comfyui-minimax-h3-audio-T8
A ComfyUI extension with 331 custom nodes.
Nodes (331)
The H3 'memory saver' that honestly tells you it might save nothing
When other packs' attention patch nodes won't talk to H3, this shim translates
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
The bouncer for H3's audio-refine chain — check everything before you burn a run
The audio-refine middleman that builds the 4-step tail and won't overclaim it
It needs to know exactly what you already ran
The node that hands SamplerCustomAdvanced a ready-made 4-step audio tail
Turns the refine plan into noise, guider, sampler and sigmas
The strict sibling of the dual-clock setup
The split that stops your next long-video segment from inheriting yesterday's audio fix
When your audio-refine pass needs a different model, this router validates the pairing
A strict four-step tail, and only two denoise points allowed
A signed, deterministic plan for re-sampling just H3's audio tail
Your refine only counts if you actually listened
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
The two-plug adapter that welds MiniMax H3's picture and sound latents together
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
CADS for H3 references — noise on the picture, hands off the audio
The v2 plan that fixed chunked two-pass H3
H3 low-sigma two-pass
Mask-preserving low-sigma refinement for H3
The planner for H3's chunked two-pass upscale — and why full-frame is the safe default
Run H3's chunked two-pass upscale, and never touch the audio tensor
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
A cache *planner* for long H3 shoots — tells you what's cached without touching a file
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'
H3 Dance Motion Source
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
Your H3 clip is 24fps and the motion looks juddery — this double it with DLSS, not RIFE
NVIDIA's denoiser-plus-upscaler as a ComfyUI post node
Why nothing in this pack runs before the audit says READY
The file-streaming DLSS-NR node
One persistent DLSS process across all your frames
An AYS schedule that knows it isn't an image-model AYS schedule
One clock for video, one for audio
The 4+4 long-video node
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
The node that forces a human to look before a face fix lands
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 H3 Face Refine v1.1.1 Sampler-Mask Patch
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
Carve one legal H3 window out of your video — context and all
Tell H3 exactly which frames are ugly — it plans the repair window
The atomic 'yes' that makes a face fix durable
Rebuild the final video from a manifest — H3 never even loads
Start a crash-safe face-fix run across many windows
Label every face track with the right character — or override by hand
FastH3's 4-step contract — apply the LoRA first, then let this wire the sampler
The T8 FlashVSR strategy node
What it actually checks (and what it won't fake)
2× or 4×, audio untouched
Stop H3 from round-tripping to the CPU every step
FreeNoise for long H3 videos — one shared noise pool, permuted per segment
Steer an H3 shot with a depth, pose or edge video — structure in, motion out
Load the ControlNet that steers MiniMax H3's video with depth, pose or edges
Does your ComfyUI actually have Generic Loops?
The tiled-VAE fix that H3 tested and rejected — now a read-only audit
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 renderer that plays every scene in order
The V2 vocal-lock renderer
The renderer behind the accepted 32-second MV
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
Killing the color seams in a multi-segment H3 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
Prompt Relay timelines plus Enhance-A-Video
H3 long video that renders every segment in a single execution
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
Refine the tail of every long-video segment without rebuilding the loop
The audit that catches H3 segments going green (or dark) at the seams
Keep the right voice on the right character across long H3 segments
The first-shot approval door for voice-plan releases
An H3 LoRA loader that stops being picky about where the LoRA came from
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
The prompt node that refuses to invent lyrics
The six sections MiniMax actually expects, compiled by hand — on purpose
The scene planner that needs two audio tracks — and why that's the point
Directing each shot like it's a contract
The CPU-only storyboard planner
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
