MiniMax H3 Audio T8
MiniMax H3 joint AV, RAVEN streaming guards, LightX2V SLA, motion recovery, FETA, latent upscale, Hybrid, multi-keyframe, long-video, and Ref2VA nodes for ComfyUI
Nodes (200)
The H3 'memory saver' that honestly tells you it might save nothing
The one conditioning node that runs every MiniMax H3 video task
An audio QC pass that never touches the audio
Lock your voice into the latent so H3 doesn't reinvent it
Blend source and generated audio without leaving the graph
Catching the 'voice suddenly sounds far away' failure
Trim your source audio to exactly the grid H3 wants
Check the decode's VRAM bill before the VAE ever runs
Turning the joint AV latent back into frames and sound
Split the joint latent without paying for a decode
Reshape the sigma schedule without changing the step count
Spend extra steps exactly where the tail ends
One extra step at the very end of H3 sampling
Check your ClipProj setup without loading a single matrix
The deterministic half of H3's scene-understanding pipeline
A reference-semantics node that defaults to fully offline
Quarantine candidate files instead
Picking the next shot from what the background job finished
Bind a Creator plan to the Long Video background controller
Where were we? The resume node that never queues anything
A delete list that refuses to delete anything
Log what actually happened, so the plan can trust it
The shot card you can edit without rebuilding the timeline
Side-by-side A/B review without the guesswork
Pick the shot and seed you're about to burn VRAM on
'render shots 3–7, seeds 0–2, and don't touch my files'
Four quality knobs that all start at zero
'Did that line actually play here?' — finding speech boundaries locally
A dialogue mix that refuses to eat your words
From a two-column script to a dialogue plan
Rerun one line, not the whole scene
One clock for video, one for audio
Turn seconds into valid H3 frame counts (22, 124, 362…)
CFG on H3 is expensive — this guider makes you say yes first
How many forwards did that guider actually burn?
Enhance-A-Video's receipts, checked after sampling
One measurement, not fifty
Enhance-A-Video that survives segment boundaries
When Prompt Relay and Enhance-A-Video want the same attention slot
Enhance the target, leave the reference alone
Enhance-A-Video plus SageAttention without the fallback roulette
FETA plus skip-block guidance on H3, with the extra forwards audited
Temporal sharpening for H3 clips, injected straight into attention
Diagnose your H3 install before you blame the workflow
Feed BlockSwap settings to the external H3 streaming sampler
Assemble a 2–3 person cast from your character profiles
Turn reference photos into a reusable character profile
Swap the face crops into your H3 latent, audio untouched
The manual-verification gate that checks you built it the author's way
Inject the parity crop latent, keep the audio exactly as-is
Replicate the reference face-refine recipe, exactly
The reference-mechanism stitch, bit-exact outside the mask
Denoise that scales with face size — 30px faces get more help
Plan the face-repair crops before you touch the sampler
An automatic 'don't ship the worse face' check before export
The low-denoise dual-clock sampler for the face-refine pass
Paste the refined face back — with a bit-exact audit of what you touched
Label every face track with the right character — or override by hand
Build a tiny content-addressed model patch, not a whole copy
Quarantine, restore and repair your hybrid model patches safely
Your Hybrid setup is probably broken somewhere — this node finds it before you burn a render
Load a clean quality base, then bolt one small hybrid artifact onto a clone — in that order
Two pruned checkpoints walk into a render — check their SHAs before you trust the blend
Two characters, one render, zero identity guarantees — the honest multi-speaker dialogue experiment
Pin a middle keyframe, pick a spot on the timeline, and let H3 interpolate around it
Stitch the repair back onto the original — repaired pixels in, untouched sound everywhere else
Preparing the mask before the sampler
