selflift-Avatar v0.1.4-experimental

selflift-Avatar v0.1.4-experimental

What's new in v0.1.4

Automatic mode now shows the actual tile count on the first line of the node status panel, not just auto.

• Before high-resolution preparation: Actual tiles: pending. Auto planning is not available during the low-resolution prefix.
• Once planned: Actual tiles: N, including odd counts selected by auto mode. 1 tile is explicitly labeled full frame / no split.
• The completed result stays visible; changing tiling controls or starting a new run clears the old count.
• Actual results are delivered as structured plan data for live events and completed/cached node UI results. Older cached text is still recognized.
• Add a numeric-node-ID lookup fallback for live updates. This is not a general subgraph-routing fix.
• The manual 2 / 4 / 6 / 8 dropdown is never overwritten by auto results.

No sampling algorithm change: tile selection, spatial overlap, audio-mask constraints, temporal upscaler windows and VRAM estimation remain unchanged. 47 CPU tests plus extended frontend callback tests with a stub DOM passed. Restart the ComfyUI backend and hard-refresh the browser after updating. These tests are not a full browser or real-model benchmark.

Download v0.1.4 · Full changelog

---

本次升级内容：v0.1.4

自动模式也会在节点状态栏第一行显示实际分块数量，不再只看到 auto。

• 高清准备前：实际分块数量：待估算。低清阶段尚未确定自动方案。
• 方案确定后：显示 实际 N 块，包括自动模式选出的奇数；1 块明确标注为整图、不拆分。
• 完成后保留结果；修改分块选项或开始新运行时清除旧数量。
• 实时事件和完成/缓存的节点 UI 结果使用结构化方案数据，并兼容旧缓存文本。
• 为实时更新增加数字节点 ID 的查找回退；不代表已修复所有子图事件路由。
• 自动结果不会覆盖手动 2 / 4 / 6 / 8 下拉值。

**本次不改采样算法：**块数选择、空间重叠、音频遮罩限制、upscaler 时间窗口和显存估算保持不变。47 项 CPU 测试及扩展的模拟 DOM 前端回调测试通过。升级后请重启后端并强制刷新网页。测试不等于完整浏览器或真实大模型效果验证。

下载 v0.1.4 · 完整更新记录

English homepage / 英文首页: https://github.com/slmonker/selflift-Avatar

中文说明: https://github.com/slmonker/selflift-Avatar/blob/main/README_CN.md

Previous releases and both v0.1.2 demo videos are retained.

selflift-Avatar v0.1.3-experimental

selflift-Avatar v0.1.3-experimental

English

Upgrade from v0.1.2: automatic-only count/direction becomes selectable auto/manual count and auto/width/height direction; console-only diagnostics gain an in-node read-only status panel. CPU regression coverage increases from 34 to 47 tests.

• Add auto/manual spatial tiling controls to the H3 sampler.
• Manual tile dropdown: 2, 4, 6, 8 (default 2). Auto mode retains the original 1–8 search, including odd counts.
• Direction: auto / width / height. Turn highres_tiling off for full-frame sampling without the tiling planner.
• A read-only node panel shows current settings, the actual plan at high-resolution preparation, effective count, ranges/overlap, and estimated available/minimum workspace. Estimates are not measured VRAM peaks or fixed GPU requirements.
• Manual mode does not silently increase the requested count and has no automatic OOM retry. Tiny dimensions can reduce the effective count.
• Existing audio-only masked tiling restriction remains: video mask all 1, audio mask all 0. Temporal upscaler windows are unchanged.
• 47 CPU tests passed, plus frontend callback tests with a stub DOM. No comprehensive browser/model/lip-sync benchmark is claimed.

Install / update

Download selflift-Avatar-v0.1.3-experimental.zip, back up the old folder outside ComfyUI/custom_nodes, and install the new selflift-Avatar folder. Git users can run git pull --ff-only. Restart the ComfyUI backend and hard-refresh the browser. If an old node does not show new controls, recreate that H3 sampler node. Re-select a dropdown value if an older workflow stored 1 or an odd count.

