I really like MiniMax H3. If you haven’t tried it, then stop what you are doing and try it today. I was in heaven earlier this year when I found that ComfyUI can make Wan 2.2 and LTX 2.3 on my PC (which admittedly is a good PC). But now MiniMax leaves them in the dust. MiniMax has different licensing requirements, so read my guide to the MiniMax H3 license for US and EU users.
The thing that MiniMax does for me, that the other engines do not do, is that it allows me to name random objects and include them in my videos. Now my private videos are full of some curious objects – like magic flying carpets, ground squirrels, and blacksmithing tools.
Why i2i Edit is needed for Video Creation
All i2v videos, as the name suggests, start with an image. Normally, I use common image-generation tools, and right now I love OpenAI’s SunFlare and Sunburst tools. These are great for producing high-quality start images. The main drawback of private generation is that OpenAI, Grok, MAI and other mainstream engines will censor images. In my tests, MAI images (Microsoft AI) do not generally allow swimsuit images for men, but do for women, which is both censored and sexist.
So, in my video creation workflow, I often take a mainstream picture and run it through an i2i edit to reconfigure it as my starting image for a video. For this, I have commonly used Qwen 2511 image edit, which comes with ComfyUI templates.
My Experience with Qwen 2511
One problem with Qwen image edit is that details on natural backgrounds look somewhat cartoonish. Using Qwen 2511 for a video yields cartoon trees, mountains and bushes. Another common problem is that the elements in the starting image aren’t “Native” to the target video.
- Native: the video model drew the object itself, so it knows what the object is and how it moves.
- In-distribution: the technical term for this. The object matches what the model saw in training.
- Out-of-distribution: the opposite. The object came from another model or a photo, so the video model has to guess.
- Domain gap: the difference between how two models draw the same thing. Qwen’s trees and H3’s trees don’t match.
- Morphing: what happens when the guess is wrong. An object changes shape or turns into something else as the video plays.
- Object permanence: the goal. Objects keep their shape and identity from first frame to last.
So a common problem for Qwen to LTX is that an object in the room gets misinterpreted and becomes something else. In one video, I made a paper tissue on a counter turn into a kitten that then walked across the counter.
Using MiniMax H3 as its own i2i Image Edit Tool
One thing I have perfected is using MiniMax as its own i2i image editor. That is – use it on my own PC to render uncensored start images for my private videos. While this is not a normal mode for MiniMax, it works fine for me. I also see they offer their own i2i workflows. Personally, I have great results using my normal Ref2VA workflow. But I add two major modifications:
- I render the largest possible 8.29MP (3840×2160) video
- I render for 0.1 seconds (One tenth of a second – which makes 5 frames)
The first is to get the highest resolution. Video, by nature, creates soft edges and tends to ghost. By rendering large and later shrinking, I can compress the soft edges so they sharpen.
And 0.1 seconds ensures a total pixel count that will render fine on my PC. Typically, this render takes about 5 minutes on my PC, where the Qwen process took about 2 minutes.
While researching this article, I learned a new trick: use Frame 0, which is noticeably brighter than Frame 3, which I had previously been using.
Here is my test
Start photos – A man hiking and a Victorian lady.
Qwen 2511 Render
The test is to place them together, change clothes, preserve one element of the backdrop but reframe it, and add a completely random element – in this case, a Circus elephant.
<image1> is a picture of a man hiking at Mt. Bachelor, Oregon with a view of Broken Top and the Oregon Cascades.
<image2> is a picture of a woman in victorian dress with black gloves on.
Instant change: The man is dressed in a dark suit, white shirt, no tie.
Instant change: The woman <image2> is dressed in a modern sleek ivory colored evening dress.
Instant change: The two people are standing at Mt. Bachelor Lodge in Oregon, with a huge view of Broken Top and the Cascade range. They are dressed and ready for an elegant dinner. No alcoholic drinks in view.
Instant change: Also in the room, without explanation is a full sized circus elephant.
So Qwen 2511 does a credible job of following my instructions. But when you look at the trees, you can see that they look more like World of Warcraft than the original trees. The mountain is an entirely different mountain (Mt. Bachelor, not Broken Top). A video made from this picture will start with plasticky faces and probably won’t improve them.
How that compares to Qwen 2.1
Updating to the brand new Qwen 2.1 model is a generational improvement.
The faces are vastly improved; the background is the correct mountain, and the trees and the elephant both look much better. The glaring error is the faces. The man’s face is from the start photo, brightly lit with shadows, and the woman’s face, from her start photo, has flat light and softer features. They do not match.
Qwen works by using the exact source image as template, without reinterpreting it.
Rendering using MinMax H3 0.1 Second Video – Frame 0
overall_soundscape: Ambient and physical sounds
non_diegetic_music: N/A
integrated_multimodal_description:
<Picture 1> is the start image.
<Picture 2> is the character reference.
Instant change: The man is dressed in a dark suit, white shirt, no tie.
Instant change: The woman <Picture 2> is dressed in a modern sleek ivory colored evening dress.
Instant change: The two people are standing at Mt. Bachelor Lodge in Oregon, with a huge view of Broken Top and the Cascade range. They are dressed and ready for an elegant dinner. No alcoholic drinks in view.
Frontal lighting on the two people clearly showing their faces. Both the man and the woman face the camera.
Instant change: Also in the room, without explanation is a full sized circus elephant.
It’s a bit harder to see here because the picture rendered large and had to be shrunk for this article. Here is the detail:
MiniMax H3 works by absorbing the reference photos and reinterpreting them,
and can change light and focus characteristics on them
Qwen 2.1 vs MiniMax H3
If you blow up just the faces, you can clearly see the differences between MiniMax H3 and Qwen 2.1 for image editing.
But I want to reiterate my main point – my need for image edit is not just faces and skin, it is to unlock the capability of MiniMax H3 to add random physical objects that the characters can use naturally. By ensuring the objects are “Native” to MiniMax, it improves the odds that the video will include them correctly and they will not become distractions.
Summary
I have not included the exact workflows, mostly because I still use the templates that ComfyUI provides for MiniMax. I have vibe-coded some tools to assist my queuing and also to capture the frame I want. But there’s really nothing special here – just a large size to compensate for softness, and a very short video time so that it fits in memory on my personal computer. I use 20 steps for video quality, and the 5-minute render time is long, but the quality makes up for it.
FAQ
Can I run MiniMax H3 locally?
Yes. MiniMax released open weights in August 2026, and ComfyUI supports H3 natively. I run it on my own laptop. For big jobs, I rent a cloud GPU. Nothing leaves my control.
How much VRAM do I need for MiniMax H3?
I run H3 on an RTX 5090 laptop with 64GB RAM and 24GB GPU. At current prices, this is about a $4,000 computer. I use int8 models and memory-saving attention. A 4K still takes about 5 minutes on the laptop. On a cloud-based RTX PRO 6000 instance, it takes under 2 minutes and costs under $0.05c per picture.
Can MiniMax H3 run on 8 GB of VRAM?
I have not tested 8 GB. The models are large. On 8 GB, expect slow renders and heavy system RAM usage. A rented cloud GPU is the practical answer.
Is MiniMax H3 free to use?
The models are free to download. But the license has conditions, and they differ for the US and EU. Read my guide to the MiniMax H3 license for US and EU users.
How good is MiniMax H3?
It is the best open video model I have used. It beats Wan 2.2 and LTX 2.3 on my PC. For image editing, Qwen-Image 2.1 is close in still quality. But for video start pictures, H3 wins. Every object it draws is native. Nothing morphs when the video starts.