Minor update in V3.4

recontruct the attention part of model subgraph.
add sol attn into workflow.
recontruct the front panel of model subgraph.
add download link of lightx2v lora in markdown.
renew the attention combination.
use Chunk FeedForward to replace Low VRAM attention.

Remark:

Make sure you had SWITCH OFF ACC Lora if you chose normal Turbo Lora, vice versa.
Beware you had to choose either (Sage attention + pytouch attention) or kitchen attention + Specturn, Sage attention and kitchen attention COULD NOT use together, SWITCH OFF those which are not using.
disable the int8_qk in Sol attn patch if you feel too much noise.
Sol attention/Fused Modulation/Chunk Feed Forward/Spectrum could be compatible with Sage attention and Kitchen attention.
Sol attention now support 30 series RTX display card

minor update in the model subgraph V3.3

add one more lora loader for twin turbo lora.(step 600 ema turbo lora + 4 to 8 step hybrid turbo lora)

add button in subgraph panel to enable option for turbo lora.

add Sage attention node and switch for user option.

keep acc turbo lora and switch for user option.

Cancel lora stack and add four style lora loader in subgraph panel.

Update quick guide/ download link in workflow.

* Turbo Lora divided into 2 pieces, one is step600 ema, another is hybrid 4-8step r48 which are ok to compatible with Dasiwa hybrid model/ B25-49 hybrid model/ official model (hybrid means both FL2VA and REF2VA could be use).

* Beware you had to choose either Sage attention + pytouch attention or kitchen attention + Specturn, SWITCH OFF those which are not using.

* Make sure you had SWITCH OFF ACC Lora if you chose normal Turbo Lora.

* ACC LoRA is an acceleration LoRA developed by Alibaba for MiniMax H3. It generates higher-quality videos but currently only supports 8 steps.

* You can choose to enable Turbo LoRA using either ACC LoRA or a standard Turbo LoRA. Alternatively, you can disable Turbo LoRA entirely and use 20–25 steps with Spectrum.

* Spectrum skips Transformer calculations, speeding up generation by about 30%, but it may negatively affect ACC LoRA.

* If you notice prompt bleeding/failing or limb distortion, you can switch to ACC LoRA combined with ComfyUI Kitchen Attention and disable Spectrum.

* ComfyUI Kitchen Attention is optimized for INT8. If you are not using INT8, you can switch back to Sage Attention, as the generation times are quite similar.

* Please note that if you are using Dasiwa's Turbo 4-step/8-step model, you don't need to add a Turbo LoRA. You can simply turn off the 'Enable Turbo LoRA' option.

---

* Turbo Lora分成二部份, 一個是 step600 ema, 另一個是Hybrid 4-8 step r48, 串連在一起使用, 兩者都適配Dasiwa/ B25-49 混合模型或官方模型(混合模型是REF2VA和FL2VA都可使用)。

* 留意你只能選擇用Sage attention + pytouch attention 或 kitchen attention + Specturn, 不用的選項務必關上。

* 如果選用了一般 Turbo Lora, 記得關上ACC Lora的選項。

* ACC Lora是由阿里巴巴開發, 針對minimax H3 的加速LORA, 生成影片的品質較好, 但目前只有8 steps。

* 可以選擇開啟 Turbo Lora, 使用ACC Lora或一般turbo Lora, 甚至關閉Turbo Lora 用20-25 steps + Spectrum。

* Spectrum 跳過Transfromer計算, 加速生成時間約30% 但有機會影響ACC Lora。

* 如果發現prompt失效, 肢體變形, 可以選擇使用acc lora + comfyui kitchen attention, 不使用Spectrum。

* comfyui kitchen attention是針對INT8優化, 如果不是使用INT8, 可以轉回sage attention, 但生成時間其實差不多。

* 特別注意的是, 如果使用Dasiwa的Turbo 4 steps/8 steps model, 是不需要加Turbo lora, 你可以把enable turbo lora 這option 關閉。

BTW, I really don't understand what is the examination and approval Guidelines of CIVITAI...

XXX and R? REALLY?

