Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

MINIMAX H3 ULTRA WORKFLOW! (Patreon)

Published:
2026-08-06 16:32:32
Edited:
2026-08-07 13:15:44
Imported:
2026-08
Tags:

Downloads

Content

Hey everyone! I’ve just released a brand new video on MiniMax H3, and OH MY GOD… this is THE BIGGEST NEW MODEL RELEASE OF THE YEAR!MiniMax H3 is simply THE BEST OPEN WEIGHT AI VIDEO MODEL EVER! Capable of generating videos from text, images as well as from other references, it is INCREDIBLE!There is just nothing more to say, it's THE open-weight AI video model we were all waiting for...

You can check out the video right here: https://youtu.be/f35KjIMsTnU

I also created a 1-click installer for the Webui for my Patreon supporters for Local AND Runpod, that you can find right here: https://www.patreon.com/aitrepreneur/posts/165939863

Down below you will also find the MINIMAX_H3_ULTRA_WORKFLOW.json and theMINIMAX_H3_ULTRA_+TURBO-LORA_WORKFLOW.json (the one with Turbo lora parameters, so you can choose) for you to use!

As always, supporting me on Patreon allows me to keep creating helpful resources like this for the community. Thank you for your support - now go have some fun😉!

Files

Comments

x:)

5090 gets maxxxxxed out hard. Going to need to bring back Sli. With Sage plus Cache times are still very slow but this model is top notch

Ronald Tang

Is there a workflow or prompts to make the character sheets on krea 2? I saw that character sheets where used in the H3 video workflow

Mike

I've been using it for couple of days and it's truly impressive, just needed the iceing which is your beautiful workflows XD

ExusOne

The Turbo workflow is fantastic and super fast. After changing the text model to qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors, generating a 12-second video at 0.6 resolution takes about 6 minutes, instead of over 20 in the official workflow on my RTX 5060TI with 16 GB VRAM and 64 GB of RAM. Fantastic work.

Vato Vegas

how long did it take you to generate a 1080P video vs a 720P video with 5090?

Cristi Tanase

Can you please also add a Lora node as many "spicy" loras need one and some of us have no idea how to add such a node to the workflows, especially the Turbo ones. Also if you can do this for Wan 2.2 too it will be super super great. Thank you.

Matt

I haven't watched the video yet so don't bother to answer if already explained but is there any value in the heretic models Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot ?

Melissa DeLaFey

Hello Aitrpreneur! Can you please point us to where it is your V3 of the LTX-2.3 workflow? Thanks!

Aitrepreneur

There is no point no, don't waste your time

Aitrepreneur

I always add the power lora node to all workflows, always

Aitrepreneur

Yeah and the lora is not even done training yet, in a few days when it's done, it will be even better :)

Bryan S. Lee

I'm getting this when trying to use sage attention: # ComfyUI Error Report ## Error Details - **Node ID:** 2057:2051 - **Node Type:** SamplerCustomAdvanced - **Exception Type:** RuntimeError - **Exception Message:** RuntimeError: native MiniMax H3 module is missing required helpers: time_shift_slope

Christopher Fox

Do you happen to have a intro/guide for podman?

RavenLand

Tried the MINIMAX_H3_ULTRA_+TURBO-LORA_WORKFLOW.json on two computers, both give error "ModuleNotFoundError: No module named 'sageattention'" Your last models (krea2 edit, krea2 training) worked perfectly on my rtx3090 and rtx3060

Benny Hill

I get an error saying The workflow does not contain any output nodes (e.g. Save Image, Preview Image) to produce a result. Do i just need to learn more?