Hard-lock the previous segment's native latent tail
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
The 'is my H3 setup actually broken' node (and why 'unknown' is an answer)
Trimming an H3 render by seconds, video and audio together
The PDD setup node that isn't an ordinary LoRA loader
The five-second check that stops an H3 render before it OOMs at step 40
Reading a Prepared Tao/LTX Input Bundle
H3's Serial Video Node
The H3 sampler that skipped 37% of my render time
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
Carry one reviewed mask through the shot — and make it stop at the cut
A read-only RAFT audit that tells you if your H3 shot actually moves
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
RealBasicVSR temporal restore — cleanup after H3, with audio treated as untouchable
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
Who decides how much VRAM H3 gets to hog? A node that plans, but doesn't touch
A save node that refuses to write a corrupt H3 MP4
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
Dichromatic specular attenuation
The frequency-split rebuild
The node that doesn't frontalize faces
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
The node that executes the cast sheet, track by track
The little node that builds the per-person skin-fix plan
Actually look at a skin-finish candidate before you accept it
The two-pass quality stream
A fail-closed audit that isn't a beauty score
A skin mask that knows eyes from cheeks, so your finish doesn't smear
The frequency split that cares about highlights (specular-aware, not a replacement)
The guided-filter skin finisher that refuses to eat the texture
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
One keyframe at a time
Keyframed skin finish, per person, per shot
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
The SLA LoRA loader that refuses to re-quantize your FP8 base
The audit node that calls the render a failure when the math didn't happen
The precision-first H3 attention patch
Before you bolt Sol-Attn onto H3, let this node check the patch ownership
Get your H3 draft ready for the LTX-2.5 refiner — audio stays out
The low-sigma LTX refiner for when the official route changes the face too much
The official LTX-2.5 3-step refiner — apply the 0.8 LoRA first, then wire the sampler
The fast decode that ends H3 Super — TAEHV wide, not the full LTX VAE
The TAEHV encoder you should mostly not use — legacy round-trip, diagnostic only
The one-input loader that hands H3 Super its fast codec
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
Subject-safe RGB compositing done honestly
Why your H3 progress preview is one frame — and how to check if it could be better
Sharpen H3 Footage Without Making the Motion Flicker
Lock a Music Bed Into an H3 Clip Past the Dialogue
Pin a reference image to an exact second — without spending an H3 slot
Give H3 a whole video of motion cues, timed to the frame
The Topaz node that stops you guessing
The Official Topaz Interpolation Node
Topaz enhancement inside ComfyUI — and why it writes a lossless master, not an MP4
Finish the Second Half of an H3 Sample You Started Elsewhere
Save Your Half-Finished H3 Latent — Only When You Say So
Plot an object's path across the clip — without touching attention
Turn your trajectory plan into conditioning H3 Fun Control can actually eat
Split an H3 Sample in Half So You Can Resume It Later
Before you burn an hour compiling TensorRT engines, run this 5-second check
12GB free VRAM, 24GB RAM, and a long coffee break
TensorRT VAE decode — the recommended TRT node, and the benchmark that says 'no total win'
The full TRT VAE (encode + decode), and why its own author tells you not to use it
The router that refuses to let one pretend to be the other
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 plan node that won't let you misstep
Grafting a distilled branch onto your H3 model
The node that plans VDN's second pass so you don't have to
A 4.28GB audit that never actually loads the 4.28GB branch
Generate one window, look at it, then decide
The two-second check that saves you an hour of sampling
Save the outpaint — and it will refuse if the first frame moved
Decode, paste back, and hand you an MP4 that actually plays
Sample the other 21 windows under the settings you approved
See the crop before you spend an hour generating it
Per-side prompts for the new canvas — and an honest person-box audit
Reopen a candidate you generated yesterday, without re-sampling it
Sampling finished, the save crashed. Don't re-run it.