Upscale your H3 latent without wrecking the aspect ratio
Real detail on the video, native audio left untouched
The sigma plan that makes 4+4 actually work
This audit fails unless every H3 forward really used the sparse path
When SLA and KJ Sage both want the attention path, this composer picks one owner per call
4-step H3 with 85% of the attention blocks skipped — the LightX2V SLA path, properly pinned
The review gate that keeps your long video honest
Load the accepted previous segment and the exact parent identity your next render depends on
The terminal node that accepts, composes, and queues exactly one next segment
Register the prompt before you sample, so a crash or OOM doesn't nuke the whole long video run
Save a segment without touching your accepted history — the candidate half of long video
Verify every accepted segment, then stream them into one MP4 without blowing up memory
The conditioning that turns 'one more segment' into 'a continuation' — context overlap done right
The node that gives your long H3 video a memory of the last segment
Save the tail of this H3 segment, and nothing else — that's the job
Type a total duration, and this node decides every H3 segment for you
MiniMax H3 doesn't do 60-second clips, so it plans them in pieces
Sneak a tiny time-bias into H3's sigma — no extra steps, no noise, just bias
The lazy gate that skips the whole repair graph when H3 got it right
The read-only motion analyzer that decides if H3 botched the movement
A dependency-free audit that flags shaky, frozen or mushy H3 footage
Put the repaired clip back on the exact world clock — audio untouched by default
The composer that turns your stretched latent into a controlled second pass
Turn an audit's flagged ranges into a repair plan, without touching the media
Stretch the bad frames before H3 re-samples them — that's the recovery trick
Slice one repair plan into VRAM-sized windows instead of one giant pass
Assemble the repaired windows, and rerun only the one that died
Stitch each repaired face back onto the untouched original
One character, one shot, one repair job — H3's answer to multi-person face fixes
First frame, last frame, and the keyframes in between — one conditioning node
Different step counts for video and audio — with one honest caveat about cost
Reload a finished H3 latent, and let it prove nothing got swapped on you
Save a finished H3 latent as a tamper-checked file — and it won't overwrite anything
Splicing H3 Long Video segments by removing the exact 5/22/39-frame head
Proving your reloaded H3 latent is bit-for-bit the one you saved
Joining two H3 clips in latent space so you don't decode doubled frames
Checkpointing H3 mid-render at exact step boundaries, then resuming in a fresh process
The immutable JSON receipt that makes NFE resume safe to trust
Trimming an H3 render by seconds, video and audio together
The five-second check that stops an H3 render before it OOMs at step 40
The 7000-character reality check your H3 prompt needs before you queue it
One structured brief that compiles for H3, Wan 2.2, or LTX-Video
From the Studio compiler's packet to a relay timeline — without rewriting a word
Prompt enhancement with zero network by default — and guarded keys if you opt in
The node that actually installs the relay — encode the plan, patch a cloned model
One action per node — chain these to build a timed H3 scene
Segment-level relay conditioning that keeps Long Video's motion context intact
Keeping relay events from restarting at every Long Video segment
One global prompt, several timed local actions — the heart of the relay system
See every event's frame range before a single model loads
The switch that extends relay routing from video to audio — paper-faithful by default
Guess H3 packed-row counts before you commit — and don't call it a VRAM certificate
The local 8B prompt rewriter that runs on 16GB and unloads after every use
The VRAM release button for MiniMax H3's 8B prompt rewriter
Stop your prompt rewriter from quietly changing the scene
Get your <Picture 1> / <Video 1> / <Audio 1> tags right before H3 eats them
