Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
121 commits
Select commit Hold shift + click to select a range
3496dbc
chore: open the next dev cycle at 3.3.0-dev
wochenlong Oct 12, 2026
9f6ae45
docs: record the natural-language tagging deferral and Ideogram 4 sta…
wochenlong Oct 12, 2026
ea9ce5d
fix(musubi): normalize legacy list-form extra_special_tokens in token…
wochenlong Oct 12, 2026
9ccb9c0
fix(anima): reject Anima 2.9B on the Kohya engine before launch (#408)
wochenlong Oct 12, 2026
5cb5705
fix(diffsynth): content-based runtime fingerprint with legacy migrati…
wochenlong Oct 12, 2026
9c0ba26
docs: changelog entries for the first batch of recent bug fixes
wochenlong Oct 12, 2026
6146abc
Merge pull request #433 from wochenlong/fix/recent-bugs-1
wochenlong Oct 12, 2026
15b15ce
fix(diffsynth): do not resolve the leaf symlink before the weight suf…
wochenlong Oct 12, 2026
3fd09ad
fix(api): unique autosave snapshot name for single submissions (#334)
wochenlong Oct 12, 2026
747e7f1
fix(anima-fast): warn when batch size will make bucket sampling drop …
wochenlong Oct 12, 2026
55474c8
Merge pull request #434 from wochenlong/fix/recent-bugs-2
wochenlong Oct 12, 2026
8b06621
feat(portable): Windows launcher NextTrainer.exe for the 2026 package…
wochenlong Oct 12, 2026
08015fc
Merge pull request #436 from wochenlong/feat/exe-launcher
wochenlong Oct 12, 2026
d6ccdb1
feat(portable): ship the launcher in every 2026 package flavor (#435)
wochenlong Oct 12, 2026
2450f1e
Merge pull request #437 from wochenlong/feat/launcher-all-packages
wochenlong Oct 12, 2026
d05b969
docs: note the main backport and the image rebuild steps for the fing…
wochenlong Oct 12, 2026
6ccde87
fix(dataset): load managed datasets when the picker is opened through…
wochenlong Oct 12, 2026
160a3f9
Merge pull request #441 from wochenlong/fix/dataset-picker-open
wochenlong Oct 12, 2026
e990415
feat(ui): show engine install state in the training engine dropdown (…
wochenlong Oct 12, 2026
fb4c9de
Merge pull request #444 from wochenlong/feat/engine-state-labels
wochenlong Oct 12, 2026
61ee899
feat(ai-toolkit): add SD 1.5 and Ideogram 4 LoRA with Comfy single-fi…
wochenlong Oct 13, 2026
1123a42
Merge pull request #447 from wochenlong/feat/446-ai-toolkit-sd15-ideo…
wochenlong Oct 13, 2026
78d8a84
fix(musubi): guard the fp8 pair on an FP8 base and settle steps vs ep…
wochenlong Oct 13, 2026
d5a4749
Merge pull request #449 from wochenlong/fix/musubi-param-guards
wochenlong Oct 13, 2026
c9cc0b8
feat(ai-toolkit): load Ideogram 4's Qwen3-VL text encoder from the Co…
wochenlong Oct 13, 2026
65e642f
Merge pull request #451 from wochenlong/feat/ai-toolkit-ideogram4-com…
wochenlong Oct 13, 2026
7dbde9c
feat(ai-toolkit): optional Ideogram 4 unconditional-sampling LoRA (#446)
wochenlong Oct 13, 2026
3a1bf5e
Merge pull request #453 from wochenlong/feat/ai-toolkit-uncond-lora
wochenlong Oct 13, 2026
caf7a2f
chore(ai-toolkit): field-run cleanups (wait_for_process, dead config …
wochenlong Oct 13, 2026
022c9df
Merge pull request #456 from wochenlong/chore/ai-toolkit-field-cleanup
wochenlong Oct 13, 2026
f9db799
feat(anima): expose DoRA as a first-class lora_type (#30)
IryNeko Oct 13, 2026
503087a
docs(anima): document DoRA adapter type vs LoKr dora_wd (#30)
IryNeko Oct 13, 2026
7b483f3
fix(diffsynth): recover ready state when install_state.json is missing
wochenlong Oct 14, 2026
ee71918
Merge pull request #460 from wochenlong/fix/459-diffsynth-missing-state
wochenlong Oct 14, 2026
58b9620
feat(diffsynth): surgical repair instead of default full wipe
wochenlong Oct 14, 2026
1bdf452
test(frontend): mock enginesApi.reinstall in EnginesSettingsPage
wochenlong Oct 14, 2026
3b470b2
Merge pull request #463 from wochenlong/feat/462-diffsynth-surgical-r…
wochenlong Oct 14, 2026
866c800
fix(anima): run bf16 DoRA modulation in fp32 (#30)
IryNeko Oct 14, 2026
b903721
Clarify DoRA training type in documentation
IryNeko Oct 14, 2026
b32a2d9