Reconnect to a prepared run without re-encoding a single frame
Get your confirmed candidate back after a restart — by ID alone
Decide the new canvas before you burn an hour of GPU
The regional-prompt version of Prepare — keep it on its own run name
Encode the source, audio and prompt once — then never again
Binding region prompts onto the H3 model — and what it refuses to bind
Every window in one queue, resumable if your PC dies
The one-click approval gate — and why it defaults to off
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
Scripting H3-World's 37-step action timeline
One scene prompt, 37 bound actions
The H3-World LoRA loader that also installs the secret runtime
The H3-World save node that refuses to write a corrupt MP4
MiniMax H3 Audio T8
简体中文 | English
2026-09-14 v1.79.6：完成星光以外的正式Topaz闭环。14个高清模型和5个插帧模型均已走通真实官方生产worker；新增明确的10-bit SDR→HEVC Main10保持路径、多卡GPU序号和“先高清再插帧”串联工作流，并修复无损母版被8-bit审计误拒绝。HDR、VFR、隔行和旋转元数据仍明确拒绝，不静默转换；商业权重只下载到本机，不随仓库发布。详见更新说明。
2026-09-14 v1.79.5：继续修复Topaz交付边界。手动参数现在按模型定义过滤，固定预设模型不再因不支持的滑块报错；权重检查遵循正式网络模板，修复Rhea、Nyx XL和Theia误报未下载；运行前按音频编码选择MP4/MKV；10bit/HDR不再静默降成8bit；ComfyUI进度条显示推理帧与审计阶段。Apollo已完成真实短片机械验证，画质仍待人审。详见更新说明。
2026-09-14 v1.79.4：修复正式Topaz把15秒视频保存成12.9GB无损MOV、再接SaveVideo写满磁盘的问题。高清节点现在默认一次性GPU编码高质量H.264 MP4，禁止再接SaveVideo；超大无损母版保留为高级审计选项。新增模型用途说明、自动/手动参数及独立中文滑块，并新增正式tvai_fi 2x/4x插帧节点和工作流。使用时无需保持Topaz软件打开，但须已登录并先下载模型。详见更新说明。
2026-09-13 GitHub 主线更新：指定LTX成片、Tao新对白5秒恢复片及最新Dance 4+4／前段成片LOW上下文8秒样片均已完整人审接受，不重复生成或审片。Dance正确示例与接缝防回归说明已保存；旧Dance与两条深度失败实验没有进入正式工作流。Tao原生成任务收尾触发内存保护仍保留失败，媒体接受不等于新封装完整GPU可靠性认证。可选深度参考仅作负实验说明；另见资源清理、AdaLN、HJL与音频边界。
本次更新 v1.79.0：新增双模型4＋潜空间放大＋4内循环、正式Topaz高清后处理和R1可靠性修复。新双模型模板默认两段8秒，已评样片的画面、声音、口型和接缝可接受，旧工作流不变。OpenVDN不再因上游Sol的注意力override单独报错；VDN仍使用自身注意力计算，不宣称两种算法叠加提速。星光暂停，未作为已完成能力发布。详见更新说明。
这是一个面向 MiniMax H3 的 ComfyUI 节点包。它不只做文生视频，还把图生视频、首尾帧、参考图、参考音频、长视频、口型、加速和成片修复整理成可以直接使用的工作流。
当前版本：1.79.6 · 331 个节点 · GPL-3.0-or-later
本版整理了仓库目录：实现代码集中到 h3_t8/，首页不再堆满 Python 文件。旧工作流、模型位置、节点参数和三个 TRT 命令入口保持不变；不需要重新下载模型。详见 目录结构与更新说明。
1.78.0 新增的 TRT VAE 可选后端 和 5 份工作流继续保留：安装检查、本机编译、仅视频解码、完整编解码，以及同潜空间双路对照。已审短片和文字对照获得整体认可，保留 EXP；长片未验证。需要独立 TensorRT 环境和本机编译引擎，旧工作流、生成模型和音频 VAE 不变。