See whether your H3 Qwen prefix cache is actually saving you anything
Cache H3's expensive vision prefix so repeated references don't re-encode
A loader that refuses to load — until it's sure the machine can survive
T2VA-only, or nothing
One knob panel for the whole RAVEN streaming chain
Re-noising H3's clean output and descending again
Render the reel to MP4 without ever holding it all in RAM
Plan the whole reel before you render a single frame of it
Pick which segment you're going to redo, without touching the rest
Track two or three people through a shot with SAM3.1 — colors are not identity
Inject your drive audio into H3's denoise window, on schedule
The 'yes, I actually reviewed this' handshake in the repair chain
Lock a repair segment to an immutable manifest revision
Render the repaired timeline without ever rebuilding the original
Preview a repair candidate without committing a single byte
Build the do-over list for your long video, non-destructively
A non-generative skin finish that never touches a pixel you didn't sign off
Per-person skin masks in a two-person shot, without rerunning SAM
Finish two people's skin in a long clip without loading SAM twice
Actually look at a skin-finish candidate before you accept it
A skin mask that knows eyes from cheeks, so your finish doesn't smear
The honest way to 'finish' H3 skin without a face-fix trip
A mechanical guard so your skin finish can't flatten shadows or clip highlights
The only Skin Finish node that re-encodes your video without touching the audio
Skin finish on a long H3 clip without ever materializing all the frames
Before you bolt Sol-Attn onto H3, let this node check the patch ownership
Plan dialogue, music, ambience and SFX on one H3 sound timeline
Assemble video and audio latents for a redraw, with denoise masks H3 understands
Cut source video into the exact H3 window the VAE can encode
STG for H3, tuned for its joint AV transformer and its conflict-prone patches
Catch ambiguous multi-speaker dialogue routing before you burn a render
Make a voiceover land on the exact frame count the scene needs
Stitch rendered speech turns into one track and get the SRT for free
The node that turns 'speak this line in this voice' into H3 conditioning
Pull just the audio out of an H3 latent, no video decode
The bookend that releases your VRAM after a speech render
A seatbelt for H3 speech renders that crash or get cancelled
Commit a good H3 segment before the next render — and don't lose it when ComfyUI dies
Sew your accepted H3 segments into one clean audio track — subtitles included
Cancel, reset, or just check on a long-form speech job — without touching the graph
Render one segment at a time, resume after a crash
Emotion, pace, pitch — with the honest caveat that none of it is calibrated
Turn a script into an H3 voice performance — without guessing how long it'll take
Condition, sample, decode, release
ASR verification and exact-target trimming
A VRAM reality check before the 33B speech sampler runs — not a guarantee, a gate
Strict 24fps, 17n+5, no stretching
Noise that doesn't change the audio when you change the video canvas
Prove which H3 checkpoint and VAE you actually ran — with a SHA-256, not a filename
SPEED's plan, without the hand-waving
Progressive-resolution H3 with audio riding along
Hold the raw H3 inputs so every SPEED resolution stage can re-encode its own references
Accumulate H3 spectrum statistics across runs without hoarding latents in VRAM
Persist your accumulated H3 spectrum dataset — atomic save, no silent overwrites
Turn a hundred accumulated H3 clips into a spectrum profile — only if the fit earns it
Measure H3's own latent spectrum instead of borrowing WAN or Flux constants
Tell it what to change, not which pixels
Steal a Still From an H3 Latent Without Decoding the Whole Clip
Check Your H3 Still Contract Before You Burn a 15-Second Generate on It
Pick One Shot Out of Your Timeline and Render Just That
Plan a Whole H3 Video as Shots, Not One Giant Prompt
Sharpen H3 Footage Without Making the Motion Flicker
Lock a Music Bed Into an H3 Clip Past the Dialogue