Aitrepreneur

You need to install sage attention first, you can use the sage attention installer from the post : https://www.patreon.com/aitrepreneur/posts/minimax-h3-is-ai-165939863 Put the OPTIONAL_SAGEATTENTION-INSTALLER.bat file inside the ComfyUI_windows_portable folder and run it, after the install it will create a special sageattention bat file launcher you can use to launch comfyui with sageattention enabled

Aitrepreneur

You need to install sage attention first, you can use the sage attention installer from the post : https://www.patreon.com/aitrepreneur/posts/minimax-h3-is-ai-165939863 Put the OPTIONAL_SAGEATTENTION-INSTALLER.bat file inside the ComfyUI_windows_portable folder and run it, after the install it will create a special sageattention bat file launcher you can use to launch comfyui with sageattention enabled

Aitrepreneur

You need to activate the workflow you want with the small button on the left, if it's purple it's not activated

RyperX

I ran into the problem that i needed to set degree and warmup_steps to 1 in the IMG Speedup Any idea why this is happening? You had much higher values in your Workflow.

Aitrepreneur

Yes, there was a spectrum node update, I updated the workflow redownload it from the post, it should make the generation even faster now

x:)

700 secs but with new 4 step lora like 70 secs for 5 secs and 240 secs for 10 secs. 720p havent tested 1080 as upscaling a lot more efficient

Bob

Try my settings: MiniMax H3 FL2VA MODEL -T2V + 12V - minimax_h3_fl2va_pruned_int8_convrot.safetensors MiniMax H3 REF2VA MODEL - REFERENCE VIDEO - minimax_h3_ref2va_pruned_int8_convrot.safetensors MiniMax H3 QWEN3-VL TEXT ENCODER - qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors minimax_h3_video_vae_fp16.safetensors minimax_h3_audio_vae_fp32.safetensors resolution keep specifically as in the graph: | megapixels | Aspect | Output (multiple=32) | |---|---|---| | 0.2 | 16:9 | 608 x 352 | | 0.3 | 16:9 | 736 x 416 | |<h2> 0.4| </h2><h2>16:9 | </h2><h2>864 x 480 | 480P | 0.5 | 16:9 | 960 x 544 | | 0.6 | 16:9 | 1056 x 608 | | 0.7 | 16:9 | 1152 x 640 | | 0.8 | 16:9 | 1216 x 672 | | </h2><h2>0.9 | </h2><h2>16:9 | </h2><h2>1280 x 736 | | 0.98 | 16:9 | 1344 x 768 | | 1.0 | 16:9 | 1376 x 768 | | 1.2 | 16:9 | 1504 x 832 | | 1.5 | 16:9 | 1664 x 928 | | 1.8 | 16:9 | 1824 x 1024 | | </h2><h2>2.0 | </h2><h2>16:9 | </h2><h2>1920 x 1088 | dont match to your image resolution, or better do opposite match the image first to the resolution. mine has naver face or body change if it still changes, i suggest try some loras, other loras find on civarchive.</h2>

Christian Wheel

Thanks for all your efforts! I tried the ULTRA+TURBO-LORA workflow with Sage Attention for a 10 second 1080p R2V with one video reference and it hangs at "Model Initializing" 0%. Two hours later it's still at 0%, GPU + VRAM usage is at 100% the whole time. RTX 5090 / 192GB RAM.

Henri

Hey sorry, I'm quite new to all this, and this is probably a dumb question. I've followed all the instructions and I'm trying to get the reference to video to work. However, I'm getting an error that I'm not seeing an obvious solution for? "1 Missing connection Required input slots have no connection feeding them. Video Combine 🎥🅥🅗🅢 - images" The "images" is clearly linked in the flowchart, so not I'm not sure why it's not reading.

flat

That’s exciting. I don’t know how much of your internal training workflow you want to make visible, but I wonder if a small “currently training” section on your landing page—with an estimated completion window or simple countdown—could become a great recurring feature. It would give supporters something concrete to follow and build anticipation around each new release without needing to reveal the technical details.

Aitrepreneur

This is not normal obviously but without more info, it difficult to say. Try restarting your pc first then try again, could just be a bug. Dm me if it keeps happening

Super Assetto Corsa

Video Reference do not work The last lines of the log: Output #0, f32le, to 'pipe:': [out#0/f32le @ 000002bf21fc1e80] Output file does not contain any stream Error opening output file -. Error opening output files: Invalid argument

LW

Hi K, thanks once again for keeping us up to date with the latest and greatest. Minimax H3 looks outstanding. In your video you recommend using ChatGTP to structure prompts. I (and maybe others) don't use ChatGTP and don't really want to start doing so. Therefore could I ask you to post the prompts you are using to get those great example videos, so I (and maybe others) can see what the structure of a good prompt actually looks like. Thanks.

