YT YouTube2Text

YouTube2Text → Transcripts → ComfyUI MiniMax H3: Best Video Generation Workflows (Ep29)

ComfyUI MiniMax H3: Best Video Generation Workflows (Ep29) — Transcript

by pixaroma · 5,914 words · 842 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:01[music]
0:02Welcome to another video tutorial. Today
0:05I will show you how to use the new
0:07Miniax H3 video model so you can create
0:11awesome videos.
0:13>> Welcome to my Comfy UI setup where every
0:15cable has a purpose probably.
0:19Perfect. I have now connected everything
0:21exactly the way I was not supposed to.
0:25So I guess we are buying more VRAM.
0:28>> What is your name? I am Pixab Bunny.
0:32>> Where are you coming from?
0:42>> You always look at me like you
0:43understand everything. Are you my little
0:45best friend?
0:46>> Of course I am. But a best friend also
0:48deserves treats.
0:49>> So that is what this is about. You are
0:51adorable and very clever.
0:53>> Clever, cute, and hungry. That is my
0:55full talent set.
0:57>> Milo, be honest. Do you think I am
0:59overthinking everything again?
1:01>> Yes, but in a very talented way.
1:04>> That is not exactly comforting, but
1:06somehow it helps.
1:22As you know, I use the Comfy UI easy
1:24install like in episode 1 because it has
1:27a lot of useful scripts that make things
1:28easier. So, before I run Miniax, I have
1:31to do some updates. I will update easy
1:33install because that will also update my
1:36Pixaroma nodes. If you have another
1:38Comfy UI version, just update Pixar Roma
1:41nodes manually with GitPull or from the
1:43manager if it works. I asked for it to
1:45be added to the manager again since it
1:47was removed for a while because of the
1:49problem I had with GitHub in the past.
1:51Press enter to exit when it shows the
1:53message. Then you need to update Comfy
1:56UI. You should have a similar BAT file
1:59somewhere, either in the update folder
2:01or in the Comfy UI folder that lets you
2:04update Comfy UI. This is necessary
2:07because the miniax nodes are new and you
2:09do not have them in older comfy UI
2:11versions. Then to speed things up, it
2:14helps to have Sage attention. If you do
2:17not have it with easy install, it is
2:19quite easy to install. From the add-ons
2:22folder, run this BAT file. And if all
2:24these things are green, it should work
2:26correctly. If you get messages in red,
2:28then maybe your video card is not
2:30supported. This is for Windows and
2:32Nvidia cards. I had some problems with
2:3420 series video cards, but on the 30
2:37series, 40 series, and 50 series, it
2:40should work correctly. So, after you run
2:43that, you will have this extra comfy UI
2:45startup BAT file with Sage Attention. If
2:48we take a quick look inside, we can see
2:51it has these arguments. So, it knows to
2:53start with Sage Attention. This means we
2:56do not need to add any extra Sage
2:58Attention nodes to our workflow. In
3:00fact, if you have Sage Attention nodes
3:03in your workflow, it is better to start
3:05Comfy UI normally and let the nodes do
3:07the job. But I prefer this method
3:09because it gives me smaller workflows
3:11and it works the same. So now in the
3:14launcher, I can start either the desktop
3:17or browser version of Comfy UI with Sage
3:20Attention. You can see that when it
3:22starts, it says it is using Sage
3:24Attention. I added a version check
3:27Pixaroma node here so you can see which
3:30Comfy UI version and which version of
3:32Pixaroma nodes I have at the moment of
3:35recording. Make sure you have this
3:37version or a newer one, not an older
3:39version. Then you can download the free
3:42workflows in the Comfy UI folder. Look
3:45for the user folder, then default and
3:48finally workflows. You can place the
3:50downloaded folder with the workflows
3:52there. I removed the resources and left
3:55only the actual workflows in that
3:57folder. So now in Comfy UI, you should
4:00see the new workflows. If not, use this
4:03refresh button. You have them organized
4:05in the same order that I present them in
4:07this tutorial. Some of you may have
4:09noticed some new buttons here, like a W
4:12letter and a question mark. This new
4:14button or the alt + W shortcut opens the
4:18workflow manager that I built for you.