Finish the Second Half of an H3 Sample You Started Elsewhere
Save Your Half-Finished H3 Latent — Only When You Say So
Split an H3 Sample in Half So You Can Resume It Later
A Guardrail That Stops Pass 2 From Silently Rewriting Your Audio
Tail, Bias, STG, Restart
Stitch the Two H3 Passes Together Without Breaking the Audio Clock
The 4+4 Sigma Split That Doubles as an H3 Hires Fix
One Character Sheet for Your Whole H3 Video
The One-Knob Fix for H3's Over-Smoothed Reference Textures
Delete a Voice the Way You Should — Into a Recoverable Trash Folder
Pull a Saved Voice Back Into Any Workflow
Save a Voice You Actually Want to Reuse (Explicitly)
Give H3 a Voice It Can Remember — Describe It or Clone It
A Reserve Policy That Actually Runs Before Load
MiniMax H3 Audio T8
面向 ComfyUI 的 MiniMax H3 视频与音频节点包。它保留原生 H3 的工作流接口，并在此基础上提供双时钟采样、音频控制、长视频、关键帧、脸部修复和显存诊断等能力。
当前版本：1.46.0 · 节点 200 个 · GPL-3.0-or-later
能做什么
- 文生音视频（T2VA）、图生视频、首尾帧、参考图/参考视频/参考音频和 Hybrid 工作流
- 双时钟视频/音频采样，以及音频锁定、重混、参考和最终混音
- 学习型双采样在非1.0起点把音频重建到自己的sigma时钟，再由第二阶段完成联合AV；显式锁源审计仍可选
13-latent-upscale中标准、native语音、lock_source、remix_source 0.20 和 reference_only 五份学习型二采工作流已统一到作者对齐的低清4步+高清4步，共8次联合AV前向- 同一清晰语音素材上的 native、lock_source、remix_source 0.20 与 reference_only 双采成片已由用户完整审核通过；结论限定于该素材
- 32 倍数对齐的 3D latent 放大，避免官方放大节点造成宽高偏移
- 长视频分段、断点续跑、时间线报告和音视频同步检查
- 单人/多人脸部修复、SAM3.1 追踪、多帧关键帧和实验性动态细节增强
- Motion Recovery动作过载分析、自动ABSTAIN懒旁路、局部时间超采样二采、原时钟恢复和分窗断点续跑（实验）
- 语音、对白、演绎、ADR 和生产辅助节点（实验功能会明确标注）
- 音频完整性、参考相对音色漂移与多人对白路由预检：报告首尾边界、DC、削波、持续频谱/响度漂移、参考音频绑定和歧义，遇到风险返回
ABSTAIN，不自动修改素材 - 提示词预算与角色媒体编译：统计字符/token、检查
<Picture>/<Video>/<Audio>数量、顺序与人物绑定覆盖，原提示词连首尾空白也不静默删除 - 提示词服务路由：默认本地原文直通，可显式连接 OpenAI/LM Studio/llama.cpp 或 Ollama；密钥仅从环境变量读取，Ollama默认请求后卸载（实验）
- 提示词语义合同审计：用用户声明的必需/禁用动作词组、精确对白和媒体标签检查重写候选；默认继续输出原文，机械通过并经人工显式接受后才切换候选（实验）
- Creator Workspace：在既有 Studio Timeline 上追加镜头覆盖、seed变体、运行区间、hold-map sidecar、显式运行回执与断点续跑计划；可选桥接现有Long Video后台取消/重试，同时提供不拉伸画面、A/B音轨分别试听的同步音画审片。候选保留计划默认只生成清单；另有双阶段、SHA锁定、可恢复的Quarantine节点，显式确认后只把文件移入输出目录内的隔离区，不提供永久删除（实验）
- ClipProj 与 Sol-Attn 外部兼容审计：只检查插件版本、矩阵维度、硬件和补丁所有权，不复制、加载或执行外部实现（实验）
- 原生 H3 latent 时间线拼接：按视频24fps/音频40Hz双时钟移除后续片段重复的5帧前缀，并在CPU组合完整AV latent（实验）
- 原生 H3 Long Video续接拼接：同时核对Planner与Conditioning报告，按5/22/39帧真实上下文移除双时钟重叠；旧5帧通用拼接保持不变（实验）
- 原生 H3 latent 恢复清单：分块计算联合AV、可选mask和元数据的精确SHA-256；重载后默认不一致即报错，不写文件、不冒充扩散内部断点恢复（实验）
- 原生 H3 latent 检查点 Save/Load：把已经形成的完整联合AV latent、可选mask和受支持元数据保存为无pickle的safetensors，并在另一进程中按文件SHA与内容清单严格恢复；默认不保存，不等于采样中断后续跑NFE（实验）
- 双时钟 Euler NFE 检查点/续跑：仅支持本项目
dual_clock_euler + native_flow；新增运行合约编译器会绑定最终提示词、媒体映射、报告和实际Conditioning张量，显式启用后在每个完整NFE边界原子保存联合AV状态，并可在新的ComfyUI进程中从剩余sigma继续。默认关闭，不支持DPM++、SDE/ancestral、第三方sampler或中断在Transformer前向内部的恢复（实验） - 预检、显存/缓存报告、模型兼容检查和异常路径清理
- LightX2V Turbo-SLA LoRA、85% 动态块稀疏 Sage2、逐次运行审计，以及可选的KJ Sage单入口组合器（实验）
- RAVEN Streaming外部运行时桥接：统一发布参数、在巨大权重载入前检查硬件/内存，并严格审计仅T2VA的因果分块请求（实验）
- Skin Finish 非生成式肤质收尾：P0在可靠遮罩内生成低频肤色/油光候选；固定ParseNet节点只选择皮肤并保护眼眉鼻唇头发，新多人语义节点复用SAM3.1逐镜轨迹、以YuNet五点对齐逐脸解析并与各自人物MASK相交，可选身份分配只作报告标签；缺模型、hash、对齐或覆盖证据时输出空MASK。P1另提供可续跑状态和不经过
GetVideoComponents的两遍文件流式节点；P2 Texture Guard保护源片深阴影/近饱和高光，并对新增裁切和高频纹理下限逐帧fail closed。显式接受后以单线程H.264重编码画面、严格解码后才原子发布，并逐包复制和校验原音频（实验）
稳定节点与 Advanced/Experimental 节点分开。高级节点采用追加式设计：未连接时不改变旧工作流、旧节点 ID、输入顺序和默认采样数学。
安装
Comfy Registry（推荐）
在 ComfyUI Manager 中搜索 MiniMax H3 Audio T8，或在终端执行：
comfy node install minimax-h3-audio-t8
手动安装
将仓库放到：
ComfyUI/custom_nodes/minimax-h3-audio-T8
本插件不自动下载 H3 权重，也没有强制的额外 pip 依赖。请先准备与任务匹配的模型、CLIP、VAE 和 LoRA，并重启 ComfyUI。Long Video 的候选保存和最终拼接要求 ffmpeg 可从 PATH 调用；常规 ComfyUI 整合包通常已经包含，缺失时节点会在写入前给出明确错误。只有使用 8B 提示词重写器时才安装 requirements-prompt-rewriter.txt；依赖被限制在 Transformers 4.x、PEFT 0.18–0.20 和 Pillow 10.x，避免无关升级影响其他节点。LightX2V SLA 实验节点还需要与当前 Torch/CUDA 匹配的 spas-sage-attn；只有使用SLA + KJ Sage组合工作流时才需要另外安装 ComfyUI-KJNodes。可选的Skin Finish语义遮罩节点还需要FaceXLib代码和固定的models/facedetection/parsing_parsenet.pth；多人路线另复用本项目现有的本地YuNet与原生SAM3.1轨迹。ParseNet节点会核对85,331,193字节及完整SHA-256，不会联网下载或接受任意模型路径。
第一次使用
- 打开
examples/workflows/，先从01-basic-generation中的基础工作流开始。 - 将工作流中的模型、图片、视频和音频替换为自己的文件。
- 初次运行建议使用 22 帧、较小画布和 stock sampler，确认能正常解码后再增加时长、分辨率或参考条件。
- 使用工作流中的 NOTE 节点作为参数说明；报告输出可用于排查标签、显存、音频和兼容性问题。
Quick Start 子图
