Can I paste the Hugging Face link from my browser?

Only if it points at the file. The address you see while browsing a repository (…/blob/…) returns a web page, not weights. Use the download link (…/resolve/…) to the .safetensors file, and swap main for a commit hash if you want the same file on every run.

Can I load a LoRA I trained myself?

Yes, as long as it was trained on the MiniMax H3 open weights, saved as .safetensors and reachable at an HTTPS address. The request has no field for credentials, so a private repository needs a pre-signed link that stays valid until the job has run.

Which of the three H3 LoRA families should I start with?

This one, in most cases: it takes text, image and reference input and renders faster than Singularity in our tests. Move a shot to Singularity when fast motion smears or distant faces break up, and work on stills in MiniMax H3 Image LoRA while you are still choosing the look. The same LoRA links work in all three.

How do I compare two LoRAs fairly?

Hold everything else still: the same prompt, seed, resolution and duration, with only the loras entry changed. Judge at 480p and three seconds, where a comparison is cheap. A fixed seed keeps runs comparable rather than identical, so compare the look instead of hunting for pixel-level differences.

Can I use the same LoRA to make the opening frame?

Yes. MiniMax H3 Image LoRA runs the same weights as a still-image model, so a character or style LoRA gives you a matching frame there, which you then animate here on image-to-video with the same loras array. Motion and camera LoRAs are the exception: their effect only shows on video.

Can I load a LoRA of a real person?

Only with that person's consent and the rights to their likeness. Nothing on the platform reviews the weights you load, so this check sits with you: every person shown must be an adult, and the output must not be used to present someone as doing something they did not do.

MiniMax H3 LoRA API: Uncensored NSFW Video, Your LoRAs

minimax/h3-lora/text-to-video

MiniMax H3 LoRA text-to-video runs the open-weights MiniMax H3 with up to three of your own LoRAs. It returns a 3 to 15 second MP4 with stereo sound at 480p to 1080p in seven aspect ratios, with no filter on the prompt, weights or result.

From $0.05/s

Run model View API

Input

Text to video

Prompt318/7000

Aspect ratio

Resolutionaffects price

Durationaffects price

Style LoRAs1/3

Vh5tape Vhs

Seed

Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run

Example result

Extend

Video upscaling

Video editing

{
  "model": "minimax/h3-lora/text-to-video",
  "input": {
    "aspect_ratio": "4:3",
    "resolution": "480p",
    "duration_seconds": 5,
    "loras": [
      {
        "path": "https://huggingface.co/KennethFal/vh5tape-vhs-lora-minimax-h3/resolve/main/vh5tape.safetensors",
        "scale": 1
      }
    ],
    "seed": 42,
    "prompt": "vh5tape, worn VHS tape with light tracking noise. A woman in a red satin dress sits alone in a late-night diner booth, turns toward the camera and slowly smiles, neon light flickering in the rain-streaked window behind her. Slow push-in. Audio: rain on glass, a jukebox ballad muffled through the tape, quiet room hum."
  }
}

Introduction

MiniMax H3 LoRA Text to video turns a written scene into a 3 to 15 second clip with a stereo soundtrack, and lets you load up to three of your own LoRA files on top of the open-weights MiniMax H3. Pick 480p, 540p, 768p or 1080p and one of seven aspect ratios. The platform adds no content filter to the prompt, the weights or the output.

Choose a mode by starting material

What you have	Mode
a written scene	MiniMax H3 LoRA Text to video
an opening image	MiniMax H3 LoRA Image to video
reference images, clips or audio	MiniMax H3 LoRA Reference to video

Key features & highlights

Your LoRAs, loaded per request

Each request takes up to three LoRA files as direct links, each with its own strength from 0 to 4. There is no upload step and no allow-list, so any H3 LoRA you can link to is usable. We checked that the weights really load: a VHS-tape style LoRA turned a clean hotel-room prompt into tracking noise, colour bleed and a faded 1980s palette.

Picture and sound together

Every clip comes back as a 24 fps MP4 with a stereo soundtrack generated in the same pass. Describe the sound on its own Audio: line, from rain on glass to a line of dialogue, and it arrives synced to the picture.

Uncensored with custom weights

We ran a LoRA built for the strongest content on this endpoint, with our strongest prompts, and the clip came back exactly as written. Nothing is softened or swapped, and the weights you load are never reviewed.

