的创作界面

  • One navigation rail. Remove duplicated page tabs, standalone feature and
    task pages, and the development diagnostics drawer. Create, inspiration,
    models, works, theme, about, and settings now form one clear hierarchy.
  • Production records in context. Expand model, seed, dimensions, attempt,
    stage, and execution details directly from a conversation result or a work.
    There is no detour through a separate task page.
  • Licensing stays discoverable. Prompt inspiration keeps source attribution
    while licence details are consolidated in Settings → About.
  • 单层导航。 删除重复的页面页签、独立功能页/任务页与开发诊断抽屉;创作、
    灵感、模型、作品、主题、关于与设置形成单一层级。
  • 制作记录留在上下文。 可在会话结果或作品中直接展开型号、种子、尺寸、
    Attempt、阶段与执行信息,不再跳到独立任务页。
  • 许可信息可追溯。 灵感库保留来源归属,通用许可说明集中到设置 → 关于。

Conversation and work management / 会话与作品管理

  • Conversation rows have both actions. Rename or delete a conversation from
    its row. A destructive action requires confirmation and is blocked while that
    conversation is still receiving, generating, cancelling, or cleaning up.
  • Deleting a conversation keeps its works. Generated works remain available
    for preview, export, favourites, and deletion, with provenance marked as
    “conversation deleted”.
  • Favourite works across pages. Work favourites are persisted separately
    from prompt inspiration, survive restart, and can drive a favourites-only
    filter with stable totals and pagination.
  • Reference-aware deletion. A work used by an active draft, a retained
    capability slot, or an in-flight request is not deleted. The app names the
    conversation, reference type, and action required to release it.
  • Controlled cleanup. Confirmed deletion removes application-managed
    originals, the companion mask, and dedicated preview caches as one recoverable
    operation. External exports and immutable execution history remain.
  • Reliable migration. The local SQLite library upgrades transactionally from
    schema v5 to v6 without changing existing asset identities or model readiness.
  • 会话行同时提供重命名与删除。 删除需确认;会话仍在接收、生成、取消或清理
    时禁止删除。删除会话后,作品继续用于预览、导出、收藏与删除,并注明
    “会话已删除”。
  • 作品收藏独立持久。 作品收藏与灵感收藏隔离,支持跨重启保存和“仅收藏”
    筛选,并保持完整总数与稳定分页。
  • 删除前检查引用。 被当前草稿、未激活能力槽位或活动请求引用的作品不会被
    删除;应用会指出会话、引用类型和解除方式。
  • 受控清理。 确认删除后,应用内原图、同次蒙版与专用预览缓存作为一组可恢复
    操作处理;外部导出副本和不可变执行历史保留。
  • 可靠升级。 本地 SQLite 通过事务从 v5 升级到 v6,不改变既有资产身份和
    模型就绪状态。

Desktop reliability / 桌面可靠性

  • Same-root second launch activates the running window. On Windows, a second
    launch using the same data root restores and focuses the existing window in
    about 186 ms and exits successfully. Different --data-dir roots remain
    independent.
  • Unresponsive instances are explained. If the existing window cannot be
    confirmed within five seconds, the second launch reports that condition
    instead of opening a conflicting WebView2 profile or touching the database.
  • Startup failures stay visible. Panic logging now forwards to the startup
    failure hook, preserving both the native error dialog and
    logs/startup-failure.log.
  • Follow-up interaction fixes. Draft capability slots remain valid across
    model switches, explicit source removal clears the outgoing slot, empty-prompt
    deletions have usable titles, and blocked work deletion explains the exact
    action needed.
  • 同数据根二次启动会激活已有窗口。 Windows 上再次启动同一数据根时,已有窗口
    会在约 186 ms 内还原并聚焦,第二个进程成功退出;不同 --data-dir 数据根仍可
    独立运行。
  • 无响应实例有明确说明。 已有窗口在 5 秒内无法确认时,第二次启动会直接说明
    原因,不会创建冲突的 WebView2 profile,也不会改写数据库。
  • 启动失败不再静默。 panic 日志钩子会继续调用启动失败钩子,原生错误对话框与
    logs/startup-failure.log 同时保留。
  • 补齐交互问题。 切换模型后能力槽位保持有效;显式移除源图会清空旧槽位;
    空提示词作品有可读的删除标题;受引用阻止的删除会给出具体解除动作。