4:20This is a long video, so I will keep
4:23this part short. Basically, you can move
4:26and arrange workflows, rename them, and
4:28do pretty much anything you want with
4:30them, so you can stay organized. But the
4:33best way to get started is to click this
4:35help button and read about everything it
4:37can do. It has a large help section, so
4:39it should include the basics you need to
4:41figure it out. You can click W again to
4:43close it. You also have this help file
4:46for all Pixaroma nodes so you can easily
4:48find out how to use them and what they
4:50do. I hope the Comfy UI team creates
4:53something similar for their nodes as
4:55well to make things easier. So, click
4:57around, search, and look at what each
4:59node does. You can also find a lot of
5:02other useful information there. If you
5:04do not want to see those buttons, you
5:06can go to settings, then look for
5:08Pixarroma and scroll down until you find
5:10these settings. From here, you can hide
5:14all the buttons or only the ones you
5:16want. You can use the shortcut for
5:18workflows and the help section can also
5:20be opened from each Pixaroma node. For a
5:23line, you need to have it enabled for it
5:25to work. So, let me move on to the
5:27MiniAX H3 model. First, I will open the
5:31folder with the workflows for this
5:32model. By the way, you can add workflows
5:35to favorites, and you can also
5:37right-click them to see more options,
5:39including one that lets you add a
5:41preview cover for that workflow. If you
5:43select a workflow, like this texttovideo
5:46workflow, it will show you a preview and
5:48useful information on the right. Scroll
5:51down and you can see which model it uses
5:54and a lot of other information. You can
5:56also add a note to the workflow so you
5:58can search for it by that note later.
6:00Double click a workflow to open it or
6:03use the open button. I tried to make the
6:05workflows simple and easy to use and to
6:08reduce the number of nodes where
6:09possible. All the things you need for
6:11these workflows are in this note. The
6:13most important one is the Miniax H3
6:16model which is quite big and you can
6:18click this button to see more versions.
6:20So here is a list of them. There are
6:22actually two models with different
6:24quantizations. One is first last frame
6:27to video and one is reference to video.
6:30No matter what video card you have, even
6:32if it is powerful, I still recommend
6:35that everyone uses this model, the
6:37pruned int 8 version. The rest are
6:39slower with not much improvement in
6:42quality. For the reference model, use
6:44this version as well, but I already have
6:46the correct model selected in the
6:47downloaded workflow, so you do not have
6:49to worry about it. Then there is one
6:51thing that many people do not seem to
6:53talk about, the license. If we read the
6:56license, we can see that some
6:58territories are excluded from it. The
7:00European Union, United Kingdom, Korea
7:03and United States of America are not
7:06covered by this license. I assume this
7:08is happening because of legislation in
7:10those countries and instead of having
7:12problems, they added this condition to
7:14the license. But does that mean no one
7:16from those countries can use it? I added
7:19another link here where you can read
7:21more about the license and they explain
7:23it in more detail. They explained that
7:25they had to choose between not releasing
7:28the model until it complied with the
7:29legislation in those countries or
7:32releasing it while excluding those
7:33countries or something similar. I am
7:36from Europe, so how can I use it if
7:39Europe is excluded? There is a link here
7:41to an application form. It asks for an
7:44organization. I am not sure if this also
7:47applies to people without an
7:48organization, but you probably do not
7:50need this license anyway if you only use
7:52it for fun and personal use. So, I
7:55clicked this application form and
7:56completed my details. I have the
7:58Pixaroma company, which is a oneperson
8:01company where I do everything myself.
8:03Then, it asks about annual revenue and
8:05obviously I do not make that amount of
8:07money. After that, you have to check a
8:10lot of terms confirming that you agree
8:12to comply with the license. I am not a
8:14lawyer but I think this means they give
8:17the responsibility to you so they cannot
8:19be held accountable for any problems
8:21caused by how you use the model. This
8:24keeps them safe and I am happy that I
8:26can use the model so everyone is happy.