不保证整流程更快。 本机短片包含加载、解码和保存时，原生约 30.54 秒、TRT 约 31.33 秒，没有总耗时收益；不能把裸解码提速当成整条生成提速。安装与限制见 1.78.0 发布说明。
1.77.0 新增的渐进首采和 DLSS 2x 插帧仍作为独立 EXP 节点提供，不替换旧工作流。视频扩画默认联合解码，部分素材可能出现条带、重复纹理或接缝，不承诺无缝扩画。
渐进首采 T2VA / I2VA先小画幅采 6 步，学习型潜空间放大后再采 2 步。本机约 3 秒短片热测少用约 37%～38% 端到端时间；已审画面评价相近，对白组声音和口型正常，但游戏组仍有音量和音效内容差异，32 秒路线未通过，不保证所有素材同样提速或保音。
DLSS 视频插帧 2x把 24fps 变成 48fps、30fps 变成 60fps，保持时长、分辨率和原音轨；不是超分，也不加速 H3 生成。已审短片可接受，不保证所有快动作和 HUD 都无伪影。需要用户自行准备 DLSSG 运行文件和依赖，见目录 README；节点不自动下载或安装。更多变更见 1.77.0 发布说明。
先从哪里开始
代码已集中到 h3_t8/，方便在首页直接阅读说明。工作流仍在 examples/workflows/，模型存放位置、节点名称、参数和连线不变；已有工作流不需要重新制作。旧的三个 TRT 命令入口也保留在根目录。开发者路径说明见 目录结构。
如果你第一次使用，建议按这个顺序：
- 从
examples/workflows/01-basic-generation跑一条普通 H3 视频。 - 需要参考图、首尾帧或音频控制时，看
02-audio-control和对应的参考工作流。 - 需要 OpenVDN 8 步时，先下载
t8star/Vdn-Minimax-H3-Comfy，再用10-speed里的OpenVDN_DMD8_*_Advanced.json。 - 需要 WASD 人物/镜头控制时，下载
t8star/Minimax-H3-World-Comfy，再看26-h3-world；首版固定为首帧 I2VA。 - 需要长视频或 MV 时，看
04-long-video和24-mv-lipsync。 - 需要图片或成片超分时，看
25-dlss-nr；这是 Windows RTX 专用的可选后处理。 - 只想修一小段崩脸、不想重绘整条视频时，看
06-face-refine中 2026-09-05 的 Window 工作流。 - 需要给现有视频扩上下左右画面时，看
27-video-outpaint；第一次先跑范围预览，再走候选、审图、确认和接续。
每个高级工作流都带画布说明。先替换模型和输入素材，再运行；不要一开始就把多个 LoRA、Attention 加速器和采样器叠在一起。
双模型 4+4 长视频现在另附三份可选外部 LoRA 示例，分别使用 Core PyTorch、KJ H3 Sage
或已审计的 ComfyUI-sol-attn / SolAttentionPatch。一采、二采必须各走独立链路：
底模 → 本路 Turbo LoRA → 本路可选外部 LoRA → 本路唯一 Attention 后端 → 对应 MODEL。
外部 LoRA 默认 disabled；启用前放入 models/loras，并使用本项目的 H3 兼容加载器。
不要改用通用 LoraLoaderModelOnly，也不要串接 Sage、PyTorch selector 和 Sol。
具体文件及限制见 04-long-video 的 README 和
加速器接线说明。
能做什么
H3 生成和参考控制
- 文生视频 T2VA
- 首帧图生视频 I2VA
- 尾帧生成 L2VA
- 首尾帧生成 FL2VA
- 单张或多张参考图 Ref2VA
- 参考视频、参考音频和混合参考
- 原生视频与音频联合生成、解码和保存
- H3-World 首帧 I2VA：用 WASD 和 IJKL/F 时间线控制人物与镜头
音频和口型
- 原声锁定、参考音色、对白和音轨混合
- Vocal Lock：用独立人声驱动画面，完整歌曲只在最终成片混入一次
- Audio Refine：可接 Turbo、PDD、EAV、Prompt Relay 和长视频路线
- 本地 ASR、说话人和 SyncNet 检查工具
项目里的 32 秒 Vocal Lock V3 样片已经完成五镜头串行生成、逐镜口型检查和真人完整观看。用户最终反馈为“32秒这个已经没问题了，完美”。这只代表该样片通过，不代表所有素材都会自动得到同样效果。
长视频
- 多关键帧和分段续写
- 一次排队、节点内串行生成
- 断点恢复和 accepted manifest
- Native Masked Context Plan B
- 可选 Color Match，默认开启，用于减轻分段接缝颜色跳变
- 显式