Challenger Seventy

I get error " 1 node — 1 error 1 Missing connection Required input slots have no connection feeding them. Video Combine 🎥🅥🅗🅢 - images "

Buib Baum

I have the same Problem.

Buib Baum

You have to klick on the SUbgraph section double click and also activate the nodes there to.

Buib Baum

You have to klick on the Subgraph section double click and also activate the nodes there to.

x:)

I found a second card i can use to load the a text encoder but cant figure out how to send the text and vae to cuda 1 and main model to cuda 0 without errors. Maybe you can figure out

Benny Hill

Can you give me a picture of this magic button? I can't find it I ask copilot and i just go around in circles the groups are greayed out. On your videos the workflow looks diffrent

H-BOX

Little question guys. Was the quality overall similar when using and not using the turbo lora ?

Aitrepreneur

Just use another llm, the script can be used for any llm

Le Linh

First of all, thank you for your hard work as always. I have one problem with the I2V first-last frame that I couldn't solve. With last frame it is perfect. But once I remove the last frame, the video quality start going haywire really fast. It starts producing artifact and the end result look broken. I'm trying ways to write prompt and try to use resize node to resize original picture to megapixel, nothing work.

Le Linh

update: just try reference mode. same issue. Final video have a bunch of weird artifact, movement blur, looks very broken. Not sure where did I do wrong. I didn't change anything from the workflow

Le Linh

update: figured it out. The speed up node is the problem. Turn it off and now the generation work

Wellington Almeida dos Santos

Before coming the minmax i was about to ask you to release a new version workflow for ltx 2.3 director 2. I saw people using it on minxmax, but using the multiple reference seems pretty close. Do you intend to test out the director 2 on minmax?

UnderToker

Isnt there a LLM node that can be used in a workflow that could insert the prompt automatically before adding the description the user inputs and creating the correct minimax prompt directly in the workflow?

LW

Hi K, I have to say how impressed I am with Minimax H3. You were not kidding when you said "MiniMax H3 is simply THE BEST OPEN WEIGHT AI VIDEO MODEL EVER!" That being said it is, unsurprisingly, resource intensive. Could you release an updated workflow that enables the production of longer videos? I assume that this would be by joining successive 10 second videos together. Thanks in advance.

Kris Miller

Looks good, can't wait to try this out.

AI Artists' Guild

OK but how do you find the subgraph section, what does it look like? Trying to look around on google and AI can't describe what things 'look' like yet...well not for my purposes when searching google...image taggers exist

Aitrepreneur

Yes but I don't recommend it, it add more generation times AND the prompt will never be as good as if you use chatgpt or Grok for example

Aitrepreneur

If you are using turbo don't use spectrum at the same timee

Daniel Ramirez

Im pretty sure I followed everything to a T, ehy would it be failing? MiniMax H3 FL2VA MODEL — T2V + I2V failed This node threw an error during execution. Check its inputs or try a different configuration. # ComfyUI Error Report ## Error Details - **Node ID:** 8 - **Node Type:** UNETLoader - **Exception Type:** RuntimeError - **Exception Message:** RuntimeError: shape '[28672, 5376]' is invalid for input of size 128915322

Yanis Rigouste

Yoo, thanks! Is the one without the Turbo LoRA better in terms of final image quality? :)

Roy Kahalany

Getting this error (Text to Video): Node threw an error during execution. MiniMax H3 QWEN3-VL TEXT ENCODER # ComfyUI Error Report ## Error Details - **Node ID:** 12 - **Node Type:** CLIPLoader - **Exception Type:** RuntimeError - **Exception Message:** RuntimeError: shape '[25600, 5120]' is invalid for input of size 47159490

Firespawn

this is no help. Clicking on the subgraph and making sure its active does not work. This is not the problem