Merge pull request #465 from wochenlong/feat/30-anima-dora-adapter
IryNeko Oct 14, 2026
a8945a7
Improve training workbench model-first UX and engine install gates.
wochenlong Oct 14, 2026
3623b72
Polish settings/network UX and simplify task/tagger defaults.
wochenlong Oct 14, 2026
72ca0af
fix: keep network retry usable and avoid pinning auto-selected engines.
wochenlong Oct 14, 2026
ac0dc18
Merge pull request #468 from wochenlong/feat/frontend-ux-network-and-…
wochenlong Oct 14, 2026
c9f0d41
fix(download-sources): default channels to "auto" so engine installer…
IryNeko Oct 14, 2026
fdc263a
feat(download-sources): backend connectivity probe API + accurate ins…
IryNeko Oct 14, 2026
ae31e16
feat(download-sources): sidebar source chips with per-channel connect…
IryNeko Oct 14, 2026
b602948
feat(settings): inline download-source panel with draft checks and pr…
IryNeko Oct 14, 2026
9cffa73
feat(download-sources): persist source settings on the backend (#310)
IryNeko Oct 15, 2026
4198912
feat(frontend): hydrate download-source prefs from the server (#310)
IryNeko Oct 15, 2026
184bbf1
feat: integrate native MOSS-Image LoRA engine
waw1w1 Oct 16, 2026
55ebaeb
Merge pull request #5 from waw1w1/feat/moss-anima-lora-integration
waw1w1 Oct 17, 2026
e24fa0b
通过[https://github.com/sorryhyun/anima_lora](https://github.com/sorryh…
waw1w1 Oct 17, 2026
5dbd8e9
通过[https://github.com/sorryhyun/anima_lora](https://github.com/sorryh…
waw1w1 Oct 17, 2026
0e23a09
feat(moss): adapt ComfyUI single-file model inputs for #474
wochenlong Oct 17, 2026
48e8808
Merge pull request #467 from wochenlong/fix/310-download-sources-inline
IryNeko Oct 17, 2026
9ae496c
merge(dev): integrate #310 download-sources; keep MOSS engine hooks
wochenlong Oct 18, 2026
8140f6f
docs(moss): add native and Comfy training guide
wochenlong Oct 18, 2026
e9de3e0
Merge pull request #475 from wochenlong/feat/474-moss-comfy-format
wochenlong Oct 18, 2026
9508fe1
fix: preserve user-owned training runtimes during updates
wochenlong Oct 18, 2026
a82ed1d
Merge pull request #476 from wochenlong/codex/preserve-kohya
wochenlong Oct 18, 2026
0f6226a
feat(moss): remap Comfy-Org fused qkv DiT and polish training UX
wochenlong Oct 18, 2026
3b0a3a2
Merge pull request #479 from wochenlong/feat/474-moss-comfy-qkv-ui
wochenlong Oct 18, 2026
b4d6181
fix: preserve kohya runtime during release updates
wochenlong Oct 18, 2026
222b965
Merge pull request #481 from wochenlong/codex/preserve-kohya-release
wochenlong Oct 18, 2026
0391de0
feat(moss): prefer Kohya MOSS-Image LoRA (#483)
wochenlong Oct 18, 2026
90c4e0e
docs(readme): mention MOSS-Image and Kohya as default engine
wochenlong Oct 18, 2026
55a5a80
fix(ui): show Repair CTA when Kohya runtime is broken (#490)
wochenlong Oct 18, 2026
b999995
docs: showcase Kohaku MOSS style LoRA results with attribution
wochenlong Oct 18, 2026
0986c75
docs: standardize showcase attribution as KohakuBlueleaf
wochenlong Oct 18, 2026
b5da8d8
fix: hide unavailable plugin marketplace
wochenlong Oct 18, 2026
2359632
Merge pull request #493 from wochenlong/codex/hide-marketplace-dev
wochenlong Oct 18, 2026
a4fa321
docs: consolidate training tutorials in README
wochenlong Oct 18, 2026
031c153
docs: consolidate training tutorials in README
wochenlong Oct 18, 2026
79793eb
docs: refresh multi-engine hero and remove unreadable homepage screen…
wochenlong Oct 18, 2026
f29647e
docs: restore mascot cover with MOSS label and place multi-engine ill…
wochenlong Oct 18, 2026
cf49d57
docs: place multi-engine illustration in training engines section
wochenlong Oct 18, 2026
4653b1e
docs: add pink engine and model illustrations with compatibility matrix
wochenlong Oct 18, 2026
ecd051a
feat(kohya): accept legacy main environment as fallback runtime
IryNeko Oct 20, 2026
cd07384
feat(frontend): label kohya main-environment compat mode
IryNeko Oct 20, 2026
30a4f29
fix(frontend): use EngineStatus.message in KohyaGatePage test fixture
IryNeko Oct 20, 2026