3.2 updated

minor update in the model subgraph

Added ACC LoRA loader.
Subgraph: Added an ACC LoRA toggle to the panel.
Subgraph: Added a Turbo LoRA toggle to the panel.
Updated the Quick Guide inside the workflow.

小更新:

增加ACC lora loader
subgraph 面板增加ACC lora開關
subgraph 面板增加Turbo lora開關
Update Workflow內的Quick guide

* ACC LoRA is an acceleration LoRA developed by Alibaba for MiniMax H3. It generates higher-quality videos but currently only supports 8 steps.

* You can choose to enable Turbo LoRA using either ACC LoRA or a standard Turbo LoRA. Alternatively, you can disable Turbo LoRA entirely and use 20–25 steps with Spectrum.

* Spectrum skips Transformer calculations, speeding up generation by about 30%, but it may negatively affect ACC LoRA.

* If you notice prompt bleeding/failing or limb distortion, you can switch to ACC LoRA combined with ComfyUI Kitchen Attention and disable Spectrum.

* ComfyUI Kitchen Attention is optimized for INT8. If you are not using INT8, you can switch back to Sage Attention, as the generation times are quite similar.

* Please note that if you are using Dasiwa's Turbo 4-step/8-step model, you don't need to add a Turbo LoRA. You can simply turn off the 'Enable Turbo LoRA' option.

---

* ACC Lora是由阿里巴巴開發, 針對minimax H3 的加速LORA, 生成影片的品質較好, 但目前只有8 steps.

* 可以選擇開啟 Turbo Lora, 使用ACC Lora或一般turbo Lora, 甚至關閉Turbo Lora 用20-25 steps + Spectrum.

* Spectrum 跳過Transfromer計算, 加速生成時間約30% 但有機會影響ACC Lora.

* 如果發現prompt失效, 肢體變形, 可以選擇使用acc lora + comfyui kitchen attention, 不使用Spectrum

* comfyui kitchen attention是針對INT8優化, 如果不是使用INT8, 可以轉回sage attention, 但生成時間其實差不多

* 特別注意的是, 如果使用Dasiwa的Turbo 4 steps/8 steps model, 是不需要加Turbo lora, 你可以把enable turbo lora 這option 關閉

And this is the big time difference using turbo lora (8 steps) or not (25 steps) in 15s video generation.

This is the time difference with and without Spectrum in 8 steps acc turbo lora generate 15s video.

This is the time difference between ACC lora (8 steps) 310.47s and normal turbo lora (8 steps) 277.18s generate 15s video.

⛩️💀 DaSiWa MiniMax H3 💀⛩️

https://civarchive.com/models/2877206/dasiwa-minimax-h3?modelVersionId=3275408

All credit belongs to the node creator!

Special Thanks:👏

3.1 updated

minor update in the model subgraph

replaced sage attention to comfyui kitchen attention.
add several control button on the model subgraph panel
add low VRAM attention option for the Low VRAM user (12GB or below)
Larger head_chunks save VRAM but increase generation time; typically set to 2, 4, or 8.

3.1 版本更新

model subgraph微調：對model subgraph進行了微幅更新。
更換注意力機制：將 Sage Attention 更換為 ComfyUI Kitchen Attention。
新增控制按鈕：在model subgraph面板上增加了數個控制按鈕。
優化低顯存設定：為低顯存用戶（12GB 或以下）新增了低顯存注意力機制選項。
head_chunks數值越大, VRAM佔用越少, 但生成時間越久, 一般可設定成2, 4, 8。

請參考下方關於 spectrum的訊息:

More information of spectrum could be found below:

https://comfyui-wiki.com/en/news/2026-08-03-comfyui-spectrum-minimax-h3

https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

在介面左下角進入terminal

安裝完成後重啟

all credit belongs to the node creator👏

3.0 updated

Add 2 button in model subgraph, lora loader and new user guide.

replaced the Ksampler to Minimax H3 Director.

Once I said, I hate those workflow too big like a trash hole, so I kept it clean and tidy.

This Minimax H3 Director is so powerful, included all the function orginial and plus, worth to try.

V3.0 額外加入了lora loader, model subgraph中多加了二個button, 加入了user guide

把原來的Ksampler更換成Minimax H3 Director.