English homepage: https://github.com/slmonker/selflift-Avatar
Chinese documentation: https://github.com/slmonker/selflift-Avatar/blob/main/README_CN.md
Existing videos remain on v0.1.2: https://github.com/slmonker/selflift-Avatar/releases/tag/v0.1.2-experimental

中文

相对 v0.1.2：从自动分块升级到自动/手动块数与方向选择；从主要看控制台升级为节点内状态显示；CPU 回归测试由 34 项增加到 47 项。

• 新增自动/手动空间分块模式、2/4/6/8 偶数下拉菜单，以及自动/宽度/高度方向选择。
• 节点显示实际块数、范围、重叠和显存估算；不把块数对应成固定显卡容量。
• 手动不会自动增加块数，无 OOM 自动重试；小尺寸时有效块数可能减少。
• 保留视频全 1、音频全 0 的限制，时间上采样窗口不变。
• 47 项 CPU 测试和模拟 DOM 前端回调测试通过；未宣称全面的真实模型效果验证。
• 升级后重启后端并强制刷新浏览器，旧节点必要时重新添加。旧工作流若保存了奇数值或 1，请重新选择下拉值。
• 英文仍为首页，中文独立页面；此前两个 Demo 保持原链接。

selflift-Avatar v0.1.2-experimental

English documentation is on the repository homepage; Chinese documentation is in README_CN.md.

v0.1.2 experimental update

• Enables automatic high-resolution spatial tiling for the validated audio-only mask mode: video mask all 1, audio mask all 0.
• Every visual tile receives the complete audio latent and complete H3 audio conditioning; audio is not split at visual tile boundaries.
• Adds constant SolidMask-to-audio compatibility, native H3 audio conditioning, H3 audio scale handling, and 34 CPU regression tests.
• The planner chooses the spatial direction and tile count from available workspace; 1–8 tiles are possible.
• Partial/soft video masks, partial/soft audio masks, dynamic denoise masks, and ControlNet with H3 tiling remain unsupported.

Demo

Two maintainer-provided H3 demo videos are attached below. They are examples of workflows, not universal lip-sync benchmarks.

• _00011-audio.mp4: 13.67 seconds, 1088x1920, 24fps.
• Selflift Avatar demo.mp4: 10.05 seconds, 1906x1080, 30fps.

Assets

• Source ZIP and both MP4 files are attached to this release.

selflift-Avatar v0.1.1-experimental

本版修正 H3 音频条件输入路径，保留原版节点 ID 和兼容接口。

更新

• 将音频遮罩传入 H3 原生条件构建和 KSamplerX0Inpaint 输入侧注入。
• 高清恢复采样状态与干净 inpaint anchor 分离，避免噪声污染已知音频。
• 保留双流遮罩、多通道遮罩及音频 SolidMask 兼容处理。

验证

• 28 项 CPU 测试通过，覆盖真实 Euler / 原生 inpaint 调用链、H3 音频尺度及两阶段条件标签。
• 维护者反馈当前工作流试用可用；尚无系统性口型准确率、多模型或多音频评估。

安装与更新

下载 ZIP，把 selflift-Avatar 文件夹放入 ComfyUI/custom_nodes；更新前备份旧版本，更新后重启 ComfyUI 后端。Git 安装可执行 git pull --ff-only。
带遮罩时保持 highres_tiling=false，使用 Euler。动态 denoise_mask_function 仍不支持。

v0.1.0-experimental 保留供回退。上游来源和许可情况见 README，未新增开源许可证。

selflift-Avatar v0.1.0-experimental

独立实验版本，原版节点可并存。

• H3 NestedTensor 音视频遮罩、多通道视频遮罩和音频保留。
• 兼容通过 SetLatentNoiseMask 附加到音频的恒定图像尺寸 SolidMask。
• 修正 H3 音频约束的采样尺度转换；修复带 mask 的纯像素 anchor 分支。
• 26 项 CPU 测试通过；尚未进行真实 H3 大模型视频生成验证。

安装：下载 ZIP，将其中的 selflift-Avatar 文件夹放入 ComfyUI/custom_nodes，重启后搜索 selflift-Avatar。
带遮罩时关闭 highres_tiling。只支持静态输入遮罩，不执行动态 denoise_mask_function。

保留上游来源说明；本次上传未添加新的开源许可证。