Yet Another Workflow : easy t2v + i2v
I've aimed at a user-friendly UI for ComfyUI. There's a balance between complexity and ease of use, and this workflow aims to give you useful controls with clear guidance on what you need to care about. I hope these will be helpful to anyone strugging with quality and the general UI-isms of ComfyUI. I've taken the time to color code and add lots of notes. Please read the notes, I've tried to make them useful!
This is the workflow I use, it's not aimed at a skill level. It's designed to be easy to use and adjust with some UI concessions and labeling to ensure you can pilot it with less experience in a way that is more sophisticated than the official example workflows, which can be easy to break.
The primary goal with this workflow is to give you a strong foundational place to generate either text to video (T2V) or image to video (I2V) outputs without having to fuss too much. Additionally, reference to video (R2V) support will be added in a future release.
The green controls are the stuff you generally want to mess with.
The secondary goal here is to provide a consistent interface to interact with different samplers and now different models.
v0.50b fixes an issue with LoRA's. If you downloaded the original version and are running into issues, please grab this new one.
Now for MiniMax H3
One of the strengths of respecting a consistency to a UI is that it allows you to change things under the hood to enable people to explore new techniques without changing the majority of their experience. Whether you're familiar with my Wan 2.2 workflow, my LTX-2.3 workflow, or whether this is your first experience with my workflow, once you're comfortable with one of them, the goal is that you can switch between them with minimal fuss.
If I'm calling my shot: MiniMax H3 is the new sherrif in town. Things are always in flux (pun intended), but until Wan announces a new open weights release, and LTX-2.3 solves its technology problems, H3 is here to rule the roost. There are some questions about how easy it will be to train for, but even baseline it's incredibly versitile.
I've made a lot of videos. A lot a lot. I am left speechless by H3's quality and capabilities with no LoRA's or additional support. It just works. Compared to LTX-2.3:
Higher quality baseline
Higher maximum quality: 2K
More consistent identities
Better prompt adherance
Better physics
The T2V model is genuinely good
Basically everything. H3 was made for real video production. You can prompt timeline sequences and edits, and it will follow. H3 offers a genuine challenge to Wan 2.2. Where as previously, you were picking one ore the other. There are still advantages in terms of LoRA's and specific featue support that leave some niches for the other models, but for video generation, this will be your new home.
And the 2K looks imacculate!
Any problems?
Dialog handling seems to be a bit clunkier compared to LTX-2.3, but that might be a prompting skill issue on my end. Early days! I've found the voice performances to be a bit worse on average, but fewer poor generations overall. Accents seem harder to tigger. Characters in H3 are prone to speaking giberish if you don't provide enough explicit instruction. LTX-2.3 has video extension capabilities. There is no formal extension mechanism for H3.
What's the catch?
The model is big. The model is power hungry. The models are enormous compared to its predecessors. A PRO 6000 is probably the baseline in terms of perfomrance and memory. And generation times are longer than equivalent LTX-2.3 videos - roughly 1.5-2x longer - closer to Wan. That delectable 2K resolution will take over 30 minutes for 12 seconds of 2K video on a PRO 6000. I've noted the gen times on each example video in the comments.
Fortunately you can do lower resolutions and shorter lengths.
Like it?
Give it a like! Tag it as a Resource when you use it! Support on Patreon or a tip on Ko-fi are also welcome. Yellow Buzz will go towards promoting awareness here on Civit.
Need help?
I like helping people get going with this stuff, so if you want help message me. If you want extended one-on-one help, there's an option on the Patreon. I'm happy to walk you through the details, answer your questions, and give you some extra tips and tricks, and scripts. I've done this for a few folks, I'll save you money and headaches.
I've also written an article here on getting it going with my Runpod template. The template will vastly expedite and simplify getting things up and running.
General Advice
Make lots of videos! Post your videos! Don't fuss with the tech! Be smart about how you spend your time with this stuff. It's easy to burn out if you spend more time trying to get things to work than making videos you like. That's really why I'm posting this.
Use RunPod. Use the RTX 5090 or PRO 6000 or the H100 SXM. Use my LTX-2.3 template. If you've not used RunPod before, sign up with my link; we'll both get some free credit. See the article for more.