subgraphs/ 提供 7 个 ComfyUI 原生 Quick Start 子图：T2VA、I2VA/FL2VA、Ref2VA、Audio Drive、Long Video、Repair 和 Creator AV Review。它们只封装现有节点并减少外露参数，不改变旧节点的输入顺序或默认值。Creator审片入口保持A/B音轨独立、并排视频无声且默认ABSTAIN，不会自动选择胜者。文件名统一使用日期开头的 ASCII 名称，避免 Registry 打包在部分 Windows/GBK 环境漏文件；导入后的子图标题和 NOTE 仍保留中英说明。详细说明见 subgraphs/README.md。
可选高级扩展
| 功能 | 入口 | 额外要求 | 当前结论 |
| --- | --- | --- | --- |
| 8B 提示词重写 | 14-prompt-relay/2026-08-22_H3_Prompt_Rewriter_8B_Advanced_EXP.json | requirements-prompt-rewriter.txt、Qwen3-VL-8B 基座和 LightX2V LoRA | 16GB 本机可运行并能生成结构化结果，但 CPU 分片很慢；默认生成后卸载 |
| LanPaint 局部 AV 修复 | 03-image-video-edit/2026-08-22_H3_LanPaint_AV_Local_Repair_Advanced_EXP.json | 单独安装 scraed/LanPaint | 画面蒙版与音频秒区间分离，未声明区域回贴原始内容；尚未做高负载质量验证 |
| 外部 BlockSwap | 12-system-memory/2026-08-22_H3_External_BlockSwap_Stock20_Advanced_EXP.json | 单独安装 xiaolibai-sys/ComfyUI-MiniMaxH3 | 只服务外部 MINIMAX_H3_* 类型，不接受官方 MODEL；16GB 未做压力认证 |
| LightX2V Turbo-SLA | 15-sla-attention/2026-08-22_H3_LightX2V_SLA_FL2VA_4Step_Advanced_EXP.json | 固定 SLA LoRA、FL2VA、4 NFE/6V/3A 和匹配的 spas-sage-attn | 256×256×22 INT8 兼容机械验证通过；尚不宣称画质、速度、音频或通用16GB优势 |
| SLA + KJ Sage组合器 | 15-sla-attention/2026-08-22_H3_LightX2V_SLA_KJ_Sage_Composer_FL2VA_4Step_Advanced_EXP.json | 上述SLA条件，加单独安装的 ComfyUI-KJNodes | 解决KJ完整forward绕过SLA的问题；每次Attention只选一个后端，当前完成结构与回归验证 |
| 音频完整性审计 | 05-speech-dialogue/2026-08-22_H3_Audio_Integrity_Audit_Advanced.json | 无；纯CPU信号分析 | 输出PASS/ABSTAIN和证据，不修音；循环音乐可能触发尾首相似提示 |
| 音色远近漂移审计 | 05-speech-dialogue/2026-08-22_H3_Audio_Perceptual_Drift_Audit_Advanced.json | 同内容、同时间线的基准和候选音频 | 真实试听异常的纯二采在1.4～3.6秒触发ABSTAIN，正常一采与本例80%混音PASS；它是声学复核提示，不诊断远场/混响/换声 |
| 多人对白路由预检 | 05-speech-dialogue/2026-08-22_H3_Speaker_Routing_Audit_Advanced.json | 每个角色独立参考音频 | 编译<Audio N>映射并检查重复波形、未结构化笑声/喘息和描述歧义 |
| 提示词预算与角色编译 | 14-prompt-relay/2026-08-22_H3_Prompt_Budget_Role_Compiler_Advanced.json | CLIP为可选输入 | 默认7000匹配当前官方H3 CLI提交上限；7000/7001与三人物映射已测。真实Qwen3-VL 8B/Boogu编译文本为140个token、规划估算153。当前ComfyUI无7000字符硬拦截，故不把API/CLI规则冒充本地tokenizer硬上限；视觉/时间戳token在Conditioning阶段另行加入 |
| 提示词服务路由 | 14-prompt-relay/2026-08-23_H3_Prompt_Provider_Router_Advanced_EXP.json | 本地服务无需密钥；远程服务的密钥通过环境变量提供 | 默认不联网；原始<d>对白先替换为不可猜测令牌，provider只有逐字返回唯一令牌且位置正确才会恢复，真实对白不上传；可选0～2次合同修复且不重传参考图，默认0保持旧请求数；CPU-only 8B已有一条严格合同通过，但把“旋转”改成“站立”，仍未通过语义质量门 |
| 提示词语义合同审计 | 14-prompt-relay/2026-08-23_H3_Prompt_Semantic_Contract_Audit_Advanced_EXP.json | 无；纯本地字符串审计 | 已知“旋转→站立”真实Provider回归会被拒绝并保留原文；空锚点ABSTAIN，非法合同、对白变化和源媒体标签丢失fail closed。词组PASS不等于通用语义等价，仍需人工复核 |
| Creator Workspace | 11-studio-production/2026-08-22_H3_Creator_Workspace_Run_Window_Advanced.json | 复用现有Studio Timeline | 真实3镜头计划、3候选seed、hold-map和显式选择已执行；不隐式排队、不写文件、不删除候选 |
| 同步A/B预览 | 11-studio-production/2026-08-22_H3_Creator_Synchronized_AB_Advanced.json | 输入两个IMAGE帧批次 | 真实39帧A/B已保持源像素并严格解码；几何不一致可强制ABSTAIN；仍不比较音频 |
| 同步音画A/B审片 | 11-studio-production/2026-08-22_H3_Creator_Synchronized_AV_AB_Advanced.json | 两条同内容、同起点且均带音轨的视频 | 保存无声并排画面并分别试听A/B音轨。1088×544近景真人合同以同一124帧latent得到243帧/10.125秒两路成片，精确双时钟、严格解码及1,125MiB最低余量通过；一名审阅者在所有画面/接缝/对白维度判平且无硬失败。该固定简单素材结果不等于普遍等价，仍不自动接受 |
| Creator运行回执、续跑与可恢复隔离 | 11-studio-production/2026-08-23_H3_Creator_Run_Receipt_Resume_Advanced.json | 复用Creator Workspace；运行结果由用户或外部执行器显式填写 | 记录completed/accepted/rejected/cancelled/failed并给出render/review/retry/complete；Retention Plan仍只生成keep/proposed-delete清单。Quarantine节点在示例中默认静音，先prepare_only核对路径/字节/SHA，再以精确plan hash、新epoch和显式确认移动到output/MiniMaxH3/creator_quarantine；同一receipt/epoch可restore或recover，不提供永久删除 |
| Creator × Long Video后台桥 | 11-studio-production/2026-08-23_H3_Creator_Long_Video_Background_Bridge_Advanced_EXP.json | 必须嵌入现有Long Video Background候选保存/自动接受成片链 | workspace hash绑定任务；accepted_count选镜头、retry_count选seed；默认review_only；真实256×256×22 H3采样在1/4步定向取消后显存回到基线+90MiB；另以合法124+119帧轻量控制链验证终态重新挂接、自动续排、Candidate Save/Auto Accept和最终AV合成，真实H3恢复成片质量仍待验收 |
| ClipProj兼容审计 | 12-system-memory/2026-08-22_H3_ClipProj_Compatibility_Audit_Advanced_EXP.json | 单独安装ComfyUI-ClipProj 0.1.13+ | 检查4B/8B投影维度、Qwen3-VL声明与加载模式；不加载模型，也不替换32B默认路径 |
| Sol-Attn兼容审计 | 12-system-memory/2026-08-22_H3_Sol_Attn_Compatibility_Audit_Advanced_EXP.json | 单独安装ComfyUI-sol-attn 0.6.2+ | 检查CUDA/BF16、架构与完整H3补丁所有权；不运行kernel，不作速度/显存结论 |