如我先前所言, 我討厭那些又大又亂像垃圾崗一樣的workflow, 所以我盡量保持簡潔

這Minimax H3 Director實在太強大, 包含了原本所有功能外還追加其他功能, 值得一試

Details please refer to the github link below:

詳細請參閱下方的連結:👇

https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director

以下是Node Creator 在github的原文使用指引:

A quick guide copy from the Node creator in github:👇

Default sampling

Canvas default 0.4MP 16:9 (864×480), 5s / 124 frames @ 24 fps (17k+5 grid)
25 steps, res_multistep + simple, CFG 1.0
Sigma shift: video 12 / audio 3

First/last frame (fl2v) — short guide

Set task type to First/Last Frame to Video (fl2v)
Click Add group, upload start and/or end frame (end-only is allowed)
Adjust duration on the shot card or timeline; write mid-shot motion / camera / transition in the prompt
Queue; with multiple groups, use Run select to sample only some of them

Reference groups (r2v) — short guide

Set task type to Reference to Video (r2v) (ref2va UNET + audio_vae)
Click Enable common params (collapsed/off by default); upload shared refs/audio and write a common prompt (e.g. character lock / subject_definitions); when enabled it is concatenated with each group prompt
Click Add material group; write per-shot prompts and optionally add group-only assets (same slot overrides common)
In prompts use <Picture N> / <Video K> / <Audio J>, or type @ (with common params on, picker includes common + group assets)
Timeline previews group duration/thumbs; Run-select stays in sync with group checkboxes

Source video (v2v / rv2v) — short guide

Choose v2v or rv2v, upload a source video and split segments (cut / equal-split / smart split)
Write a prompt per segment; the source clip is bound as <Video 1> automatically
For rv2v, optionally add reference images / audio; audio mode can be generate / source / mute

Refine / upscale — short guide

Add MiniMax H3 Director Refine and wire refine into Director refine. Leave it unconnected for the original single pass
mode=refine: same-resolution second sample. mode=upscale: enlarge to a target canvas then second-sample. mode=latent_upscale: enlarge H3 video latent only (no second sample). Resolution widgets appear for upscale / latent_upscale (follow Director, aspect + megapixels, or custom W×H). Director canvas is the first-pass size; Refine target is the enlarge size
passes: refine rounds, default 1, max 9999. In upscale mode only the first round enlarges; later rounds stay on that canvas. latent_upscale does not sample
Optional refine_model (second-pass UNET); unwired uses the Director model. Typical: Turbo LoRA on pass 1, a clean / other LoRA UNET on refine
Director images is the refined clip; images_pre_refine is the first pass before upscale (for A/B). source_images is still the timeline source, not the first-pass generate
Second sample uses SIGMAS: wire BasicScheduler or ManualSigmas into Refine sigmas
fl2v skips refine by default (protects pinned first/last frames); turn off skip_fl2v on Refine to include those shots
Upscale default is h3_latent: pick the 3D weights in Refine (dropdown under upscale_method; also shown for mode=latent_upscale). Put the file in ComfyUI/models/latent_upscale_models/. lanczos can take optional upscale_model (RealESRGAN etc.); or use nvidia_rtx_vsr
Segment export with passes>1 also writes seg_XXXX_pN.mp4 per round; export-all still only keeps first-pass and the final clip

Example: example_workflows/minimax_h3_director_二采_加速.json

External multi-group wiring

Mirror the two official conditioning nodes and feed multi-group batches into the Director:

Add MiniMax H3 Director Group (Image to Video) or (Reference to Video)
Wire per group: prompt / duration_sec; I2V family uses first_frame / last_frame (none=t2v, first only=i2v, last only or both=fl2v); R2V uses Autogrow slots (same as official Reference to Video: images ≤9, videos ≤3, audios ≤3). Output size is set on the Director
Batch with Director Groups Combine (Autogrow slots, same UX as official Reference to Video) → Director i2v_groups / r2v_groups; a single group can connect to the Director directly
Match task_type to the port (t2v/i2v/fl2v ↔ i2v_groups; r2v ↔ r2v_groups); do not connect both ports at once
When linked, graph wiring overrides UI cards (external priority); Run-select still applies by group index