8:29That is how I did it so I could use the
8:31models on my channel. I received an
8:33email confirming that I was approved and
8:35that was all. So let me show you how to
8:37get the models. Use this button to
8:40download the model. It will take a while
8:42because of its size. After downloading
8:44it, navigate to your Comfy UI folder. Go
8:47to models, then diffusion models, and
8:50create an H3 folder to keep things
8:52organized. Place the model for this
8:54workflow inside that folder. I have two
8:57models because I also downloaded the
8:59reference model for the other workflows.
9:01Now, you have the Miniax model. Next, we
9:04need a text encoder. Download this Quen
9:07model which is also big but very smart
9:09with 32 billion parameters. Until now we
9:13used models with four and 8 billion
9:15parameters. Place it in the text
9:17encoders folder. For the VAE we have two
9:20models because this model also supports
9:23audio. Download both of them into the
9:25VAE folder and then you can load them
9:27into these nodes. As for the custom
9:30nodes, you know me, I try to keep things
9:32simple. Only Pixaroma nodes are used. I
9:35also included a custom chat GPT link I
9:38made for prompting, but I will talk more
9:40about that later. After all the models
9:42have finished downloading, you can press
9:44R to refresh the node definitions, and
9:47everything should work if the files have
9:49the same names and are placed in the
9:51correct folders. If not, you will need
9:54to select them manually inside the
9:55nodes. Like I said, I like simple
9:58workflows, so compared with what you
10:00find on the internet, this should be
10:02much easier to use. I added some steps
10:05in blue for you to follow. And this
10:07workflow only needs three steps from
10:08you. Let me start with the first one
10:10where you can choose the orientation and
10:12size. I added some of my favorite sizes
10:15here, which I recommend depending on
10:17your video card. The more VRAM you have,
10:20the larger the size you can generate.
10:22People have tested this even on a video
10:24card with 8 GB of VRAM, and it worked if
10:26they had enough system RAM and dynamic
10:28VRAM was enabled. Remember to switch
10:31between portrait and landscape since
10:33that decides the ratio. Because we are
10:35making videos, I did not include a
10:37square ratio, but you can add one as
10:40well. If you go to settings, you can add
10:42the sizes you want here, change their
10:44order, remove sizes, or add more. I
10:47included all the sizes the model can
10:49generate, and I added a star next to the
10:52ones that I think work well. For the
10:54optimal size, I think this one is the
10:56best, but I did many of my tests using
10:59this size. Miniax also prefers sizes
11:02that are multiples of 32, so I enabled
11:05that option. No matter what size you
11:07enter, it will round it to a compatible
11:10value. So, that is the Pixar sizes node.
11:13I only changed the title to make it
11:15easier to use. For this video recording,
11:17I will select portrait and this size.
11:20Then, I created another node for the
11:22duration. If we go to settings, you can
11:25add your custom durations here in
11:26seconds. I also added a few ready-made
11:29formulas directly into the node. By
11:32default, it uses this Miniax H3 formula,
11:36but you can use others for different
11:37workflows or create your own formula.