accepted_picture_low_context_v1：上一段实际成片末39帧经原resize与视频VAE，作为下一段LOW视频参考；HIGH、音频、4+4及3D放大器不改。旧JSON默认仍为独立low-x0续接。见原接缝示例与防回归说明及新增Dance单项GPU对照／人审接受记录。两份指定样例接受，不是任意素材、Depth或长时长保证。
视频扩画
- 支持上下左右自定义扩边，也可以按目标宽高比和锚点自动计算画布
- 新节点默认
joint_decode（联合解码）：整幅画面一起经过 VAE 解码，避免硬贴原片造成的轮廓断口。原片区域也会重建，细节可能变化，不能保证像素不变 - 可选
preserve_source（保留原片）：在有损编码前精确回贴原片像素，但新增画面与原片之间可能仍有接缝；两种模式都保留已有原音轨 - 先生成第一个窗口作为候选，人工审图并确认后才继续整条视频；接续和保存沿用候选模式，不会偷偷换模式。旧候选没有模式记录时仍按保留原片处理
- 支持逐镜切点、逐镜提示词、区域提示词和原片人物框审计。Color Match 默认开启，但只在保留原片模式处理扩区，并在切镜时重置；联合解码跳过边缘修色和几何校正
- 取消或重启后可以复用已完成窗口，但需要核对素材、模型和运行配置；更新 Core、KJ 或节点代码后，不能直接跳过缓存身份检查
- 最终 MP4 使用更保守的全帧内 H.264，文件会比常规编码大，但可避开本机已经复现的多线程解码坏帧问题
- 可在成片后单独接 DLSS-NR 2x；超分会改变整幅画面质感，所以仍需要另行看片
扩画不需要新的专用模型或转换权重，直接使用现有 H3 FL2VA 主模型、Qwen3-VL、视频 VAE 和音频 VAE。生成工作流还需要单独安装 ComfyUI-KJNodes，当前测试路线固定使用 Stock20 与 KJ 的 H3 低显存 Attention/FFN。Turbo、SPEED、SLA、OpenVDN、FastH3 等组合暂不允许直接叠加；兼容审计工作流会提前说明冲突。已实际生成完整 32 秒并核对原音轨，但部分扩画区仍有瑕疵；本次按已知限制发布，不等于全部素材画质通过。详见 扩画工作流说明。
加速和成片修复
- OpenVDN DMD 8 步 / Stage B 50 步
- PDD、SLA、SPEED、FastH3 VSA、Enhance-A-Video
- 稳定双时钟采样节点内置可选
beta57调度器，无需安装 RES4LYF；默认仍是native_flow - DLSS-NR 图片/短视频帧/长视频文件超分，以及 FlashVSR、RealBasicVSR、RAFT、Skin Finish
- Face Refine Window：只生成选中的连续坏脸时间窗，预览后由人决定接受或拒绝
- NVIDIA H3 + LTX-2.5 两阶段超分实验路线
带 Advanced 或 EXP 的功能需要使用对应工作流。不同加速方案通常不能叠加；节点会尽量提前拒绝明显冲突。
安装
ComfyUI Manager
在 Manager 中搜索 MiniMax H3 Audio T8，安装后完全重启 ComfyUI。
手动安装
cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8
本节点依赖较新的 ComfyUI 原生 MiniMax H3 支持。如果节点全部变红、工作流提示缺少节点，先更新 ComfyUI 本体、前端和 Manager，再彻底退出并重新启动。
模型放哪里
常用目录如下：
| 模型 | ComfyUI 目录 |
| --- | --- |
| H3 主模型 | models/diffusion_models |
| Qwen3-VL 文本编码器 | models/text_encoders |
| 视频与音频 VAE | models/vae |
| Turbo、PDD、SLA 等 LoRA | models/loras |
| OpenVDN 完整模型包 | t8star/Vdn-Minimax-H3-Comfy；仓库已按 models 下的正确目录整理 |
| H3-World 动作 LoRA | t8star/Minimax-H3-World-Comfy 已按 models/loras/minimax/H3-World 的相对目录整理；原始来源为 DANNY621/H3-World |
| FlashVSR / RealBasicVSR | models/upscale_models 或工作流注明的专用目录 |
| DLSS-NR v1.3 外部运行时（不是模型） | models/DLSS-NR/1.3（用户自行取得，节点不下载或分发） |
| 人脸、光流和分割模型 | 工作流或对应文档注明的目录 |
不同 H3 基模和 LoRA 不是随便混用的。文件名相近也不代表结构兼容。
Face Refine Window：只修选中的时间窗
examples/workflows/06-face-refine 里有三份 2026-09-05 工作流：