默认采样参数

画布默认 0.4MP 16:9（864×480），5 秒 / 124 帧 @ 24 fps（17k+5 网格）
25 steps，res_multistep + simple，CFG 1.0
Sigma shift：video 12 / audio 3

首尾帧 fl2v 用法摘要

任务类型选 「首尾帧生视频 (fl2v)」
点击「添加一组」，上传首帧和/或尾帧（可只传尾帧）
在镜卡片或时间轴上调整时长；提示词写中间运动 / 镜头 / 过渡
Queue 生成；多组可勾选「选择运行」只跑部分组

参考主体 r2v 用法摘要

任务类型选 「参考主体生视频 (r2v)」（需 ref2va UNET + audio_vae）
点击 「启用公共参数」 展开面板（默认折叠/关闭）；上传共用参考图/音频并写公共提示词（如角色锁定 / subject_definitions）；启用后会与每组提示词拼接
点击「添加素材组」；每组写分镜提示词，可按需再挂本组独有素材（同槽位覆盖公共素材）
提示词中用 <Picture N> / <Video K> / <Audio J>，或输入 @（启用公共参数时可引用公共 + 本组素材）
时间轴可预览各组时长与缩略图；「选择运行」与素材组勾选同步

源视频 v2v / rv2v 用法摘要

选 v2v 或 rv2v，上传源视频并分段（切分 / 均分 / 智能分割）
每段写提示词；系统自动将源片段绑定为 <Video 1>
rv2v 可额外上传参考图 / 参考音频；声音模式可选生成 / 原声 / 静音

二采 / 放大 Refine 用法摘要

添加 MiniMax H3 Director Refine，把 refine 接到导演台 refine 口。不接则仍是原来的一采
mode=refine：同分辨率再采一遍（精修）。mode=upscale：先放大到目标画布再二采。mode=latent_upscale：只放大 H3 视频 latent，不再二采。分辨率控件在 upscale / latent_upscale 时显示（可跟随导演台、按比例+百万像素，或自定义宽高）。导演台是一采分辨率，Refine 目标才是放大后的宽高
passes：精修次数，默认 1、最多 9999。upscale 只在第 1 次放大，后面都是同分辨率精修；latent_upscale 不二采
可选接 refine_model（二采 UNET）；不接则用导演台主模型。适合一采挂 Turbo LoRA、二采卸掉或换另一套
导演台 images 是二采后成片；images_pre_refine 是一采、放大前的画面，便于对比。source_images 仍是时间轴原片，不是一采结果
二采用 SIGMAS：把 BasicScheduler 或 ManualSigmas 接到 Refine 的 sigmas 口
fl2v 默认跳过二采（保护钉死的首尾帧）；关掉 Refine 上的 skip_fl2v 才会采
upscale 默认 h3_latent：在 Refine 节点里选 3D 权重（upscale_method 下方下拉框；mode=latent_upscale 时同样出现）。权重放 ComfyUI/models/latent_upscale_models/。lanczos 可另接 upscale_model（RealESRGAN 等），不接则纯插值；也可改 nvidia_rtx_vsr
「分段导出」且 passes>1 时，每轮会另落 seg_XXXX_pN.mp4；「全部导出」只出一采和终稿

示例：example_workflows/minimax_h3_director_二采_加速.json

外部多组接线（第三方节点接入）

对齐官网两个 conditioning 节点，把扩写 / 抠图 / Load Video 等处理结果以多组形式送进导演台：

添加 MiniMax H3 Director Group (Image to Video) 或 (Reference to Video)
按组接线：prompt / duration_sec；I2V 系接 first_frame / last_frame（无帧=t2v，仅首=i2v，仅尾或首+尾=fl2v）；R2V 为 Autogrow（同官方 Reference to Video）：接图/视频/音频会自动多出空口（图≤9、视频≤3、音频≤3）。输出宽高在导演台统一设置
多组：用 Director Groups Combine（Autogrow：接满最后一个口会自动多出新口，同官方 Reference to Video）→ 导演台 i2v_groups / r2v_groups；单组可直接把 group 连到导演台
导演台 task_type 与口一致（t2v/i2v/fl2v ↔ i2v_groups；r2v ↔ r2v_groups）；不要两口同时连接
连接后执行以图中接线为准（外部优先）；UI 卡片变淡，仍可用「选择运行」按组序勾选

This node made by AImixer from China, all credit to him👏

這強大的Node是由AImixer這中國開源大神製作, 感謝!