11:40The cool part is that you can turn it
11:41into a slider if you want and adjust the
11:44values or you can just type them if you
11:46prefer that. I like this button version
11:48because I can add only the values I
11:50need. So, I will use a 3-second video
11:53for this test. Then comes the long
11:56detailed prompt. The more details you
11:58add, the better it can understand what
12:00you are asking for. You can write all of
12:02that by hand, but I cannot write prompts
12:04that are so detailed and structured in a
12:07way that AI understands perfectly. So,
12:09why not use AI to create prompts for AI,
12:12right? Even if a short prompt might work
12:15well, you give the model too much
12:17freedom to invent things. That is why I
12:19built this custom chat GPT. You can
12:22click on it and use it. It should also
12:24work with a free subscription, but you
12:26can use another LLM if you want. I will
12:29also provide some prompt formulas inside
12:31the workflow archive. Since I created
12:33this GPT, I can also edit it so you can
12:36see how I set it up. In the configure
12:38tab, I added instructions telling it to
12:41use the attached knowledge. Here I
12:44uploaded this text file with the long
12:46formula that explains how to write the
12:48prompts. The formulas are also inside
12:50this folder that you downloaded with the
12:52workflows. There is one for the
12:54reference model and one for text to
12:56video and first last frame to video. You
12:59can look at them, change them, and adapt
13:01them however you want. So back to the
13:04custom GPT, you need to tell it what
13:06type of prompt you want. Is it for text
13:08to video? Is it for first frame to
13:10video? or is it for first frame last
13:12frame to video. This helps it choose the
13:15correct prompt structure. In this case,
13:17I want a textto video prompt. Then I
13:20describe in my own words what I want and
13:22I hope it understands me. And I got a
13:24long detailed prompt for that. Some
13:27parts look like instructions, but that
13:29is fine because it helps the model
13:31understand what I want it to do. You can
13:33see how it describes everything and also
13:35how it uses English. I tried it with
13:38other languages and it works. It says it
13:40supports 11 languages, but I also tried
13:43Romanian and it worked. Sometimes it
13:46mixed up one word, but overall it worked
13:48well. I added a run timer here so I can
13:51see how long it takes, but more recently
13:53I prefer using a run log because I can
13:56see the history for each run. Let me run
13:58it and see how it works. It loads all
14:01the models and then the prompt and
14:03settings go into this new miniax node
14:05created by the Comfy UI team. After
14:08that, it goes into the K sampler where I
14:10used 20 steps and set CFG to one because
14:13we do not have a negative prompt. For
14:15the sampler, I used res multi-step. And
14:18for theuler, I used simple. Some people
14:21also recommend beta. So, you can try
14:23that one as well. Then, I ran into a
14:26problem and got an error. What could it
14:28be, I wonder? I recorded this because
14:30some of you might get the same error,
14:32and I will show you what fixed it for
14:34me. What I did was close Comfy UI. Then
14:39inside the BAT file I used to start
14:41Comfy UI, I had this argument that
14:44disabled dynamic VRAMm. You could remove
14:47it manually and save the file, but to
14:49make sure all the BAT files are set
14:50correctly and have that argument
14:52removed, I will do something else. If
14:54you use a GGUF model, you might want
14:56dynamic VRAM disabled. But in this case,
14:59I want it enabled. So, I go to add-ons,
15:02then tools, and toggle it easily using
15:05this BAT file. Now, it removed that
15:07argument. If I go back and check, you
15:10can see that the argument is no longer
15:12there. So, everything should work now.
15:14But let me test it. I open comfy UI
15:17again and run the workflow. And now I
15:19got the video I described.
15:23>> Do you like my cat?
15:28>> Do you like my cat?
15:30I will run it again and then I can add a
15:33label for each recorded runtime. By the
15:35way, you can now rightclick the node, go
15:38to node settings and enable this option.
15:41Then you can close the settings and look
15:44now it shows my video card along with
15:46the VRAM and RAM I have. Pretty cool,
15:49right? And the second video looks like
15:52this.
15:53>> Do you like my cat?
15:58Do you like my cat?
16:00>> I also did some tests while I was not
16:02recording so the recording would not
16:04affect the generation times and here is
16:07what I got. All the tests are for a
16:105-second texttovideo workflow. I did 10
16:13runs and for each run I increased the
16:16resolution to see how it affected the
16:18total running time. The higher the
16:20resolution, the more time it takes. But
16:22the increase is not linear. For example,
16:24for this resolution, it took around 3
16:26minutes, but for full HD, it took 10
16:29minutes. So, I prefer to create a
16:3115-second video at a lower resolution
16:33instead.
16:37[music]
16:48[music]
16:50Heat. Heat.