Window_Manual_Review：一次处理一个窗口，先预览，再手工接受或拒绝。Window_Studio_Serial：把多个窗口的决定写入可恢复清单；仍然一次只跑一个 H3 任务。Window_Studio_Compose：所有决定完成后，或提交后在保存前崩溃时，不加载 H3，直接从清单重建成片。
这条路线不会把整段正常人脸一起重画。你先用 0 基、闭区间帧号填写坏脸范围，例如 0-23；节点会补足
H3 所需的连续上下文，但上下文和边缘补帧永远不会自动进入最终接受区。预览和拒绝会逐像素返回原片；接受
必须显式勾选确认。最终视频始终使用完整原始音轨，窗口生成出来的音频会被丢弃。
Studio 版会把每个窗口的接受/拒绝结果按顺序保存。接受后的窗口不能回退或重复执行；崩溃后会从第一个未决定 窗口继续。同一项目有进程级锁，避免两个任务同时抢显存。它不会替你判断哪张脸更好，也不会自动接受候选。
生成工作流提供可选的 MiniMaxH3FaceRefineSamplerMaskPatchV11T8Advanced，enabled 默认关闭，
关闭时 MODEL 和 LATENT 原样直通，保持原有路线。手动开启才应用上游
ComfyUI-H3-FaceRefine v1.1.1
的采样遮罩修正：视频 noise_mask 只交给采样器，保留帧按当前 sigma 重加噪；锁定音频的
mask 条件仍原样送入 H3。开启时会核对 LATENT 中的音视频遮罩哈希和去噪报告，拒绝遮罩不匹配、
不兼容模型或冲突 MODEL patch。Parity、Window Manual 和 Studio Serial
工作流已接好；Compose-only 不加载 H3，所以不需要该节点。这个修正不需要新模型、pip 包或外部程序。
实测边界：RTX 4060 Ti 16GB 上，2 GiB 预留配置完成 90 帧、124 帧各 3 次冷启动和 3 次暖启动，再连续 执行 3 个窗口，共 15/15 次成功；最低空闲显存 678 MiB，最终进程 private 增量分别为 40.74、34.20 和 35.12 MiB，没有超过预注册的 256 MiB 阶梯门。这个数字只适用于该机器和固定工作流，不代表所有 16GB 显卡都安全。当前功能仍标为 Advanced EXP；首个真实坏脸样本和音频机械检查已通过，但多素材人脸质量仍需 用户看完整盲测后决定。v1.1.1 修正后的同素材 90 帧实测也已严格串行跑通，原始 PCM 完全一致、最低空闲显存 717.8 MiB。2026-09-06 盲评中，用户认为 B（旧路线）好一点：总体、前 24 帧五官、网格/发虚均选 B， 时序与接缝持平。因此保留旧路线为默认，新修正仅供手动对比；本次结果不代表所有素材，也不能证明整个人脸修复功能的画质已经全面验收。
DLSS-NR：Windows RTX 可选后处理
examples/workflows/25-dlss-nr 提供运行时检查、图片、短视频帧序列和
文件视频四份独立工作流。默认使用 v1.3 Standard + 2x。
DLSS-NR 不需要新的 PyTorch / safetensors 模型，但需要外部程序。 请从
video2dlssnr v1.3 官方 Release
取得 video2dlssnr_release.zip 完整包；不要用 light 包，也不需要安装上游的 ComfyUI 节点包。本项目
不会自动下载、安装或分发其中的 EXE 和 NVIDIA DLL。
运行条件：
- Windows 10/11、NVIDIA RTX 显卡、NVIDIA 驱动 616.56 或更新版本
- 图片和帧序列路线不增加新的 pip 依赖；使用 ComfyUI 已有的 Torch、NumPy 和 PyAV 环境
Video File路线还要求ffprobe可在PATH中找到；通常安装 FFmpeg 后即可获得- 外部运行时及 NVIDIA 的适用许可仍需遵守；节点不再要求勾选接受，环境检查通过即可运行
正确目录结构如下。保留完整 ZIP，把其中 out 目录的四个文件复制到这里的 bin，并把本仓库的
examples/runtime-manifests/dlss-nr-v1.3.json
复制为 t8-runtime-manifest.json：
ComfyUI/models/DLSS-NR/1.3/
├── t8-runtime-manifest.json
├── video2dlssnr_release.zip
└── bin/
├── video2dlssnr.exe
├── nvngx_dlss.dll
├── nvngx_dlssnr.dll
└── nvngx.dll_dlssnr.dll
先运行 Runtime_Audit。节点会校验完整包及已解压文件的哈希、驱动、GPU 映射和真实 feature probe；
只有显示 READY 后才运行超分。v1.2 只允许 1× NR-only，默认 2× 工作流必须使用 v1.3。
三类固定素材的四路盲测均通过非退化门。人审结论是不同高清方法会带来不同皮肤和质感，没有一种在所有 素材上一定更好。因此这里的Standard只是推荐起点；超分不修复源片已有的身份、口型或真实纹理问题。
