v0.1.11 — Transformers 4.57.1 compatibility

修复

支持 Transformers 4.57.1，移除节点对 5.19 的最低版本要求。修复 Qwen 词级对齐在 4.57.1 中因掩码参数差异导致的真实前向错误，并保留现有 5.19 环境兼容性。

推荐组合：Transformers 4.57.1、Tokenizers 0.22.2、Hugging Face Hub 0.36.2。整合包 requirements 默认采用该组合；程序自动更新保留已安装的 Python 包，不会自动运行 pip 或更换运行环境。模型无需重新下载。

验证：4.57.1 与 5.19 各 479 项 CPU 测试通过、1 项跳过，包括新增的真实 Qwen 前向、缓存、因果遮罩与 tokenizer 测试。4.57.1 另完成 10 次实际浏览器画布运行，覆盖语音生成、参考音色、改稿、局部编辑、流式、Codec、Qwen 转写、R2T2 与词级对齐；转写内容、11 字时间戳及复杂场景两段台词均核对通过。

Fixes

Support Transformers 4.57.1 without requiring a 5.19 upgrade. Correct the causal-mask API call used by Qwen word alignment on 4.57.1, while retaining compatibility with existing 5.19 environments.

Recommended combination: Transformers 4.57.1, Tokenizers 0.22.2, Hugging Face Hub 0.36.2. Portable requirements default to these versions. Code updates preserve installed Python packages and do not run pip or switch environments automatically. Model weights are unchanged.

Validation: 479 CPU tests passed and 1 skipped on each of 4.57.1 and 5.19, including real Qwen forward/cache, causality and tokenizer checks. On 4.57.1, 10 actual browser canvas runs covered speech, references, revisions, infill, streaming, codec, Qwen ASR, R2T2 and word alignment. Transcription content, the 11-character alignment and both lines of the reference scene revision were checked.

Downloads / 下载

• ComfyUI-Drama3-T8-0.1.11.zip: independent native nodes and 15 canvas workflows.
• Drama3-T8-Portable-0.1.11-code.zip: cumulative program update for an existing complete Windows environment.
• portable-update.json, SHA256SUMS.txt: automatic-update manifest and checksums.

These assets contain no model weights or Python runtime. The verified R2T2 native runtime remains pinned to its unchanged v0.1.7 asset.

v0.1.10 — Scene revision and speaker casting fixes

修复

修复带参考音色的多角色场景改稿在首句提前结束、遗漏后续台词的问题。全局导演说明使用正确的提示范围，提前结束时进行每行最多一次的换行恢复，随后由模型生成真实分隔符与后续音频。空段、多段和帧预算不足仍会报错。

参考场景的命名角色现在使用稳定的 speaker 编号，连续改稿沿用映射，避免把角色名读进音频。原始参考、完整工程历史与旧音频均保留；模型权重无需更新。

验证：472 项 CPU 测试通过、1 项跳过；浏览器画布点击运行完成 9 个音频成功用例与 1 个预期预算拒绝，17 段英文/中文音频经本地 R2T2 转写核对台词一致。覆盖原先两个失败种子和连续改稿。该结果仅代表所测用例。

Fixes

Reference-conditioned scene revisions could stop after the first line and omit later dialogue. Global director notes now use the correct prompt scope. A bounded, once-per-line newline recovery lets the model produce the actual separator and subsequent audio. Empty, extra and frame-budget-limited incomplete segments remain errors.

Named roles in reference scenes now use stable speaker IDs across revisions, preventing role names from leaking into spoken dialogue. Original references, complete session history and archived audio are retained. Model weights are unchanged.

Validation: 472 CPU tests passed, 1 skipped; 9 successful audio cases and 1 expected budget rejection through actual browser canvas Run clicks. Local R2T2 transcription matched all 17 English/Chinese dialogue segments, including both previously failing seeds and repeated revisions. Results apply to these fixtures.

Downloads / 下载

• ComfyUI-Drama3-T8-0.1.10.zip: independent native nodes and 15 canvas workflows.
• Drama3-T8-Portable-0.1.10-code.zip: code update for an existing complete Windows bundle; close the studio before updating.
• portable-update.json and SHA256SUMS.txt: verified automatic updates and checksums.