87a98b7
feat(frontend): clarify uninstall scope in kohya main-env mode
IryNeko Oct 20, 2026
af6042f
fix(kohya): unpin tokenizers from the unresolvable 0.22.2
IryNeko Oct 20, 2026
cfc672e
merge: fix kohya tokenizers pin into main-env fallback branch
IryNeko Oct 20, 2026
124aa73
test(kohya): verify dependency graph resolves
wochenlong Oct 20, 2026
c7f198a
Merge pull request #503 from wochenlong/fix/kohya-tokenizers-pin
wochenlong Oct 20, 2026
f609b3a
feat(kohya): non-blocking pin drift detection with incremental upgrade
IryNeko Oct 20, 2026
0d4f933
feat(kohya): gate MOSS submissions on a transformers 5.x stack
IryNeko Oct 20, 2026
733c32f
feat(frontend): surface kohya pin drift with version detail and upgra…
IryNeko Oct 20, 2026
dc10821
fix(kohya): harden fallback audit execution and preflight (PR #502 re…
IryNeko Oct 20, 2026
d53ca71
fix(frontend): keep the engine state badge chain contiguous (PR #502 …
IryNeko Oct 20, 2026
1287ddb
feat(kohya): record main-env upgrades in a patch file the launcher ho…
IryNeko Oct 20, 2026
b9ea0f9
fix(kohya): match the training launcher's user-site policy in main-en…
IryNeko Oct 20, 2026
3cb0d18
Merge pull request #502 from wochenlong/feat/kohya-main-env-fallback
IryNeko Oct 20, 2026
eb3aa90
feat(user-data): tolerate unknown settings fields on read
IryNeko Oct 21, 2026
95d5738
fix(diffsynth): make conversion cache dir configurable (#391)
IryNeko Oct 21, 2026
9bd4594
feat(frontend): expose diffsynth cache dir in training and engine set…
IryNeko Oct 21, 2026
0b2d18d
fix(user-data): keep record allow-lists strict, permit partial select…
IryNeko Oct 21, 2026
e131817
fix(diffsynth): treat configured cache path as root for models and en…
IryNeko Oct 21, 2026
fe3e480
fix(startup): probe ports with SO_REUSEADDR so TIME_WAIT does not blo…
IryNeko Oct 21, 2026
bb509d8
feat: complete optional KohakuMoss engine integration on dev
waw1w1 Oct 22, 2026
10d3944
fix(kohaku): align runtime dependencies, config and task lifecycle
waw1w1 Oct 23, 2026
9dd3a0c
fix(kohaku): complete engine lifecycle and isolate Comfy cache
Oct 23, 2026
d473d74
chore(frontend): rebuild dist for MOSS-Image and dev UI updates
wochenlong Oct 25, 2026
7c71120
Merge pull request #508 from wochenlong/chore/rebuild-dist-dev
wochenlong Oct 25, 2026
14e6067
chore: bump VERSION to 3.3.1-dev
wochenlong Oct 25, 2026
8ce49bf
Merge pull request #511 from wochenlong/chore/bump-331-dev-branch
wochenlong Oct 25, 2026
96275e0
fix(portable): bundle uv in requirements and resolve uv executable fo…
wochenlong Oct 26, 2026
e20c77f
Merge pull request #513 from wochenlong/fix/portable-uv-kohya-install…
wochenlong Oct 26, 2026
56f3334
merge: resolve dev conflicts, wire uv resolution, and rebuild dist fo…
wochenlong Oct 26, 2026
9883f75
fix(kohaku): pin kohakuvault<1.0, probe lycoris kernel fallback, and …
wochenlong Oct 26, 2026
6a899cb
docs: add KohakuMossTrainer to READMEs, MOSS guide, and credits
wochenlong Oct 27, 2026
43d3c81
Merge pull request #505 from wochenlong/fix/391-diffsynth-cache-dir
wochenlong Oct 27, 2026
53d4e6b
chore(frontend): rebuild dist for #505 diffsynth cache dir
wochenlong Oct 27, 2026
e6fc9c5
fix(kohya): restore MOSS cache strategy contract
wochenlong Oct 27, 2026
49bc513
docs: credit waw1w1 as a community contributor
wochenlong Oct 28, 2026
44bbd4e
docs: describe niangao2331 by his team role
wochenlong Oct 28, 2026
9ca2de9
docs: drop the unfilled 3.0.0 beta thanks section
wochenlong Oct 28, 2026
0e9b0cd
docs: clarify project origins, independent maintenance and attribution
wochenlong Nov 3, 2026
fd64823
docs: move project origins and license context from homepage to credits
wochenlong Nov 3, 2026
549c898
merge: reconcile main with dev for KohakuMossTrainer review
wochenlong Nov 3, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions .gitattributes
Original file line number Diff line number Diff line change
Expand Up @@ -16,3 +16,5 @@ frontend/index.html text eol=lf