| ClipProj 4B低负载T2VA桥 | 12-system-memory/2026-08-23_H3_ClipProj_4B_T2VA_Bridge_Advanced_EXP.json | 官方ComfyUI格式Qwen3-VL 4B、4B v3.1矩阵和外部ClipProj | 短T2VA和1088×544近景真人I2VA均已真实生成。固定简单素材中4B与8B、4B与原生32B的单人盲评均为全维度平局且无硬失败，但这不建立普遍等价或非劣；32B继续默认。原生32B本次最低余量643MiB，仅证明这一条通过512MiB门，不代表普遍16GB安全；纯文本qwen_3_4b仍不可替代 |
| ClipProj 8B完整T2VA桥 | 12-system-memory/2026-08-22_H3_ClipProj_8B_T2VA_Bridge_Advanced_EXP.json | 8B Qwen3-VL、v3.1投影矩阵和外部ClipProj | 本机固定seed短T2VA已真实编码、生成并严格解码；只证明链路，其他模态和质量A/B仍待补 |
| ClipProj 8B完整I2VA桥 | 12-system-memory/2026-08-22_H3_ClipProj_8B_I2VA_Bridge_Advanced_EXP.json | 同上，并连接一张首帧图 | 本机256×256×22、4步I2VA已走通视觉塔与联合AV生成并严格解码3/3；不等于32B画质或对白遵循度已通过 |
| ClipProj 8B完整FL2VA桥 | 12-system-memory/2026-08-22_H3_ClipProj_8B_FL2VA_Bridge_Advanced_EXP.json | 同上，并连接独立首帧与尾帧 | 本机256×256×22、4步FL2VA已走通双视觉关键帧与联合AV生成并严格解码3/3；只证明短链执行，不证明长插值或32B画质等价 |
| ClipProj 8B完整Ref2VA桥 | 12-system-memory/2026-08-22_H3_ClipProj_8B_Ref2VA_Bridge_Advanced_EXP.json | 8B ClipProj、Ref2VA pruned INT8与一张参考图 | 本机Stock20短链及原生32B同seed机械对照均通过严格解码；身份代理未稳定过线且8B只余约62MiB，故不宣称质量等价、省显存或16GB安全 |
| Sol-Attn保守T2VA桥 | 12-system-memory/2026-08-22_H3_Sol_Attn_T2VA_Conservative_Advanced_EXP.json | 外部Sol-Attn 0.6.2+ | 4步默认dense百分比为0，首三/末块dense、exact conditioning KV、首次strict；本机0.737MP实跑记录5139 tokens并进入kernel，但单次顺序A/B未显示速度或显存优势；一名审阅者给出全平且没有失败备注，仍不足以提升为推荐路线 |
| RAVEN Streaming受保护T2VA | 16-raven-streaming/2026-08-23_H3_RAVEN_Streaming_T2VA_Guarded_Advanced_EXP.json | 单独安装外部RAVEN插件0.1.0、完整BF16 H3和 mandatory RAVEN LoRA | 真正采样/预览仍由外部节点执行；T8统一4步、12/3、2/2、cpu_pinned参数并检查T2VA合同。当前16GB/128GB机器低于已审阅资源范围，默认会在加载前拒绝，不宣称可运行 |
| Skin Finish肤质收尾 | 17-skin-finish/2026-08-24_H3_Skin_Finish_External_Mask_Advanced_EXP.json | 可靠外部MASK，或同一来源帧的Face Refine Plan | Basic/Advanced/Preview-Audit均默认保留source；仅做SDR非生成式低频肤色和油光候选，不修复五官、模糊、身份或口型。缺mask、面积异常、来源不一致和音频不一致均fail closed |
| 多人Skin Finish与原音频封装 | 17-skin-finish/2026-08-24_H3_Skin_Finish_MultiPerson_Video_Finalize_Advanced_EXP.json | 原生SAM3.1 track plan、CPU YuNet、未裁切8-bit SDR文件VIDEO | 复用一次追踪结果，不重复加载SAM；镜头内bbox EMA和绝对帧state支持重叠续块。最终候选默认不保存，人工接受后仅重编码视频，兼容音频包payload逐包复制并SHA复核；HDR/10-bit、旋转、裁切和未知codec拒绝 |
| 两遍文件流式Skin Finish | 17-skin-finish/2026-08-24_H3_Skin_Finish_Two_Pass_Video_Stream_Advanced_EXP.json | Long Video/Studio最终未裁切8-bit SDR文件VIDEO、固定YuNet | 不连接完整IMAGE：第一遍只保存脸框/切镜/来源摘要，第二遍按默认4帧chunk处理并以单线程H.264立即编码，发布前执行FFmpeg严格解码。1088×544×124单次机械验证通过；默认false时不分析、不写文件，共享中性色、无身份识别/语义parser，不等于去模糊或修脸 |
| Skin Finish纹理/曝光机械护栏 | 17-skin-finish/2026-08-24_H3_Skin_Finish_Texture_Guard_Advanced_EXP.json | 已生成并审核来源绑定的source/candidate/used mask | 默认保护源片深阴影和近饱和高光；新增裁切或源片相对高通RMS下限失败时整帧回退。0.592MP×124单次默认门机械通过且音频PCM一致；它不是自然肤质评分、语义parser、去模糊或毛孔生成器 |
| Skin Finish固定ParseNet语义遮罩 | 17-skin-finish/2026-08-24_H3_Skin_Finish_Semantic_Mask_Advanced_EXP.json | FaceXLib代码、固定v0.2.2 ParseNet权重和同一来源帧的Face Refine Plan | 仅CPU逐帧解析并在结束后卸载；默认选择skin，排除鼻、眼、眉、嘴唇、头发和配饰。真实3帧固定权重机械检查通过；现有plan无五点关键点且只代表单轨人脸，尚未通过多人/跨镜和人工肤质验收 |
| 多人五点ParseNet语义遮罩 | 17-skin-finish/2026-08-24_H3_Skin_Finish_MultiPerson_Semantic_Mask_Advanced_EXP.json | 同一来源IMAGE、hash-valid SAM3.1逐镜track plan、本地YuNet、固定ParseNet；identity assignment可选 | 每个可靠脸只匹配一个人物轨迹，五点对齐到FFHQ 512后解析并反投影，与各自人物MASK相交；覆盖不足整批返回空MASK。0.67584MP六帧双人真实YuNet/ParseNet机械检查12/12 READY；该低负载检查使用来源绑定的人物区域夹具，不冒充一次新的SAM实跑或人工肤质通过 |
| 原生latent时间线拼接 | 04-long-video/2026-08-22_H3_Native_Latent_Timeline_Concat_Advanced_EXP.json | 输入完整H3嵌套AV latent，batch=1且画布一致 | 真实22+22→39帧单次解码、52000采样无损音频和重复hash已通过；旧32B路线两次未过512MiB余量门，另一个固定8B ClipProj短链以605MiB最低余量通过一次0.25秒采样门。独立段仍会换景，不能外推为无缝或通用16GB安全 |
| 原生latent Long Video续接拼接 | 04-long-video/2026-08-23_H3_Native_Latent_Continuation_Concat_Advanced_EXP.json | 上一段/累计时间线latent、本段采样完成latent，以及同一Planner与Conditioning的直连报告 | 精确移除5/22/39帧上下文；124+124在22帧context下得到226帧，默认要求音画上下文。最终隐藏尾帧须一次解码后再裁；不等于NFE断点恢复、无缝质量或省显存证明 |
| 原生latent恢复清单 | 04-long-video/2026-08-23_H3_Native_Latent_Resume_Manifest_Advanced_EXP.json | 输入已保存/重载的完整H3嵌套AV latent；恢复核对时同时粘贴旧manifest_json | SHA-256分块大小不影响摘要；默认error阻断内容、shape、dtype、mask或checkpoint ID不一致。节点不保存latent、不恢复NFE内部状态，也不作续段质量/显存保证 |