Validation / 验证

The release keeps build, UI, and real inference evidence separate:

  • macOS arm64: the current native validation covers the simplified workspace,
    seven ready models, real text-to-image, BiRefNet, Klein reference editing,
    cancellation, export, v5→v6 upgrade, conversation/work deletion, and
    favourites.
  • Windows 10 / RTX 3080 Ti 12 GB: the current-build acceptance completed
    9/9 real generation cases across seven models and eight capabilities. It also
    covered seeded and real-network installation with resume, work/conversation
    management, v5→v6 upgrade, cancellation, forced-kill recovery, startup
    diagnostics, and second-launch activation.
  • Windows 11 / RTX 2060 SUPER 8 GB: the compact tier supports Qwen-Image 2.1,
    Qwen-Image 2.1 Turbo, MOSS-Image-SFT, and FLUX.2 Klein 4B (text-to-image and
    edit) at 768² with float16. The 0.2.0 functional regression found no
    execution-surface change and passed the current UI, storage, and lifecycle
    checks.
  • GitHub Actions: the release workflow checks Rust formatting, Clippy,
    workspace tests, frontend tests/typecheck/build, OpenSpec, package identity,
    architecture, checksums, and tag/version provenance. CI does not download
    model weights or execute GPU inference.

本次发布继续把构建、界面与真实推理证据分开记录:

  • macOS arm64: 当前原生验收覆盖精简工作区、七个模型就绪、真实文生图、
    BiRefNet、Klein 参考图编辑、取消、导出、v5→v6 升级、会话/作品删除与收藏。
  • Windows 10 / RTX 3080 Ti 12 GB: 当前构建完成七个模型、八项能力共 9/9 个
    真实生成用例,并覆盖种子缓存与真实网络安装/续传、作品与会话管理、v5→v6
    升级、取消、强杀恢复、启动诊断与二次启动激活。
  • Windows 11 / RTX 2060 SUPER 8 GB: 紧凑档支持 Qwen-Image 2.1 / Turbo、
    MOSS-Image-SFT 与 FLUX.2 Klein 4B(文生图与编辑),使用 768²、float16;
    0.2.0 功能回归确认执行面无变化,并通过当前界面、存储与生命周期检查。
  • GitHub Actions: 发布流程检查 Rust 格式、Clippy、workspace 测试、前端
    测试/类型检查/构建、OpenSpec、包身份、架构、checksum 和 tag/版本来源。
    CI 不下载模型权重,也不执行 GPU 推理。

Supported Models / 支持模型

Model / 型号 Capability / 能力 Fixed recipe / 固定配方
Z-Image Turbo Text to image / 文生图 4-bit, 9 steps, 1:1 / 4:3 / 3:4
FLUX.2 Klein 4B distilled Text to image, single-image edit / 文生图、单图编辑 4-bit, 4 steps, guidance 1.0
BiRefNet general epoch_244 Background removal, RGBA + mask / 去背景、RGBA + mask ONNX Runtime CPU, float32
Real-ESRGAN x4plus Exact 4× upscale / 精确 4× 放大 ncnn Vulkan, tile 128
Qwen-Image 2.1 Text to image / 文生图 NF4, 40 steps, guidance 1.0
Qwen-Image 2.1 Turbo Text to image / 文生图 NF4, 8 steps, official Turbo sampling
MOSS-Image-SFT Text to image / 文生图 NF4, 50 steps, guidance 3.0

The standard tier targets at least 12 GB VRAM at 1024² with bfloat16; the
compact tier targets at least 8 GB VRAM at 768² with float16. Z-Image Turbo
requires 12 GB because it overflows in float16 on BF16-less Turing cards.