# WSL/bash scripts stay LF (see agent-assets/.gitattributes same rule)
*.sh text eol=lf

mikazuki/engines/moss/patches/*.patch text eol=lf -whitespace
20 changes: 19 additions & 1 deletion .gitignore
Original file line number Diff line number Diff line change
Expand Up @@ -145,15 +145,33 @@ extensions/diffsynth/
/extensions/moss/
/extensions/.moss.lock

# Isolated KohakuMoss runtime and cross-process lock
/extensions/kohaku-moss/
/extensions/.kohaku-moss.lock

# Local project memory
/MEMORY.md

# Local Qwen-Image-2.1 processor assets
/qwen-image-2.1-processor/
/tokenizer-cache/

# Generated frontend assets (existing tracked release assets remain tracked)
# Generated frontend assets
/frontend/dist/

# User-owned settings, credentials, presets and task archives.
/user_data/

# launcher (.NET) build artifacts
launcher/bin/
launcher/obj/

# Local verification output
/junit*.xml
/pytest*.xml
/frontend/.vite/
/frontend/.vitest/
/frontend/coverage/

# Local MOSS integration notes
/mikazuki/engines/moss/FIELD_NOTES.md
49 changes: 31 additions & 18 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -3,33 +3,46 @@
本文件记录 **wochenlong/lora-scripts-next**(产品名 **Next Trainer**)面向镜像与 AutoDL 的发行说明;上游 kohya-ss/sd-scripts 的变更请见其仓库。

---
## 未发布
## 未发布(dev)