| 原生latent检查点Save/Load | 04-long-video/2026-08-23_H3_Native_Latent_Checkpoint_Save_Load_Advanced_EXP.json | 输入完整H3嵌套AV latent；首次保存需显式打开confirm_save | 文件限制在ComfyUI输出目录，唯一文件名、写入后校验和原子放置；独立进程严格恢复已逐tensor/mask/元数据核对一致。它只恢复已完成latent，不保存Transformer、采样器、队列或CUDA状态 |
| 双时钟Euler NFE检查点/续跑 | 04-long-video/2026-08-23_H3_Dual_Clock_NFE_Checkpoint_Resume_Advanced_EXP.json | 仅限dual_clock_euler + native_flow；必须填写精确模型合同，并把Conditioning的4路输出接入NFE Run Contract编译器，不能把普通report文本直接接到JSON输入 | 默认disabled且不写文件；一条256×256×22、4步真实H3任务已在第2步中断，由新ComfyUI进程续跑，最终联合AV latent、解码RGB与PCM均和不中断控制逐位一致。仅证明该固定合同，不支持多步历史、随机SDE或任意wrapper |
这些扩展均为追加节点。未放入旧工作流时不会改变旧采样路径。外部项目及模型权重不随本仓库分发。
以后生成的匿名评审页会先要求选择“可判断 / 原素材不足 / 播放问题 / 不确定”。只有“可判断”
的组才统计 A/B/平；其余组明确记为 ABSTAIN，不会被误算成方法平局或质量通过。旧版评审 JSON
仍可解析，已经完成的盲测页和私钥不会被重建工具覆写。
工作流目录
| 目录 | 用途 |
| --- | --- |
| 01-basic-generation | 基础 T2VA、I2VA、FL2VA 和 Ref2VA |
| 02-audio-control | 音频驱动、音频锁定、重混、参考音频和双时钟 |
| 03-image-video-edit | 图像编辑、视频编辑和音视频参考 |
| 04-long-video | 分段生成、长视频、断点恢复和拼接 |
| 05-speech-dialogue | 单人语音、对白、演绎和 ADR（实验） |
| 06-face-refine | 单人/多人脸部修复与 SAM3.1 追踪（实验） |
| 07-motion-detail | 动态细节、尾段采样、FETA 与 Motion Recovery 二采修复（实验） |
| 08-multi-keyframe | 中间关键帧和关键帧计划（实验） |
| 09-hybrid-model | FL2VA/Ref2VA 混合模型研究（实验） |
| 10-speed | SPEED 多分辨率采样研究（实验） |
| 11-studio-production | 生产、字幕、时间线和批量辅助 |
| 12-system-memory | 预检、显存、缓存、VBAR 和释放策略 |
| 13-latent-upscale | 32 倍数对齐的 latent 放大 |
| 14-prompt-relay | 全局提示词常驻、局部事件按时间接力（实验） |
| 15-sla-attention | LightX2V Turbo-SLA 动态块稀疏 attention（实验） |
| 16-raven-streaming | RAVEN因果分块流式T2VA与资源/合同保护（实验） |
| 17-skin-finish | 解码后肤质收尾、遮罩门禁和人工审计（实验） |
每个目录都有自己的 README.md，说明适用场景、参数建议和已知限制。完整索引见 examples/workflows/README.md。
常用稳定节点
| 节点 | 作用 |
| --- | --- |
| MiniMaxH3AudioConditioningT8 | 统一处理任务类型、提示词、图像/视频/音频参考和 AV 条件 |
| MiniMaxH3DualClockSamplerT8 | 使用独立的视频/音频时钟采样 |
| MiniMaxH3AVDecodeT8 | 解码 AV latent 并保持音视频时间线 |
| MiniMaxH3AudioLatentControlT8 | 音频锁定、重混和参考策略 |
| MiniMaxH3LatentUpscaleBy32T8 | 按 32 像素网格放大 3D latent |
| MiniMaxH3PreflightT8 | 在生成前检查模型、标签、尺寸、帧数和显存风险 |
提示词媒体标签
显式引用建议使用官方标签：<Picture 1>、<Video 1>、<Audio 1>。标签编号必须与实际输入对应。节点会规范化大小写和空白，并在报告中列出最终绑定结果。
- 只有一个明确的旧式参考输入时，可兼容历史工作流的零编号/单参考写法。
- 多个参考输入出现歧义、缺失或重复标签时，严格模式会直接报错，不会静默猜测。
- 看到
prompt media tag validation failed时，请先检查任务类型、标签编号和对应输入是否真的接入。
模型与尺寸
常见目录：
models/diffusion_models H3 DiT 模型
models/text_encoders Qwen/MiniMax CLIP
models/vae 视频 VAE、音频 VAE
models/loras Turbo、质量或任务 LoRA
models/latent_upscale_models
models/face_detection/checkpoints
不要只按文件名判断兼容性；任务类型、模型家族、LoRA、VAE 和 ComfyUI 版本必须匹配。工作流中的 NOTE 和预检报告会提示推荐组合。
重要约束：
- H3 是联合 AV Transformer，视频和音频共享部分计算；参考图、参考视频、参考音频越多，显存和序列长度越高。
- 宽高应为 32 的倍数；
1920×1088作为官方高分辨率参考面积，而不是学习型放大节点的执行禁令。该节点允许用户选择更大画布并报告高显存风险，实际能否完成取决于显卡、帧数和参考数量。 - 常用帧数遵循 H3 的网格约束，例如 22、124、362；不要只按“秒数”猜可用帧数。
- 16GB 显卡不要默认视为所有任务安全。先用较小画布/短片段通过预检，再逐步增加规模。
- 人脸检测模型可以通过
ComfyUI/models下的目录符号链接或 Junction 指向其他盘；节点仍拒绝绝对路径和路径穿越，并在报告中标记外部存储。 - 环境审计会标记 RTX 50 系/compute capability 12.x 的 SageAttention 高 token 风险；Strict Sage 在 50,000 packed rows 以上直接拒绝执行，避免把可导入误当成高分辨率输出正确。
- SPEED、Prompt Relay、动态细节、Hybrid、多帧关键帧、多人脸修复和语音节点目前按实验功能使用；未通过本机验证的组合会 fail-closed。
- Motion Recovery 是独立的“pass-1分析 → 问题段扩时 → 部分去噪V2V → 恢复原时钟”链，不是光流插帧、修脸或单采样器增强。Analyzer默认使用
auto_conservative_exp，Auto Gate通过ComfyUI原生lazy input实现真正旁路：平静片ABSTAIN时不请求任何二采节点，并原样输出pass-1画面和音频。736×416×124真实平静T2VA已验证second_pass_requested=false，旁路与pass-1的MP4、解码画面及PCM哈希均一致。Stock20 T2VA、I2VA、FL2VA和独立Ref2VA模型现各有一条真实二采成片，均恢复为124帧并通过严格音视频解码；I2VA还生成了pass1_original、pass2_recovered_exp和blend_exp三种完整对白音轨。完整人工试听确认默认原音正常；纯pass2_recovered_exp在中段会突然变成远处声音再恢复，已降为诊断用途；blend_exp在本次pass1_mix=0.8素材中正常，但仍只是单素材EXP结论。多模态单条成功不代表稳定提质或通用16GB安全，FL2VA/Ref2VA/I2VA粗采余量均曾低于512MiB，不要在首轮叠加EAV/STG/RF Restart/BlockCache。 - Enhance-A-Video / FETA 节点按论文公式从目标视频 Q/K 计算跨帧 CFI，只直接增强目标视频 attention 输出。原节点继续支持 Stock20 的 T2VA / I2VA / FL2VA / L2VA，以及严格限定的修正 Alpha8 Turbo8 T2VA；独立 Reference Composer 只开放原生 Stock20 Ref2VA 和任务型 Hybrid，不改变旧节点合同。上述七类路线现均至少完成一组 0.7MP 同输入、同 seed 的 disabled/apply 机械与媒体对照；Ref2VA 增强端最低显存余量仅约 417MiB，低于项目 512MiB 门槛。自动指标只能证明轨迹和联合音频发生变化，不能证明画质、参考遵循或声音更好，因此所有路线仍不宣称稳定提质、音频非劣或通用16GB安全。