2.0 updated

Detail could be found in the link below:

新版本更新詳細可以參考下方連結

https://civarchive.com/articles/33990/minimax-h3-simple-workflow-multiref-for-low-vram-updated-20

Video Upscale simple workflow for low VRAM (Extreme FAST)

Low VRAM專用簡單影片放大Workflow (極快)

https://civarchive.com/models/2871874/video-upscale-workflow-for-low-vram-extreme-fast

How to merge videos with FFmpeg in ComfyUI Desktop (2026)?

如何使用FFmpeg在ComfyUi Desktop(2026)中合拼片段

https://civarchive.com/articles/34119/how-to-merge-videos-with-ffmpeg-in-comfyui-desktop-2026

This is a simple workflow for VoxCPM2, featuring a single-speaker setup and a multi-speaker setup.

這是作為輸助語音生成VoxCPM2的Workflow

https://civarchive.com/models/2865730/voxcpm2-simple-workflow-for-low-v-ram

I've also written an article about VoxCPM2—feel free to check it out if you're interested!

還有關於VoxCPM2的文章, 有興趣可以看看

https://civarchive.com/articles/34023/why-do-we-still-need-a-audio-generation-model-even-minimax-h3-could-produce-audio

Krea 2 Quadview Simple workflow

Qaudview might lock the character appearence during mulit reference in long video generation.

Krea 2 多角度參考圖 Workflow

多角度參考圖可以在影片生成時鎖死角色外表, 特別在長片分多次生成中作為參考圖使用

https://civarchive.com/models/2867878/krea-2-quadview-simple-workflow-for-low-vram

This is the simplest Minimax H3 Multi Ref workflow, and you can easily convert it to Image to Video or First-Last Frame to Video if needed. Detailed instructions can be found in the article linked below, so I will not repeat them here.

這是最簡單的 Minimax H3 Multi Ref Workflow, 如果想轉成image to Video 或 First Last Frame to Video就可以了

詳細說明可以參考下方連結的文章, 這兒就不重覆了

https://civarchive.com/articles/33959/for-someone-whom-want-to-use-minimax-h3-in-low-vram

📊 Hardware Requirements

8GB VRAM is usable but suboptimal.
12GB VRAM minimum is highly recommended.
32GB+ System RAM is required.
Note: VRAM will inevitably fill up, forcing the system to offload data to system RAM.

➡️ Model & Text Encoder Selection

Model: Choose the INT8 convrot version (Civitai Link). Avoid the GGUF model due to excessive quality degradation.
Text Encoder: Stick to INT8. Using GGUF doubles the generation time and carries a very high risk of OOM (Out of Memory) errors.
CLIP: GGUF is perfectly fine here; Q4KM is the top choice.
Model Versions:
- fl2v version: For Image to Video / First Image / First-Last Image tasks.
- ref2va version: For Multi-reference tasks (Video / Dialogue Audio / BGM Audio).

💡 LoRA Settings

Low VRAM Adjustments: You can consider 4 steps, but it is not recommended. It degrades the quality too much, turning H3's best features—the dialogue and sound effects—into complete garbage.
Recommended: Use 8 steps if your hardware allows.
768p Version: Even for the 4-step 768p model, the creator still recommends 6–8 steps. This specific 768p version delivers highly impressive image and audio quality.
Note: LoRAs are also split into fl2v and ref2v versions. Download the one that matches your specific workflow needs.

🗒 VAE Configuration

Video VAE: Consider using the INT8 version. It has a slightly smaller file size with negligible difference in final quality.
Audio VAE: The file size is already inherently small, so the version choice is inconsequential.

🛠 Workflow & Speed Optimization

Keep it Simple: Due to Low VRAM limitations, keep your main workflow minimal. Move non-generation nodes like Prompt Enhancer and Upscale into a separate standalone workflow to prevent OOM errors.
Prompt Enhancer (Chinese): Chinese language users can consider T8. It supports running local models natively without requiring an API.
Sage Attention: It is strongly recommended to integrate Sage attention into your workflow, as it drastically improves generation speed. A tutorial link is provided below.
Easy Cache Node: This node yields marginal performance gains when paired with 4–8 step LoRAs. However, it is retained in the workflow since it introduces zero visual degradation.