### 新增

- **MOSS-Image LoRA(#480 / #483)**:接入 MOSS-Image 训练,Kohya 为推荐及默认引擎,内置原生 trainer、缓存与 ConvRot INT8 支持;`MOSS (anima_lora)` 保留为次选独立运行时。支持原生 Diffusers 与 Comfy 单文件输入,详见 [MOSS-Image 训练指南](docs/MOSS.md)。

### 修复

- **DiffSynth 指纹不再包含 mtime(#410)**:目录拷贝、云镜像打包、解压或启动脚本触碰文件都会改写时间戳,导致可用环境被误报「环境或源码已变化」并要求重装(AutoDL 定制镜像上会反复复发)。指纹改为**内容哈希**(`content-v2`)并持久化方案标记:时间戳变化不再误报,旧状态**一次性迁移**(不要求重装),内容真的被改动、训练入口/独立 Python 缺失时仍会正确报错。
- **预装 DiffSynth 的定制镜像需要同步这两个文件后重建镜像才生效**:`mikazuki/engines/diffsynth/extension_state.py`、`mikazuki/engines/diffsynth/install_worker.py`;只更新 `install_state.json` 不算修复(下次时间戳再变还会复发)。
(下一个开发周期:AI Toolkit 模型格式与训练体验,详见 docs/roadmap.md)

- **Krea2 tokenizer 的 `extra_special_tokens` 为 list 时加载失败(#416)**:探测/下载/启动前会把旧的 list 形态规范化为 transformers 期望的 mapping(`{"additional_special_tokens": [...]}`),手动放置的 tokenizer 同样会被治好;已是 mapping 的配置不动,坏 JSON 安全跳过。
- **Kohya 加载 Anima 2.9B 前明确拦截(#408)**:`validate_model` 只读 safetensors 头部统计 DiT block 数,超过 28 即拒绝并提示「请改用 Anima Fast(设置 → 训练引擎 安装),或改用 Anima 2B」,不再等到加载时抛一串 `blocks.28.*` 权重名;2B 与 Anima Fast 路径不受影响,坏文件/无法识别不会被误判。
- **标签编辑的「选择已有数据集」弹窗为空(#440)**:`ManagedDatasetPicker` 现在会监听受控的 `open` 属性(标签编辑隐藏内部按钮、直接控制弹窗),弹窗打开时执行与内部按钮相同的初始化并拉取数据集列表;关闭后重开正常刷新,`browse()` 的 generation 机制继续防止并发响应覆盖
- **训练引擎下拉显示安装状态(#443)**:每个引擎带状态后缀(已就绪 / 未安装 / 需修复 / 安装中 / 已停用 / 未验证),模型不支持的引擎继续灰显并保留「该模型暂不支持」提示,另加一条可操作提示列出「当前模型支持但未安装/待修复」的引擎;引擎状态接口失败时静默降级为纯引擎名
- **AI Toolkit 新增 SD 1.5 与 Ideogram 4 LoRA(#446)**:两个模型都直接读取 Comfy 格式单文件,无需用户转换 diffusers 目录
- **SD 1.5(arch `sd1`)**:`single_file` 模式加载 Comfy/LDM 全量 checkpoint(`model.diffusion_model.*` / `cond_stage_model.*` / `first_stage_model.*`),由驱动层绑定本地 diffusers 配置目录(仅 config 与 tokenizer,不含权重,默认 `sd-models/sd15/config`,可由「训练用模型」面板从 HF / ModelScope 下载),离线启动不访问 Hub
- **Ideogram 4(arch `ideogram4`)**:`comfyui_files` 模式直接使用 Comfy-Org 单文件(条件 DiT + Flux2 VAE,与 Musubi 同一套下载源),驱动层在内存中整理 FP8 `weight_scale` 形状并剔除 `comfy_quant` 标记后交给上游扩展,不写 diffusers 目录、不产生缓存副本;文本编码器使用本地 Qwen3-VL-8B BF16 目录。沿用**非商用**许可提示,整合包不预置权重
- **底模已是 FP8 时自动关闭 fp8_base/fp8_scaled(#417)**:只读 safetensors 头部探测 DiT dtype,已是 FP8 权重时自动关闭这两项并提示;非 FP8 底模的成对约束不变
- **max_train_steps 与 max_train_epochs 同时填写(#418)**:按**步数优先**只保留 max_train_steps 并明确提示(上游以 epoch 优先,会让「steps=100」实跑到按 epoch 推导的步数),表单文案同步写清
- **AI Toolkit 的 Ideogram 4 文本编码器改为 Comfy FP8 单文件(#446 跟进)**:驱动层把 tokenizer/config 仍指向本地目录、权重走 Comfy 单文件(上游 mixin 原生支持 `comfy_quant` 单文件加载),资源清单改为「TE 单文件 + config/tokenizer 目录(约几 MB)」,用户不再需要 ~16 GB 的 bf16 目录;`model_directory` 分支与 `config/model_assets.toml` 覆盖仍可回退到 bf16 目录