16:52[music]
16:57[music]
17:03[music]
17:09[music]
17:14>> [music]
17:25[music]
17:26>> So you saw that the higher the
17:28resolution the better it looks but it
17:30also takes more time to generate. Let me
17:33move now to the first frame to video
17:35workflow since we also want to control
17:37the first frame so we can animate
17:38exactly the image we want. This workflow
17:41uses the same models and nodes. There is
17:44one extra thing here, an image, and that
17:47image is connected to the first frame
17:49input in the miniax node. After the
17:51image is loaded, step two is to choose
17:54the duration. The longer the video, the
17:56more time it takes, but at least the
17:59increase is more linear. So, if you know
18:01the generation time for 3 seconds, it
18:03will take around twice as long for 6
18:05seconds, three times as long for 9
18:07seconds, and so on. To make things
18:10easier, instead of constantly changing
18:12the video size, checking the image
18:14ratio, and adjusting everything
18:16manually, we can do all of that with a
18:18single click. So, I built another node
18:21called longest side. It takes the image
18:23from the load image node, makes the
18:25dimensions multiples of 32, and sets the
18:28longest side of the video to the value
18:30you choose here while keeping the
18:32original ratio. Then, step four is the
18:34prompt. For longest side, we have some
18:37settings here. You can see that for this
18:39workflow I used 32. You can change that
18:43from here by clicking the button
18:44multiple times. You can choose the
18:47longest side of the video from here. I
18:49added the recommended values, but you
18:51can add more. You can even use a
18:54different ratio if you want to crop the
18:56image. For example, if I want a square
18:59video, I can select that. Then from the
19:02settings, I can choose whether the crop
19:04should be centered or positioned
19:06differently. In this case, because the
19:08woman is near the top, a centered crop
19:11would remove her face. So, a top crop
19:13would work better. But I do not plan to
19:15crop the image. So, most of the time, I
19:17prefer to keep a ratio that is as close
19:20as possible to the original image. For
19:22the size tabs, you can also add your own
19:24values here. So, I can change them
19:27however I want. And then the buttons are
19:29updated here. The next time I click that
19:31button, it will use the new size. I will
19:34put it back to the default settings for
19:36this workflow. Then for the prompt, I
19:39usually take a screenshot of the load
19:41image node. After that, I go to chat GPT
19:44and open the custom chat GPT. I ask for
19:48a first frame to video prompt, paste the
19:50screenshot, and then describe what I
19:52want to happen. Chat GPT is quite good
19:55at this. Now, I have a long prompt that
19:57I can copy and paste here. I will use 5
20:00seconds and then run the workflow. I
20:03just noticed that I forgot to remove the
20:04square crop. So, I will cancel the
20:06generation, set it to keep the original
20:09ratio because I do not want to crop and
20:11run it again. It took around a minute
20:13and a half. Let me check the result.
20:18Well, hello there little sweetheart.
20:23Well, hello there little sweetheart.
20:25>> If I check the prompt, you can see that
20:27it used the English tag there along with
20:29the words in English. But like I said,
20:32you can use other languages, too. Here
20:35is an example in Romanian.
20:49>> [laughter]
20:52>> I did tests with 5, 10, and 15-second
20:55videos, and the time is around double
20:58for each extra 5 seconds, but the
21:00maximum for miniax is 15 seconds.
21:02Anyway, I tried to increase the
21:04resolution to this one. And now for 15
21:07seconds, I got 13 minutes. So, it is
21:09almost like 1 minute of running time per
21:11second of video at this resolution. Here
21:14are a few more tests for image to video.
21:17And it is almost the same as text to
21:19video just with a few extra seconds,
21:21maybe around five extra seconds at HD
21:23resolution. I now have some volunteers
21:26on Discord who test my workflows called
21:29bug hunters. And here is an example of
21:31tests on 16 GB VR RAM. They tried it
21:35with Sage Attention and also with some
21:37extra nodes. With easy cache, it reduced
21:40the time, but the results were worse. So
21:43I prefer the quality. The generation
21:45times with Sage attention as arguments
21:48and with nodes are similar. So I prefer
21:51using the arguments until someone
21:52releases better nodes. It is time to
21:55move to the next workflows that use
21:57besides the first frame also a last
21:59frame so you have control over where it
22:02ends as well. For this example, I tried
22:05two different bunny images to see if it
22:07can do a transition. So I started with
22:09this bunny sitting as the first frame.