H3-World：WASD 人物与镜头控制
上游项目：Danzer1xxxxChan/H3-World ·
ComfyUI 模型包：t8star/Minimax-H3-World-Comfy ·
原始模型：DANNY621/H3-World
首版只做一件明确的事：输入一张首帧图，以 832×480、124 帧、24fps 生成一条约 5.17 秒的 I2VA，
并让 37 个潜空间时间点分别接收人物与镜头动作。预设包括前进、后退、左右移动、上下倾斜、左右摇镜和
快速摇镜；custom 可以用 JSON 分段组合动作。非 832×480 的首帧会等比覆盖后居中裁剪，不会直接拉伸。
下载 LoRA：
hf download t8star/Minimax-H3-World-Comfy --include "loras/**" --local-dir ComfyUI/models
下载后的完整路径应为
ComfyUI/models/loras/minimax/H3-World/step-10000.safetensors。模型包中的文件与上游固定 revision
逐字节一致，SHA-256 为
DDD9187B920B1E52C2D090F4E264FD83D8D433EFC2A5B159E58883AEAF96E526；我们没有转换、合并或量化它。
这个 LoRA 已经是 ComfyUI 可直接加载的 104 对 A/B 权重。工作流还需要现有的完整
minimax_h3_fl2va_int8_convrot.safetensors、Qwen3-VL 编码器、视频 VAE 和音频 VAE。它不增加新的
pip 依赖，但最终安全保存要求 ffmpeg 可在 PATH 中找到。请直接使用
examples/workflows/26-h3-world 的工作流，不要叠加 OpenVDN、SLA、
VSA、Sol-Attn、BlockCache 或另一个模型/Attention 接管节点。
固定停车场样本的匿名盲测已通过：用户正确识别出持续前进的 H3-World 版本，确认动作稳定、两边声音 正常，画面质量持平。该结论支持这条固定合同转为正式 Advanced 功能，不代表任意人物、动作或显存配置 都能得到相同结果。
OpenVDN：推荐的 8 步路线
新版 Core 兼容与二次采样（已在 v1.75.0 发布）
这一轮解决的是两件事：更新 ComfyUI 后仍能使用原有节点，以及让 VDN 先生成小图视频、再做潜空间放大和第二次采样。现有单采工作流继续保留，不需要重新转换底模。
- 全局 Sage 开关可以保留。
--use-sage-attention是默认注意力后端，不等于另一个节点接管了 VDN。VDN 自己需要的注意力仍按它的算法执行，不会因此变成 Sage 版 VDN。 - Core 自带的稀疏注意力节点与外部 Sol 插件不是一回事。 对已识别的 Core 补丁，新兼容代码只在 VDN 分支上避让它；其他 MODEL 分支不变。未知的模型或注意力替换仍会提示冲突，不能随意叠加。
- 二采有两条路线。 一条是 VDN 8 步 → 学习型 2x 潜空间放大 → VDN 4 步；另一条在第二次采样改用独立原生 H3 分支和新版 EMA B。不要给 VDN 分支再加通用 EMA LoRA。
- 默认保留第一遍的声音。 第二遍主要细化画面，不重新生成音轨。两条路线都需要已有的
models/latent_upscale_models/minimax_h3_latent_upscaler_3d_fp16.safetensors；只有原生 H3 二采另需models/loras/minimax_h3_turbo_v4_step600_ema_comfyui_B.safetensors。保存需要 FFmpeg，没有新增自动下载或安装步骤。
九类输入 × 完整/pruned 底模的二采测试已跑完；0.52MP 对照中，古典音乐、人声和两边口型正常，I2VA 环境声也无杂音，两组画面都差不多。四张二采工作流已交付到项目和用户目录，使用方法见 二采工作流说明。独立 Core/VDN 版本通过 2435 项完整回归、真实模型 DynamicVRAM 二采和解包导入，已在提交 3769d70 发布；本轮未完成的扩画修订不在该次发布中。二采不保证总比单采更清楚，也不保证所有 16GB 显卡和插件组合安全；兼容范围中列出实际测试边界。
完整模型包：t8star/Vdn-Minimax-H3-Comfy
模型仓库已经按 ComfyUI 的 diffusion_models、text_encoders 和 vae 目录整理好。获得模型访问权限后，可以直接下载到 ComfyUI/models：
hf auth login
hf download t8star/Vdn-Minimax-H3-Comfy --local-dir ComfyUI/models
