MiniMax H3Pick a LoRA, type a prompt, get video with sound. Nothing to install.
Reference Media
Max 12 total: 9 images, 3 videos, 3 audio. Media 2–15s; files 10/50/10MB.
Interactive controls become available after authentication finishes loading.
Neon Pop Anime Fashion Film
A rebellious neon-pop heroine transforms through three distinct anime fashion identities in a high-impact 15-second editorial film.
Every other way to run a MiniMax H3 LoRA asks you to do the same three things first: download the weights, set up ComfyUI, and own a GPU big enough to hold a 33-billion-parameter model. This page skips all of it. Choose a LoRA below, describe your shot, and the render happens on our hardware. Signing up puts 20 credits on your account — enough for one 5-second 768p LoRA clip, or two at 480p, no card needed. One thing to know up front: LoRA runs output at 768p. The open-source LoRAs built for MiniMax H3 do not support 2K yet. Everything else about the model is unchanged, including the audio. Independent third-party service. Not affiliated with, endorsed by, sponsored by, or operated by MiniMax.
A LoRA changes what the base model is good at. The same prompt through two different LoRAs gives you two different films. Below is what is currently loaded, grouped by what each one actually does.
Each option below is available in the generator. Most card samples are our own output. The MiniMax H3 Acc LoRAs comparison was published by Alibaba PAI and is identified on its dedicated page.
Turns flat prompts into moving shots. The base model tends toward a locked-off camera unless you fight it; this LoRA responds to dolly, crane, orbit, and push-in language the way a camera operator would.
Use when: The subject is right but the shot feels static.
Martial arts choreography with weight behind it — strikes that land, stances that hold, wire-work that reads as deliberate rather than floaty. Trained on wushu and kung fu footage.
Use when: You need fight sequences the base model turns into flailing.
Objects behave like they have mass. Things fall, collide, splash, and settle instead of drifting through each other. This is the one that fixes the “AI video” tell most viewers notice first.
Use when: Anything with object interaction — pouring, dropping, breaking, colliding.
Clean line art anime rendering. Describe an ordinary scene and get it back in a specific drawn language rather than the base model’s semi-realistic default.
Use when: Anime output that holds its style across a whole clip.
Replace one thing inside a finished clip and leave everything else alone. Mask the subject, supply a reference image, and the background, the camera move, and the original soundtrack all survive untouched.
Use when: A clip is right except for one element — swap the product, change the animal, replace a logo — and re-rolling the whole shot would lose what already works.
Improves human-centered shots with natural skin texture, coherent close-up faces, believable expressions and gestures, film-style lighting, and subtle documentary camera movement.
Use when: Portraits, conversations, workers, travel scenes, crowds, or any shot where people need to feel less synthetic.
Read the Realism People LoRA guideHelps supported MiniMax H3 workflows finish sooner while keeping the scene close to the original result. It is a practical choice when you want to test prompts, compare directions, and refine a video with less waiting.
Use when: Iteration speed matters for prompt, keyframe, or reference-guided video creation.
Use MiniMax H3 Acc LoRAsA MiniMax H3 LoRA is a small trained adapter that attaches to the MiniMax H3 video model and biases its output toward a specific style, motion pattern, or subject, without retraining the model itself.
The base model is a generalist. It has seen enough footage to make a competent version of almost anything and a specialist version of almost nothing. A LoRA narrows that. Someone trains one on a few hundred clips of wire-fu choreography, and from then on the model knows what that kind of motion looks like at a level no prompt could describe.
The file itself is small — usually tens or low hundreds of megabytes against the 33-billion-parameter base. That is the whole idea: you are not shipping a new model, you are shipping a correction to an existing one. It also means a LoRA only works with the model it was trained against. A MiniMax H3 LoRA will not load into Seedance or Wan.
Most style and motion LoRAs come with trigger words — specific phrases used during training. Putting those words in your prompt helps activate the intended behaviour, and each relevant card lists them. MiniMax H3 Acc LoRAs are the exception: select the option and use a normal video prompt without a trigger word.
MiniMax H3 Acc LoRAs focus on faster generation rather than changing the visual style. They are available directly in the generator, and you can compare their output on the MiniMax H3 Acc LoRAs page .
The LoRA decides what kind of prompt works. A camera-motion LoRA wants shot language — dolly in, handheld, rack focus. A style LoRA wants subject and scene and will handle the look itself. Choosing first means writing once.
Most style and motion LoRAs use trigger words, and each relevant card lists them. Paste those words into your prompt alongside the shot description. MiniMax H3 Acc LoRAs are different: select the option and write a normal prompt without adding a trigger word.
LoRA runs output at 768p and cost 3 credits per second (2 at 480p), so a 5-second 768p test is 15 credits. Weight is the setting most likely to be wrong on a first attempt — find the right one on 5-second clips before you commit to a 15-second render, which costs three times as much for the same mistake.
Most LoRAs have a weight you can move. Low weight leaves the base model mostly intact and nudges it. High weight commits to the LoRA's look and can overpower the rest of your prompt. When output goes strange — melted faces, ignored instructions — the weight is usually the cause, not the prompt.