📥 Workflow Downloads & Resources

Sample Workflow: The attached file contains a simple workflow that I am currently using. Feel free to download and use it if needed.
This Workflow: This is the simplest Minimax H3 Multi Ref workflow. You can easily convert it to Image to Video or First-Last Frame to Video based on your needs.
Detailed Guide: For a comprehensive breakdown, please refer to the article linked below, so I will not repeat the details here.

🌟 Minimax H3 完整設定與優化指南

📊 硬體需求

8GB VRAM：勉強可用，但體驗不佳。
建議最低 12GB VRAM：強烈推薦的硬體門檻。
系統 RAM 32GB 以上：不可或缺的配置。
備註：不論如何調整，VRAM 遲早都會裝不下，系統最後還是會自動 offload（分流）到系統 RAM。

➡️ 模型與 Text Encoder 選擇

主要模型：請認準並選用 INT8 convrot 版本（Civitai 連結）。請別選 GGUF 格式，因為畫面劣化實在太嚴重。
Text Encoder：同樣請堅持使用 INT8。如果換成 GGUF，生成時間足足會慢上一倍，而且有極大機率會直接爆顯存（OOM）。
CLIP：這邊可以使用 GGUF，首選為 Q4KM。
模型版本對應：
- fl2v 版本：適用於 Image to Video（圖生影片）/ First Image（首幀）/ First Last Image（首尾幀）任務。
- ref2va 版本：適用於 Multi-reference（多重參考）任務（如：影片 / 對白音訊 / BGM 背景音樂）。

💡 LoRA 設定

低顯存（Low VRAM）調整：雖然可以考慮設定為 4 steps，但非常不建議。因為劣化太過嚴重，會導致 H3 最強大的音效與對白直接變成垃圾。
推薦設定：在硬體條件允許的情況下，最好使用 8 steps。
768p 版本：就算是 4 steps 的 768p 版本，原創作者的官方建議依舊是 6–8 steps。其中這個 768p 版本的畫質與聲音表現都非常出色。
備註：LoRA 同樣有區分 fl2v 和 ref2v 兩種版本，請務必根據自己的工作流需求對應下載。

🗒 VAE 配置

Video VAE（影片 VAE）：可以考慮使用 INT8 版本。檔案體積略小一些，且與原版相比幾乎沒有顯著差別。
Audio VAE（音訊 VAE）：因為這個元件本來的體積就非常小，所以使用什麼版本都無所謂。

🛠 工作流與速度優化

保持簡單：由於是低顯存（Low VRAM）環境，主要工作流請盡量保持精簡。像 Prompt Enhancer（提示詞增強）、Upscale（放大） 這類不涉及核心生成的節點，請獨立拆分到另一個工作流中，否則極度容易引發 OOM 錯誤。
Prompt Enhancer（中文使用者推薦）：中文使用者可以考慮使用 T8 這個 Prompt Enhancer，它支援本地模型運作，不需要另外串接 API。
Sage Attention：強烈建議在工作流中加入並啟用 Sage attention，這會讓整體的生成速度產生極其明顯的差距。如果需要教學教程，可以參考下方提供的連結。
Easy Cache 節點：這個節點在搭配 4–8 steps 的 LoRA 時提速效果雖然沒有很顯著，但因為它不會造成圖片劣化，所以依然選擇保留在工作流中。

📥 工作流下載與資源

附帶工作流：附件中包含了我目前正在使用的簡單工作流，有需要的人可以自行下載使用。
本工作流說明：這是目前最精簡的 Minimax H3 Multi Ref（多重參考）工作流，如果後續想要自行轉換成 Image to Video 或 First Last Frame to Video 也完全沒問題。
詳細教學：更多細節說明與圖文教學可以參考下方連結的文章，此處就不再重複贅述。

Files with SHA256: f80714ac46328b1d3264d1ed1d5855ceb2c46cd64b680935f82dd81c4babbd22