标准档面向 ≥12 GB 显存、1024² 与 bfloat16;紧凑档面向 ≥8 GB 显存、768² 与
float16。Z-Image Turbo 在无 BF16 的 Turing 架构上存在 fp16 溢出,因此仍要求
12 GB 显存。

Downloads / 下载

  • Windows x64: Image.Studio_0.2.0_x64-setup.exe
  • macOS 11+ Apple Silicon: Image-Studio-0.2.0-macos-arm64.zip
  • Verify with SHA256SUMS-windows.txt or SHA256SUMS-macos.txt.
  • 请使用对应的 SHA256SUMS-*.txt 校验下载文件。

Installation / 安装

Windows

The installer uses a current-user NSIS package and embeds the WebView2 offline
installer. The build is unsigned, so SmartScreen may show “Windows protected
your PC”. Verify the SHA-256 first, then use More info → Run anyway only if
the digest matches.

安装包使用 current-user NSIS,并内置 WebView2 离线安装器。由于产物未签名,
SmartScreen 可能显示“Windows 已保护你的电脑”。请先核对 SHA-256,只有完全一致时
才选择“更多信息 → 仍要运行”。

macOS

The macOS build is unsigned and not notarized. If opening a checksum-verified
app shows “Image Studio.app is damaged and can’t be opened”, move it to
Applications and remove only the quarantine attribute:

sudo xattr -rd com.apple.quarantine "/Applications/Image Studio.app"
open "/Applications/Image Studio.app"

Do not run this command when the checksum is missing or does not match. If the
app is installed elsewhere, replace the quoted path.

macOS 构建未签名、未公证。如果已核对 checksum 的应用显示
“Image Studio.app 已损坏,无法打开”,请把应用放入“应用程序”,然后只移除
该应用包的 quarantine 属性:

sudo xattr -rd com.apple.quarantine "/Applications/Image Studio.app"
open "/Applications/Image Studio.app"

checksum 缺失或与下载文件不一致时不要执行;应用安装在其他位置时请替换引号中的
路径。

Known Limitations / 已知限制

  • The builds are unsigned and not notarized. Windows SmartScreen and macOS
    Gatekeeper can block first launch; verify the checksum before proceeding.
  • Z-Image Turbo requires a 12 GB card.
  • The 8 GB compact tier is certified on Turing sm75
    (RTX 2060 SUPER). Other 8 GB architectures are not certified by this release.
  • Windows install-time NF4 preparation is RAM/commit based. Machines with very
    little commit headroom may need a larger virtual-memory setting before
    installing Qwen-Image 2.1 or MOSS-Image-SFT.
  • Model weights and private Python runtimes are not bundled. Generative models
    download 25–33 GB each and retain prepared caches.
  • Hardware results are scoped to the recorded machines and architectures. CI
    does not replace real GPU generation acceptance.
  • 产物未签名、未公证;Windows SmartScreen 与 macOS Gatekeeper 可能拦截首次启动,
    请先核对 checksum。
  • Z-Image Turbo 仍要求 12 GB 显卡。
  • 8 GB 紧凑档仅在 Turing sm75(RTX 2060 SUPER)上完成认证,不外推到其他
    8 GB 架构。
  • Windows 安装期 NF4 准备取决于可用内存与 commit 余量;余量很小的机器可能需要在
    安装 Qwen-Image 2.1 或 MOSS-Image-SFT 前调大虚拟内存。
  • 模型权重与私有 Python 运行时不随安装包分发;每个生成模型约需下载 25–33 GB
    并保留准备好的缓存。
  • 硬件结果仅限已记录机器与架构;CI 不能替代真实 GPU 生成验收。

License / 许可证

Image Studio is licensed under Apache-2.0. Third-party libraries, fonts,
example prompts/images, model weights, and runtimes retain their upstream
licenses. See LICENSE, NOTICE, and THIRD_PARTY_NOTICES.md.

Image Studio 采用 Apache-2.0;第三方库、字体、示例提示词/图片、模型权重与运行时
继续遵循各自上游许可证,见 LICENSE、NOTICE 与 THIRD_PARTY_NOTICES.md。