22:11Then this ballerina bunny as the second
22:13frame. I set it to four seconds. I let
22:16it keep the ratio and this longest side
22:19and that is taken from the first frame
22:21since that usually decides the ratio of
22:24the image. So that first frame is
22:26connected here and then it goes into the
22:28miniax node into the first frame input.
22:31The second image is connected to the
22:33last frame input. I will run this
22:35workflow and just to give you an idea of
22:38how I prompted it, I took a screenshot
22:40of both nodes and in chat GPT I asked
22:43for a prompt for first frame last frame
22:45to video. Then you add how you want the
22:47transformation to happen. I wanted a
22:505-second video and it gave me a long
22:52prompt and the result for this workflow
22:55looks like this.
22:57[music]
23:02[music]
23:04For the next workflow, you can use it to
23:06get the last frame of a video. So, you
23:08decide how the video should end. So, for
23:10example, I have this ballerina bunny. I
23:13set the duration to 4 seconds and I kept
23:16the ratio and generated the video at
23:18this longest side. With a long prompt,
23:20the image should be connected only to
23:22the last frame input. And you leave the
23:24first frame input empty so the model can
23:26invent that part. So, if I run this
23:29workflow, I get this result.
23:33>> [music]
23:39>> So it is very easy to make the video end
23:42on a certain image and you do not have
23:44to generate two images all the time. So
23:46all these four workflows use the first
23:48model the first last frame to video
23:50model. Now I am moving to one that uses
23:53reference instead. I have one workflow
23:55here for two images and one for three
23:58images. Let me start with the two image
24:00one first. Make sure you downloaded this
24:02reference video model because it will
24:04not work correctly with the first one.
24:06Then for chat GPT, I made a different
24:09one because reference works a little
24:11differently. It does not start with a
24:13certain frame or end with it. It looks
24:15at the references and generates
24:16something new from them. So for the
24:19prompt, I will just take a screenshot of
24:21these two images and then in chat GPT I
24:24paste it and ask for a reference to
24:26video prompt for these two images. Then
24:28you describe how you want it. I already
24:30did that for this long prompt and the
24:32result is this.
24:36>> Are you ready to escape this summer?
24:41>> Are you ready to escape this summer?
24:43>> That is just a simple example, but you
24:45can do more complex things. Let me show
24:48you another example with three images.
24:50It can do more than three. You just
24:52connect more images there in the miniax
24:54node. It can do a maximum of nine
24:57images, three short clips of around 5 to
24:5915 seconds, but for me that took forever
25:02to generate. I am not sure if that is a
25:04bug, but I could not get it working
25:06properly. It can also use three audio
25:09files with a maximum of 12 files
25:11combined. I will show audio examples in
25:14a few minutes. Since this uses images as
25:17references, the ratio of the images does
25:20not matter that much. So, I can use
25:22custom ratios here. So decide on the
25:25maximum resolution you want to use and
25:27do not forget to choose portrait or
25:29landscape. I tell you this because I
25:31forgot many times to switch it. I could
25:34make it follow the ratio of the first
25:36image like in the other workflows, but
25:38then I have to keep in mind that the
25:40first image must have the correct ratio.
25:42This way I know I can use images with
25:45any ratio and I choose the video ratio
25:47myself. So, I want this bunny wearing
25:50that orange outfit in an exotic location
25:53like the last image. I used chat GPT for
25:56the prompt and the result is this one.
25:58>> Hello from paradise.
26:03Hello from paradise.