No model weights or Python runtime are included. Existing model paths, projects, voice libraries and outputs are preserved by the portable updater. The unchanged R2T2 native runtime remains pinned to its verified v0.1.7 asset.

v0.1.9 — Scene context and workspace fixes

中文

修复场景任务丢失会话方向/参考/历史、生成与对齐请求乱序覆盖当前选择、导入/上传登记失败遗留文件。拒绝空音色存档及非整数分支参数。保留15张纯画布示例。

验证：378项CPU检查通过，1项Windows权限跳过；15项新浏览器断言通过。实际画布默认场景、带方向的场景改稿、参考克隆成功。带参考的复杂场景改稿两次缺少分段符，被运行时拒绝，已记录为已知限制。

附件不含模型和Python，已有整合包支持程序自动更新。模型、依赖及原生库不变，R2T2安装器沿用已校验的v0.1.7资产。

English

Preserves scene session direction, references and history. Fixes stale generation/alignment responses overriding current selections and unregistered files left by failed imports/uploads. Rejects empty voice references and noninteger branch counts. Includes 15 canvas-only examples.

Validation: 378 CPU checks passed (one Windows privilege skip), plus 15 new browser assertions. Default scenes, directed scene revisions and reference cloning passed actual canvas runs. Two reference-conditioned scene-revision variants omitted separators and were correctly rejected; this limitation is documented.

Assets contain no model weights or Python runtime. Existing complete bundles support automatic code updates. Models, dependencies and the verified v0.1.7 R2T2 runtime are unchanged.

v0.1.8 — Canvas workflows and exact seeds

中文

提供15个经过实际浏览器画布运行的工作流，覆盖17个原生节点；移除API示例。修复长文仅输出末段、转写/对齐预览不可见、工程说明遮挡和浏览器大整数种子精度问题。大种子使用 seed_text。

验证：23个实际画布场景；346项CPU检查通过、1项因Windows权限跳过。测试覆盖文字显示、音频播放、实时分包、局部编辑保留、流式续接和工程回读。

附件不含模型和Python。完整环境由作者另行提供，已有整合包支持程序自动更新。R2T2原生库内容不变，安装器沿用v0.1.7的已校验下载。

English

Includes 15 browser-tested canvas workflows covering all 17 native nodes. Removes API examples and fixes incomplete long-text output, missing transcription/alignment previews, overlapping session instructions and browser seed precision. Use seed_text for exact uint64 seeds.

Validated with 23 actual canvas runs and 346 passing CPU checks (one Windows privilege skip). Release assets contain no models or Python runtime. Existing complete bundles support automatic code updates; the unchanged R2T2 runtime remains pinned to its verified v0.1.7 asset.

v0.1.7 · Native nodes and portable auto-updates

简体中文

首次公开发布：17 个 DRAMA-3 原生节点、白粉本地工作台，以及整合包自动更新。

• GitHub Release 不含模型和 Python；完整运行环境由作者单独提供。
• 原生节点包：解压到 ComfyUI/custom_nodes/，用宿主 Python 安装 requirements.txt 并运行 install.py，然后重启。安装器校验并安装单独分发的 R2T2 原生运行库，不下载模型。
• 整合包更新：关闭现有工作台，将 Drama3-T8-Portable-0.1.7-code.zip 解压到包含 Start-Drama3.cmd 的目录，适用于基于 0.1.3 或后续版本的完整包。此后启动自动检查正式版；也可关闭工作台后运行 Update-Drama3.cmd。
• 自动更新校验 SHA-256 与大小，保留模型、Python、自定义路径、工程、音色库和输出；更新失败回滚。
• 实测 Windows x64 / RTX 5090 Laptop 24 GB / 64 GB RAM / Torch 2.8 + CUDA 12.8。R2T2 其他 GPU 架构可能需要重编译原生库。

模型与辅助模型 · 中文说明

English

First public release: 17 native DRAMA-3 nodes, a white-and-pink local studio, and transactional portable code updates.

No release asset includes model weights or a Python runtime. The author distributes complete environments separately. Extract the node ZIP into ComfyUI/custom_nodes/, install dependencies using the host Python, run install.py, then restart. The installer verifies the separately distributed R2T2 native runtime and does not download models.

For a complete portable bundle based on 0.1.3 or later, close the studio and extract the code ZIP beside Start-Drama3.cmd. Future launches check stable releases automatically; Update-Drama3.cmd checks manually. Updates verify SHA-256 and sizes, preserve models/runtime/user data and roll back on failure.

Validation: 328 passed, 1 skipped; 43 updater and 12 native installer cases included. Exported source and actual 0.1.3 source upgrade checked. Tested on Windows x64 / RTX 5090 Laptop 24 GB / 64 GB RAM / Torch 2.8 + CUDA 12.8.

Built with Fish Audio. See repository and model cards for component licenses and custom model path settings.