然后打开 examples/workflows/10-speed 中的正式 OpenVDN 工作流。目前提供：
- T2VA
- I2VA
- L2VA
- FL2VA
- 单图 Ref2VA
- 多图 Ref2VA
- 参考视频加原音轨
- 独立参考音频
- 首帧加参考音频 Hybrid
OpenVDN 上游公开说明的是 T2VA；其他模式是本项目利用 ComfyUI 原生 H3 条件布局实现并真实跑通的扩展。
完整底模和 pruned 底模都可以用
正式工作流默认使用：
models/diffusion_models/minimax_h3_fl2va_int8_convrot.safetensors
这是完整、非 pruned、AdaLN 输入宽度为 2688 的底模。安装新版模型包后，也可以在同一工作流里改用：
models/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors
Composer 会读取底模里的 adaln_t_table，按内容 SHA 自动选择配套的 curve-projected Turbo adapter；用户不需要再接一个 LoRA 节点。该适配器保留 208 个原样可用的目标，并把 51 个 AdaLN 目标转换成 8 维 LoRA 加 51 个偏置残差，共实际应用 310 个补丁。完整底模仍走原生 2688 维 adapter。
兼容是按曲线表内容识别，不是只看文件名。目前支持模型包中的 FL2VA pruned INT8/ConvRot（同曲线表的 FL2VA pruned FP8 也通过静态签名门）；未知 pruned/curve-basis 模型仍会在采样前给出明确错误，避免套错适配器后静默生成。
已完成的验证
使用完整底模，项目在同一台 RTX 4060 Ti 16GB 上严格串行测试了 I2VA、L2VA、FL2VA、单图、双图、视频加音频、独立音频和首帧加音频共八条路线。随后又用 pruned INT8 底模在 320×192×39 下串行复跑了 T2VA 和同样八条多模态路线。每条 pruned 测试都完成：
- 800 个 OpenVDN 分支张量
- 104 个 default adapter 目标
- 259 个逻辑 turbo adapter 目标，其中 51 个 AdaLN 目标带独立偏置残差，共 310 个实际补丁
- 8 个 Euler/native-flow 步骤，video/audio shift 为 12/3
- 原生 H.264 视频、AAC 音频和联合严格解码
- 运行日志中
ERROR lora = 0
这些结果证明工作流和两类底模都能正确组合，不等于所有提示词、参考图、声音或显卡都已经通过画质验收。完整底模八条短测试最低剩余显存为 535–890MiB；pruned 九条短测试中最低只有 290MiB，仅 T2VA 和 I2VA 超过本项目的 512MiB 余量门。16GB 显卡仍应一次只跑一个任务，并根据实际占用降低分辨率或帧数。
常见问题
节点全红或找不到
先更新 ComfyUI、前端和 Manager，再完全退出重启。只更新本插件通常不够。
一运行就显存不足
先降低分辨率、帧数和参考素材数量，关闭同时运行的其他生成任务。16GB 显卡不要并发跑两条 H3。
声音变成噪音或音量异常
先检查工作流指定的 sampler、scheduler、步数、video/audio shift 和 LoRA。不要把通用 EMA、Ref2VA LoRA、OpenVDN turbo 或其他加速 LoRA随意互换。
OpenVDN 提示 AdaLN 不兼容
先确认新版模型包中存在 stage-dmd-step-250/adapters/turbo_pruned_curve_fl2va/adapter_model.safetensors。如果文件存在仍报错，说明所选 pruned 底模的曲线表与已支持版本不同；换用模型包中的 FL2VA pruned INT8，或暂时换回完整 minimax_h3_fl2va_int8_convrot.safetensors。不要靠改文件名绕过签名检查。
能不能叠加 SLA、VSA、Sol-Attn 或 BlockCache
OpenVDN 自己接管模型分支和 adapter。不要再叠加其他 MODEL 或 Attention 接管节点；这不是“多加一个会更快”，通常只会造成冲突或错误结果。
文档
许可证
节点源码使用 GPL-3.0-or-later。
模型有各自的许可证。MiniMax H3 及其衍生模型遵循 MiniMax H3 Community License Agreement；协议定义的适用地区不包括欧盟、英国、韩国和美国。下载、运行或再分发模型前，请阅读模型仓库中的完整协议和 Acceptable Use Policy。
本 GitHub 仓库不包含模型权重。OpenVDN 模型包在 Hugging Face 单独提供，并保留原始许可、NOTICE、来源和修改说明。