26:08>> Add more load image nodes there. Connect
26:10them and save your own workflows based
26:12on these. Now let me move to the next
26:15group which can also use an audio
26:17reference combined with an image. It
26:19does not do audio to audio. So I have
26:22one workflow here for speaking and one
26:24for singing. Let me start with the
26:26singing workflow first. Like I said it
26:29uses the same reference video model. You
26:31set the duration of the video and I made
26:34it so that it also affects the uploaded
26:36audio which helps you get better
26:37results. From this load audio pixeloma
26:40node, you can upload an audio file or
26:43select one from the input folder. You
26:45can see that when I select a different
26:47duration here, it also adjusts this
26:50orange selection. To make that work, the
26:52seconds output is connected here. So,
26:55the selected audio section always
26:56matches the chosen duration.
26:59[music]
27:04second selection to the right to start
27:06at a different point.
27:10>> So everything is controlled. If I
27:12disconnect the seconds input now I have
27:14more control and can adjust it manually
27:16however I want. I am showing you this
27:19because it can be useful for other
27:20workflows. But for this one it is easier
27:23to leave the seconds input connected.
27:26Also check the settings and help
27:28sections for more information. If you
27:30set a duration longer than the actual
27:32audio, you can either add silence or
27:34loop the audio. So, I will reconnect the
27:37seconds input. Let me pick 5 seconds and
27:40play it to see which part will be used
27:41for singing by looking at that white
27:43line.
27:45>> [music]
27:54[music]
27:58>> Maybe I will make it shorter and use 3
28:00seconds.
28:03[music]
28:04>> Okay, that part seems good for the
28:06input. I have a universal prompt here
28:08that works with almost anything. You can
28:10adjust it if needed, but it should work
28:12most of the time as it is. Now I will
28:15run it. I forgot to mention that you set
28:17the size of the final video from here
28:19using the longest side. This workflow
28:22also has this new H3 audio sync node
28:24that I made. If you click more, you can
28:27see that it helps keep the audio
28:29consistent. It is too complex for me to
28:31fully understand what it does, but
28:33Claude figured it out and I got good
28:35results. And the final result is this
28:38one. [music]
28:40[singing]
28:45You can adjust the prompt if you want
28:47some extra movement there. For example,
28:50I can add that the bunny puts his hands
28:51in the air while dancing or something
28:53similar. Then I run it again and I got
28:56this. [music]
29:01[singing]
29:03>> For some seeds, I got barely visible
29:05lines in the generated video. I am not
29:08sure if it is because I used a low
29:10resolution, if the model sometimes does
29:13that, or if there is a wrong setting
29:15somewhere. It is just something that
29:17appears sometimes, usually around the
29:19middle of the screen. Moving now to the
29:21second workflow, which is similar to the
29:23first one, but has a different prompt.
29:26So, I add the voice here.
29:29I can talk now since we have the MiniAX
29:31H3 model. Maybe I will make it shorter
29:34and use only 3 seconds. I can talk now
29:37since we have the Miniax H3 model. I set
29:40this size for the longest side so it
29:42generates faster. And now I will run it.
29:45You can see how I set up the prompt, but
29:47feel free to adjust it. Maybe you can
29:49find a better one, but this worked well
29:51for me. You also have more settings in
29:54the audio sync node. I made it so the
29:56audio does not go beyond 15 seconds
29:59since Miniax can only generate a maximum
30:0115-second video. Anyway, if the audio
30:04clip is longer, I set it to warn me, but
30:07you can also make it stop the run if you
30:09want. You also have the same options to
30:11add silence or loop the audio if the
30:13clip is too short. And the result is
30:16this one. I can talk now since we have
30:18the Miniax H3 model. I can talk now
30:20since we have the Miniax H3 model.
30:24Even if it is a video model, MiniAX can
30:26generate images or edit images, but it
30:29is a little slow. So I still prefer the
30:31Crea 2 model for that. But if you
30:34already have the model, at least you
30:36know it is an option. So for this
30:38portrait landscape node, I added a
30:41setting so it is constrained to
30:42multiples of 32. And you put your sizes
30:45there and just flip it to get the right
30:46ratio. You add a long prompt and you can
30:49run it. So it took 45 seconds for a full
30:52HD image and the result is not bad if
30:54you do full HD.