If you use a service like RunPod, if you're doing I2V, it can be smart to have your images ready in advance to make sure the server stays busy while you are using it.
If you run this outside of Runpod, you'll need to install some custom nodes. To do that, click the "Manager" button at the top of the Comfy interface, and then click the "Install Missing Custom Nodes". Click "Install" on each one - I recommend in order; you'll need to wait till each has installed. Do not bother restarting ComfyUI until they are all installed. The RunPod template has them preinstalled.
If the wires bother you, there's a button in the bottom right on the floating UI that will hide them.
This workflow is setup for .safetensors models, but you can use GGUF if you want to make the changes node changes.
Costs?
Still benchmarking. Will add some data soon. Expect similar to Wan.
Troubleshooting
If a node is missing (bright thick red outline with a warning when you open the workflow), you can install them by going to Manager > Install Missing Custom Nodes, and pressing Install on any the nodes that show up there.
If you are getting any errors related to a custom node, it's possible something has changed recently in the software. It might be useful to change a version back to the last "stable" build in these situations.
For example, the nightly build of WanVideoWrapper might introduce an error that wasn't there last time. With a workflow open, you can go to Manager > Custom Nodes in Workflow. This will show you all of the custom nodes. If you click, Switch Ver, you can see all of the releases. Consider trying the first numbered on at the top of the list.
If that doesn't work, or there seem to be more significant problems and you are using RunPod, you may have forgotten to select CUDA 12.8. Try restarting the server. If that doesn't work, terminate the pod, and make a new one. This will fix a surprising number of possible issues.
Sound
H3 has decent sound support, but it's a double edged sword. Sometimes the motion will be great and the sound gets weird. Sometimes the sound is great, but the character does something strange in correct. There's no good way to fix this. I want to call out that there's not a great way way to get a consistent voice across prompts.
Description
Initial release of the v0.50 model for H3
Updated to fix LoRA routing
FAQ
Comments (8)
So far the only workflow that didn't OOM on me!
Works perfectly! Any chance ref2v can be supported?
Yep. I'm working on it. It's one node change, but it's a little more complicated to explain and make fool proof in the way I want workflows to "feel".
Essentially hook up your references and then reference them like: "The man <Picture 1> is leaning against the rail on the pier <Picture 2>."
But you need to manually hook up each image and it's a bit weirder than just muting the last frame, since the number of references is variable. So it's a bit tricky from a "make this easy to understand and use perspective", but on the technical side it's a trivial change.
@boobkake22 Looks forward to it! Tried a lot of workflows. Others may be faster, but yours generates the best videos.
@skyofspyof894f Yeah, people are trying to pitch tricks to make it work faster, but that doesn't make MiniMax look great. Wan 2.2 is completely unusable without acceleration, where as MiniMax is just slow. It's still reasonable to want faster gens, but it also means trade offs in quality and versatility.
And to be clear, my workflow isn't doing anything fancy, I've just set it up to be smart and reliable. (That's what all of my workflows are about.)
Anywho! Refs are coming!
Is it possible to add upscaling in the next version and additional Sage settings? I found the Patch Sage node hidden under the Lora Loader and it was set to disabled. I end up using KJ loader that allow Sage setting, added both the KJ experimental H3 Sage node and Sol Attn node. Got 1M 4:3 20 second video completed in 2100 seconds on a RTX 3090 capped at 220W. Drop down to 0.4M and it's done around 630 seconds.
Yeah, there's will be some changes in the next version. Waiting for a few things to settle. I don't like spamming updates on this stuff, and things are changing pretty fast. Expect one sooner than later.
It's pretty clear that Sol Attention has a negative effect on the output, so that's an unlikely include. Sage is disabled because without adjustments it causes CUDA errors that crash Comfy at higher resolutions. (This is a temporary measure while these things settle untill the "right way to do it" becomes more clear.)
Additionally, there are now a mess of Turbo LoRA's, and again, waiting to see where the concensus settles, but one will make it's way into the workflow officially eventually.
I've actually gone ahead and made a beta version with both the "Larry" 600 Turbo LoRA and the H3 SageAttention patch. It's available on the Runpod template and on the Discord server. (Link in the workflow.)