### 修复(近一个月 issue 第三批)

- **标签编辑的「选择已有数据集」弹窗为空(#440)**:`ManagedDatasetPicker` 现在会监听受控的 `open` 属性(标签编辑隐藏内部按钮、直接控制弹窗),弹窗打开时执行与内部按钮相同的初始化并拉取数据集列表;关闭后重开正常刷新,`browse()` 的 generation 机制继续防止并发响应覆盖

### 修复(近一个月 issue 第四批)

- **底模已是 FP8 时自动关闭 fp8_base/fp8_scaled(#417)**:只读 safetensors 头部探测 DiT 的 dtype,已是 FP8 权重时自动关闭这两项并提示(两个都关本就是合法状态),不再等到上游拒绝;非 FP8 底模的成对约束不变
- **max_train_steps 与 max_train_epochs 同时填写(#418)**:musubi 侧以 epoch 优先,会让「steps=100」实跑到按 epoch 推导的步数;现在按**步数优先**只保留 max_train_steps 并明确提示,表单两处文案也写清「以步数为准」

### 新增
- **Ideogram 4(AI Toolkit)无条件采样 LoRA 设为可选(#446 跟进)**:新增可选资源 `unconditional_lora`(第三方产物 `ostris/ideogram_4_unconditional_lora`,103.9 MB)与表单可选字段「无条件 LoRA 路径」——填了走非对称 CFG 预览,不填走单 DiT CFG;顺带修掉 `text_encoder_tokenizer_path` 被字段白名单误报「未支持」的噪音警告
- **DiffSynth 软链权重误报(#390)**:路径规范化不再跟随叶子软链,AutoDL 那类「`*.safetensors` 名字指向无后缀实体」的挂载不再被后缀校验误拒
- **AI Toolkit 的 Ideogram 4 文本编码器改为 Comfy FP8 单文件(#446 跟进)**:不再要求用户准备 ~16 GB 的 bf16 Qwen3-VL-8B 目录——驱动层把 tokenizer/config 仍指向本地目录、权重改走 Comfy 单文件(上游 mixin 原生支持 `comfy_quant` 单文件加载),资源清单同步改为「TE 单文件 + config/tokenizer 目录(约几 MB)」

- **Windows 启动器 `NextTrainer.exe`(#435)**:放在整合包根目录(与 `启动.bat` 并存)。双击启动等价于 `启动.bat`,但接管进程树:关窗 / `Q` 真正停掉服务,端口占用或启动失败给中文提示,`U` 一键更新(先停服 → 运行更新脚本 → `R` 重启)。C# NativeAOT 编译(约 5 MB,无 .NET 运行时依赖)。构建脚本 `build-scripts/06-build-launcher.ps1`,已接入 2026 全量 / musubi-only / kohya-only / resume 的打包流程