30:57If you reduce the size, you get faster
30:59generation, like only 17 seconds. But
31:02the result is not great, more like a
31:05frame from a lowquality video, which in
31:07fact it is. It generates like a batch of
31:10five frames and picks the first one,
31:12index zero, from that. I could not set
31:15the length to one because the node gave
31:17an error. I think Miniax likes five
31:19frames at once. And now let me try the
31:22edit workflow. It is similar to that
31:24one, but this time we load an image. And
31:27I added here what to change, but longer
31:30prompts that give details about what
31:32should change and what should stay the
31:33same could work better. I tried with a
31:36smaller size this time for edit using
31:38this one. It has the same get image from
31:40batch like before. And the loaded image
31:43is connected to the first frame. And
31:46this time it took around a minute. So it
31:48changed the hair like I asked, but it is
31:50losing a bit of quality. You can get
31:52away with small details, but sometimes
31:54it can change more things. For example,
31:57sometimes it changed the text to pink
31:59also. Not only the hair, but it depends
32:02on the seed. In this case, it worked
32:04even with a short prompt. So, do not
32:07forget to check the help for each
32:08Pixaroma node. So, you can learn more
32:11about the node. That help is updated
32:13when I update the node and change
32:15something in it. Search for the things
32:17you need. And I hope all these nodes are
32:19useful for you. I will leave you with a
32:21few more examples of videos generated
32:23with Miniax.
32:25>> Meet the Pixa time machine because
32:27waiting for the weekend is basically
32:29unbearable. Destination selected.
32:32Medieval era.
32:33>> Okay, definitely older than my
32:36apartment.
32:37Picks a time machine. Travel smarter.
32:40Everyone has a smartphone. I wanted
32:42something fresher.
32:44Hey, Apple phone. What's on my schedule?
32:46>> First a call, then a snack. Hopefully
32:49not in that order.
32:51>> Finally, a phone with appeal.
32:53>> I was going to [music] say that.
32:55>> Hey guys, welcome with me into this
32:56amazing jungle. It's so beautiful here.
32:59Look at this flower. The color is
33:00incredible. It's even more beautiful in
33:02real life. Oh my god, look, a little
33:04monkey. He's so cute. This is definitely
33:07the best part of the trip. Say bye,
33:09little guy. Welcome back to the vlog.
33:11Today we are exploring the jungle where
33:13everything is beautiful, wet, and
33:14probably trying to eat something. Look
33:16at these flowers. They are gorgeous,
33:17peaceful, and suspiciously perfect.
33:19Okay, that flower has teeth and it just
33:21caught a dragonfly. I officially take
33:22back the word peaceful. Let's leave
33:24before it discovers food delivery.
33:28[music]
33:33[music]
33:40>> I knew it. The cheese thief always
33:42returns to the scene of the crime.
33:44>> Thief is such a harsh word. I prefer
33:45Tiny Cheese Consultant.
33:47>> And what exactly does a cheese
33:48consultant do?
33:49>> Quality control. Very brave work.
33:51Extremely delicious work.
33:52>> Fine, but I am supervising the
33:54investigation.
33:57[music]
34:04[music]
34:08[music]
34:12>> [snorts]
34:18[music]
34:27[music]
34:31>> Frontline tested. Planet approved. This
34:34is the Vorax [music]
34:359. Dense alloy frame, stabilized plasma
34:38chamber, beautiful recoil. management in
34:41the right hands. It ends arguments
34:43before they begin. And yes, it comes in
34:45matte black.
35:01[music]
35:11>> [music]
35:16>> That is all for today. Leave a like and
35:18a comment if you found this useful.
35:20Thank you legends and everyone who
35:22supports this channel. You are amazing.
35:25Have a great day.

About this transcript

This page contains the full transcript of ComfyUI MiniMax H3: Best Video Generation Workflows (Ep29) by pixaroma, generated from the public captions YouTube serves with the video. The transcript has 5,914 words across 842 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.