Enhance-A-Video + Strict Sage Advanced EXP由一个组合节点同时拥有 FETA 路由和本机 SageAttention HND 后端，避免第三方整块 Attention patch 绕过 FETA。它不会静默回退到 PyTorch attention；本机一条 1152×640×124、Stock20 实测完成 1000 次 FETA 测量和 1000 次 Sage 调用，失败/回退为 0，并通过三轮严格音视频解码。该单条机械验证不代表画质更好、声音非劣、速度更快或通用 16GB 安全；使用时不要再叠加 KJ Sage、BlockCache、STG 或其他全局 attention patch。Enhance-A-Video + Prompt Relay Composer Advanced EXP解决两个独立节点争用同一Attention入口的问题：它验证现有Relay绑定后，在一次路由中先执行局部事件Relay，再只对目标视频输出行应用FETA；关闭FETA时保留原Relay MODEL。当前仅开放Stock20 T2VA，一组736×416×124、20步基础机械对照及严格媒体检查已通过，但余量低于512MiB；不作为稳定提质、音频非劣或16GB安全路线宣传。Enhance-A-Video + BlockCache / STG / Long Video是三个隔离追加组合器：BlockCache只接受已核对合同的CPU缓存并按full=50/hit=1审计实际执行块；STG把同一FETA规则用于主分支和弱分支，默认审计50/49块及额外联合AV前向；Long Video保留原上下文/layout拥有者，并为每个segment_index/context_frames创建独立Stock20审计。三者当前只通过低负载确定性合同、注册和工作流导入检查；按要求未做压力测试，不宣称提质、音频非劣、提速、省显存或通用16GB安全。- 基础
MiniMax H3 LightX2V SLA Loader + Attention (Advanced EXP)仍必须直接接干净的Dual-Clock模型并独占attention。若要保留KJ MiniMax H3 Sage，请使用专门的SLA + KJ Sage Composer：连接顺序为Dual-Clock → KJ Sage → Composer，不要再插入ModelAttentionBackend或Sol-Attn。SLA生成路径运行block-sparse Sage2，KJ只负责非SLA调用或dense_lora_control；同一次Attention不会重复跑两个kernel。当前结论仍不等于画质更好、速度更快、音频非劣或普遍16GB安全。 - Prompt Relay 与 Turbo8 组合时，连接顺序必须是
UNET → Prompt Relay Conditioning → 修正 Alpha8 Bypass LoRA → DualClock Sampler；把 LoRA 接在 Relay 前面会被主动拒绝。 - Prompt Relay 需要保留输入原声时使用
lock_source，并把节点的mux_audio接到最终保存节点；native/remix/reference_only仍可能因 H3 联合 AV Transformer 而改变声音。 drive_audio是联合生成条件，不是确定性音素/口型约束。希望尽量保留原声时，使用audio_mode=lock_source、add_source_as_reference=true，提示词引用实际编号的<Audio N>； 已知台词时再把逐字内容放入<d>...</d>。最终保存必须接mux_audio，不要误接AV Decode的generated_audio。这些设置能够保留最终输入音轨并提高语义提示强度，但不能保证嘴型逐音素同步； 广播级精确口型需要在H3生成后使用专用唇形/人脸重定向工具。- Prompt Relay 默认
video_only_paper不改变旧工作流；只有显式插入Query Route并选择joint_av_exp，才会把局部事件时间扩展到目标音频。该模式是H3实验扩展，不是论文已验证能力。 14-prompt-relay已提供 T2VA、I2VA、FL2VA、L2VA、Ref2VA、Hybrid 六种视觉任务模板，以及参考视频+同编号音轨、独立参考音频、联合AV和Turbo8模板；各任务的图片/视频/音频接线、模型顺序和已验证边界写在画布 NOTE 中。- 六种视觉任务以及“参考视频+同音轨”“独立参考音频”现均至少有一组同输入、同seed、同NFE的baseline/Relay机械对照：16条成片均为736×416×124、24fps、32kHz双声道，严格音视频解码通过且无黑帧、冻结或削波；I2VA/FL2VA/L2VA/Hybrid的首尾帧锚点代理没有出现明显回退。该结果只关闭媒体与锚点合同，不代表Relay必然改善动作、身份、画质或声音；参考音轨不是原波形复制，感知优劣仍需观看和试听。
- 局部动作可用多个
Prompt Relay Event Advanced节点逐项复制串联，并连接普通 Plan 或 Studio Packet桥接；普通Plan未连接时继续读取旧local_prompts/time_ranges，Packet桥接未连接时继续读取旧events_json。Studio 的Unified Cast + Sound Canvas + Prompt Compiler编译结果仍是全局画面/声音事实源，节点不会自动拆剧情或改写事件。 - Prompt Relay显式时间范围必须按开始时间排列；
frames只接受整数帧，percent只接受0..100。零个事件（包括整条Event链全部关闭）会只保留全局提示并无补丁直通；只有一个事件时也不安装补丁，通常应直接并入global_prompt。 Prompt Relay Preview Advanced可以在不加载UNET、CLIP、VAE且不执行采样的情况下检查事件帧段、秒数、覆盖关系和Plan哈希；Prompt Relay Resource Estimate Advanced还能按画布、参考素材数量和query chunk估算H3 packed行数及Relay显式bias峰值，并在同一报告中列出736×416、1152×640、1920×1088三档矩阵。后者只是一项内存规划代理，不包含模型权重、完整attention激活、VAE/CLIP、VBAR或碎片，绝不等同于“16GB安全”。14-prompt-relay里的11份生成模板已内联Preview，另附一份纯时间线/资源预检工作流。- Long Video需要分段事件时，使用独立的
Prompt Relay Long Video Window/Conditioning Advanced；全局Plan只创建一次，本地渲染起点按accepted timeline start - context overlap计算。旧Long Video节点与旧工作流未改。本机已完成一条736×416、Turbo8、22帧上下文的segment 0→1真实链：输出124+102帧、事件没有从头重启、视频/音频各3轮严格解码通过，整卡峰值约15478/14984MiB。该单条结果仍只证明机械与时间线可用；最终音频接缝需试听，不能外推为普遍画质或16GB安全。
故障排查
遇到 OOM、NaN、花屏、音频丢失或人脸跳动时，请保留完整 ComfyUI Error Report，并记录：任务类型、模型文件、宽高、帧数、采样器/调度器、是否双时钟、参考输入、音频采样率和所有显式媒体标签。常见原因是：输入尺寸未对齐、重复连接 Sigma/采样器、参考标签错位、把实验节点接到不兼容的旧 core，或显存不足后继续复用缓存。
若KJNodes的MiniMax H3 Mem Eff Sage Attention Patch报
sageattention is not new enough ... or could not determine CUDA architecture：
- 先运行
MiniMax H3 Environment Audit / 环境兼容审计 (Advanced)，把attention_backend设为sage_attention、enforcement保持report_only。 - 查看报告中的
environment.sageattention_runtime。诊断会分别显示包/core导入错误、当前 KJNodes所需的六个core符号、wheel报告的smXX架构和当前GPU架构是否一致；它不加载H3模型， 也不执行attention kernel。 - “已安装最新版”本身不能证明兼容；Torch、CUDA、Python、GPU架构与Sage wheel必须同时匹配。 若诊断不通过，先移除KJ Sage节点改用stock attention，或为当前ComfyUI Python重装匹配wheel并 完整重启。不要让工作流静默降级后仍宣称正在使用Sage。
开发与回归
$env:PYTHONPATH = 'F:\AI-T8-video-onekey\ComfyUI'
& 'F:\AI-T8-video-onekey\python\python.exe' -m pytest -q
修改节点前请先阅读本地 SKILL.md、features.json、meta.json 和对应工作流目录说明。新增高级功能应保持旧工作流可加载；检测到未知 ComfyUI core 或第三方全局 patch 时，应明确报错而不是静默降级。
链接与许可
社区与服务
插件代码采用 GPL-3.0-or-later。H3 权重、Qwen、ASR、SAM3.1 和第三方 LoRA 各自遵循其上游许可；本仓库不随插件分发这些权重。使用人声、人物参考图或克隆功能时，请自行确认素材权利并遵守适用法律和模型许可。