### 修复(近一个月 issue 第二批)

- **DiffSynth 软链权重误报(#390)**:路径规范化不再跟随叶子软链,AutoDL 那类「`*.safetensors` 名字指向无后缀实体」的挂载不再被后缀校验拒绝
- **单条提交的 autosave 快照重名(#334)**:`/api/run` 的配置快照名加入随机后缀,同一秒内连续提交不再互相覆盖(批量提交此前已修)
- **anima-fast 小数据集丢图提示(#338)**:预检在「图片数偏少 + train_batch_size>1」时预警上游 bucket 采样会整桶丢弃图片(启动日志里只有一条 WARNING 就会被忽略掉),并给出 batch_size=1 / num_repeats / bucket_reso_steps / validation_split_num 的处理建议
- **Anima TE 缓存固化 caption_dropout_rate(#333)**:文档写明「改动该值需删除 `_anima_te.npz` 再重训」的限制([开发进度](docs/dev-progress.md));上游缓存键的一致性校验待跟进
- **anima-fast 小数据集丢图提示(#338)**:预检在「图片数偏少 + train_batch_size>1」时明确警告上游 bucket 采样会整桶丢弃图片(及 1 batch 步数远低于预期的原因),并给出 batch_size=1 / num_repeats / bucket_reso_steps / validation_split_num 的处理建议
- **Anima TE 缓存固化 caption_dropout_rate(#333)**:先在文档中写明「改动该值需删除 `_anima_te.npz` 再重训」的限制;上游缓存键的一致性校验待跟进

### 修复(近一个月 issue 第一批)

- **Krea2 tokenizer 的 `extra_special_tokens` 为 list 时加载失败(#416)**:探测/下载/启动前会把旧的 list 形态规范化为 transformers 期望的 mapping,手动放置的 tokenizer 同样会被治好
- **Kohya 加载 Anima 2.9B 前明确拦截(#408)**:只读 safetensors 头部统计 block 数,超过 28 即拒绝并提示改用 Anima Fast,不再等到加载时抛一串 `blocks.28.*` 权重名
- **DiffSynth 指纹不再含 mtime(#410)**:目录拷贝 / 云镜像 / 备份还原导致的时间戳变化不再误报环境异常;旧状态一次性迁移,不要求重装;内容改动仍能识别
- 已回移到 `main`(PR #438)。**预装 DiffSynth 的定制镜像需同步 `mikazuki/engines/diffsynth/extension_state.py` 与 `install_worker.py` 后重建镜像**;只改 `install_state.json` 不算修复(时间戳再变仍会复发)——文件到位后,现存实例在下一次状态检查时会自动迁移,无需重装

---
## v3.2.0 — 2026-10-11

### 整合包更新修复

Expand Down
12 changes: 12 additions & 0 deletions NOTICE.md
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,18 @@ Next Trainer ships this engine as an **optional plugin** (`extensions/anima_lora

The referenced repository features per-block or full-model `torch.compile`, static-shape token bucketing, and compile-friendly forward paths. On consumer GPUs this yields substantially faster step times than the default Kohya/sd-scripts path at the cost of higher VRAM and a separate Python/CUDA runtime. See **Performance** in `docs/anima-fast.md` for measured comparisons in this repo.

### KohakuMossTrainer(实验性高性能可选引擎)

The optional **KohakuMossTrainer** engine (`model_train_type: kohaku-moss-image-lora`) integrates, through an isolated Python environment and the upstream `kogine` CLI:

- `KohakuBlueleaf/KohakuMossTrainer`: https://github.com/KohakuBlueleaf/KohakuMossTrainer
- pinned integration commit: `5bb1a81ed030cc87c540d5e02fe60aab887199b4`
- integration coordination: https://github.com/KohakuBlueleaf/KohakuMossTrainer/issues/1

KohakuMossTrainer is distributed under **KohakuMoss License 1.0**, described upstream as Apache License 2.0 with an additional naming requirement. Next Trainer exposes the upstream name prominently in the engine name and UI, installs the pinned source into `extensions/kohaku-moss/`, and does not create a MOSS/Kohya fork. Please retain the upstream license and notices when redistributing an installed runtime.

We thank **Kohaku-Blueleaf(Kohaku / 龙娘)** for KohakuMossTrainer and the broader LyCORIS ecosystem. The integration directly depends on `lycoris-lora>=4.1.2`.

### LyCORIS

LoKr, LoHa, and other advanced LoRA variants are powered by:
Expand Down
Loading
Loading