For option-specific setup, recommended strengths, ready-to-use prompts, and video examples, read the MiniMax H3 LoRA user guide .
If you got here after reading a ComfyUI workflow thread or a Hugging Face model card, you already know the local route exists. It is worth being clear about why a lot of people cannot take it.
MiniMax published H3's base weights on August 3, 2026. Open weights and unrestricted use are not the same thing: the MiniMax H3 Community License excludes local deployment in the United States, the European Union, the United Kingdom, and South Korea.
Most write-ups about running H3 locally skip this line. If you are reading from one of those places, the LoRA is not the obstacle — the base model underneath it is.
H3 is a 33-billion-parameter model that decodes video and audio in the same pass. Checking it against a consumer GPU usually ends the conversation. You can quantize far enough to fit, but the output degrades as you do, and a LoRA on top does not improve the memory picture.
No weights on your drive. No ComfyUI graph to debug when a node updates. No GPU to buy or rent for an afternoon of testing. You pick a LoRA from a list, and the render runs on managed hardware.
The trade is real and worth naming: you do not control the runtime, and you pay per render rather than per month. If you are outside the restricted regions, own the hardware, and generate all day, running it yourself is the better deal and we would rather say so.
If you do not need a LoRA at all, the base MiniMax H3 AI video generator runs 2K with the same native audio.
Using a LoRA does not change how billing works. Output is charged by the second: 768p costs 3 credits per second and 480p costs 2, so a LoRA run comes in below the same clip at 2K.
Your 20 free credits cover one 5-second 768p LoRA clip, or two at 480p.
Text prompts cost nothing extra. Reference images and audio cost 1 credit each. Reference videos add 2 credits per second at 480p or 3 at 768p, so a long clip attached to a short render can cost more than the render itself.
Failed or rejected generations are refunded automatically — including the ones where a LoRA weight was set too high and the output came back unusable.
New MiniMax H3 LoRAs appear on Hugging Face most weeks. Most of them are not on this page, and that is deliberate — the list is what survived three filters, not everything we could have loaded.
The license has to permit commercial use. A lot of community LoRAs ship non-commercial or with no license at all, and no license means all rights reserved rather than free to use. We leave those alone even when they are the most popular thing in the category.
It has to work at more than one setting. A LoRA that only produces something usable at one exact weight is not worth spending your credits to discover.
And we do not host LoRAs trained on real people's likenesses, regardless of how they are licensed.
We usually leave out speed-only experiments that cannot be used through this website. MiniMax H3 Acc LoRAs are included because they are available in the online generator and have clear comparison videos that users can review before generating.
A MiniMax H3 LoRA is a small trained adapter that attaches to the MiniMax H3 video model and biases its output toward a particular style, motion pattern, or subject. It does not replace the base model — it adjusts one. Files are usually tens to low hundreds of megabytes, and each LoRA only works with the model it was trained against, so a MiniMax H3 LoRA will not load into another video model.
Not here. Running LoRAs locally normally means a ComfyUI graph, downloaded weights, and a GPU with enough memory for a 33-billion-parameter model. On this page you select a LoRA from a list and generate in the browser. ComfyUI remains a good option if you are outside the license-restricted regions and own the hardware.
Depends where you are. The MiniMax H3 Community License excludes local deployment in the United States, the European Union, the United Kingdom, and South Korea. Outside those regions, the base weights are on Hugging Face and the model runs through ComfyUI, vLLM, SGLang, and diffusers — the constraint then becomes VRAM rather than licensing.
For style and motion LoRAs, first check the trigger words shown on the selected card and then adjust the strength. A low strength may leave the result almost unchanged, while a high strength can distort it or overpower the prompt. MiniMax H3 Acc LoRAs do not need trigger words; select the option and describe the video normally.
No — a LoRA run actually costs less than 2K. LoRA output at 768p costs 3 credits per second, so a 5-second clip is 15 credits against 20 for the same clip at 2K without a LoRA. At 480p, LoRA output is 2 credits per second. Your 20 free credits cover one 768p clip, or two at 480p. Failed generations are refunded automatically, including runs ruined by a bad weight setting.
Not currently. You can request that we load a public one by sending the model card link to support@minimaxh3video.app. It needs a license permitting commercial use and it has to hold up in our testing before it goes on the list.
Videos generated on a paid pack come with commercial use rights from us, but the LoRA’s own license also applies. We only host LoRAs whose licenses permit commercial use, and each card states the license so you can check it before building a campaign on it.
Pick by the problem, not the look. If your shots feel static, start with a camera-motion LoRA. If the subject drifts between generations, start with a consistency one. If the content is right but the aesthetic is generic, start with a style LoRA. Trying to fix all three at once with a single LoRA is how people conclude LoRAs do not work.
Start with a 768p draft so a wrong guess costs 10 credits instead of 20. Pick an option, follow the setup shown on its card, and describe the shot. Your 20 free credits are already on the account.