Request parameters

Request parameters for MiniMax H3 LoRA · Text to Video
Parameter	Required	Description
`prompt`	Yes	Write the shot as a short timeline: who is in it, what they do, how the camera moves. Put the soundtrack on its own Audio: line. at least 1 characters · at most 7000 characters
`aspect_ratio`	No	16:9, 9:16, 1:1, 4:3, 3:4, 21:9 or 9:21. The ratio changes the frame, not the price: at the lowest tier a 16:9 clip comes back at 832 x 480. `16:9` · `9:16` · `1:1` · `4:3` · `3:4` · `21:9` · `9:21` · Default: `16:9`
`resolution`	No	480p, 540p, 768p or 1080p. Price rises with each step. `480p` · `540p` · `768p` · `1080p` · Default: `480p`
`duration_seconds`	No	Any whole second from 3 to 15. You pay for the seconds you ask for, so test at 3 to 5. `≥ 3` · `≤ 15` · Default: `5`
`loras`	No	Up to three {path, scale} entries. path is a direct HTTPS link to a .safetensors file trained for MiniMax H3; scale runs 0 to 4, 1 is neutral. Default: `[]` at most 3 items
`seed`	No	Keeps a series close when you compare LoRA scales or prompt edits. Change one thing per run. `≥ 0` · `≤ 2147483647`

Pricing

Billed per second of output at the chosen resolution. Loading LoRAs adds no separate charge.

Pricing
Specification	Price
`480p`	$0.05 / second
`540p`	$0.075 / second
`768p`	$0.10 / second
`1080p`	$0.20 / second

Use cases

Style tests before a campaign

Run the same short prompt with three different style LoRAs at 480p for three seconds and compare them side by side. It costs a few cents per look, and once you pick one you can re-run it at 1080p with the same seed.

Short scenes for companion and creator platforms

Load the concept or motion LoRA your audience expects, write the scene and the sound, and deliver at 768p or 1080p. There is no content flag to toggle and no second review after generation.

Retro and animated shorts

Film-stock, VHS and cartoon LoRAs change the medium, not just the grade. A 1939 hand-drawn cartoon LoRA and a VHS LoRA turn the same rooftop scene into two different productions.

Explore key use-case tutorials

Prompt guide & pro tips

Trigger first, then the shot, then the sound

Most community LoRAs expect a trigger word at the very start, such as vh5tape for the VHS LoRA. After it, write the shot as a short timeline and finish with an Audio: line. Restating what the LoRA already does usually fights it.

One scale change per run

With two or three LoRAs loaded, move one scale at a time and keep the seed fixed. Start style LoRAs at 1 and concept LoRAs at 0.6 to 0.8. If a clip looks overcooked, drop the strongest weight first.

Draft cheap, deliver once

Three to five seconds at 480p is the cheapest shape and usually comes back in well under a minute. Settle the look, the camera and the sound there, then spend on the 1080p version.

API quick start

# Generate the idempotency key once; reuse it when retrying after a lost response.
export SPICY_IDEMPOTENCY_KEY="${SPICY_IDEMPOTENCY_KEY:-$(uuidgen)}"
curl -X POST https://api.spicyapi.ai/api/v1/jobs/createTask \
  -H "Authorization: Bearer $SPICY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: $SPICY_IDEMPOTENCY_KEY" \
  -d '{
  "model": "minimax/h3-lora/text-to-video",
  "input": {
    "aspect_ratio": "16:9",
    "duration_seconds": 5,
    "prompt": "A slow aerial shot drifting over a misty pine forest at sunrise",
    "resolution": "480p"
  }
}'

# Copy data.taskId from the response into TASK_ID, then poll the status.
TASK_ID="TASK_ID_FROM_RESPONSE"
curl "https://api.spicyapi.ai/api/v1/jobs/recordInfo?taskId=$TASK_ID" \
  -H "Authorization: Bearer $SPICY_API_KEY"

See the API reference for authentication, the complete schema, polling, and error handling.

Related models and next tasks

Missing the model you need?

Tell us what to add and where you would use it. Requests go to the people who decide what gets onboarded next.

Your LoRAs on MiniMax H3, from the first test to the final cut

MiniMax H3 LoRA runs the open-weights H3 with up to three LoRA files you choose, on a written scene, a single opening frame or a set of references, and nothing on our side filters any of it. Loading the weights is the easy part. This page covers the part that decides whether the clip comes out right: which file to trust, how to point at it so it never changes under you, which endpoint should own the picture, and when a shot belongs on the Singularity or Image LoRA families instead.

Open the playground

Your LoRAs on MiniMax H3, from the first test to the final cut

Start here for: Text-to-video, and any shot where a quick render matters more than razor-sharp fast motion
Move to Singularity for: Spins, strikes and wide shots where faces have to hold
Move to Image LoRA for: Settling a look on stills before paying for seconds of video
One file, three families: The same H3 LoRA link loads on all three; we ran one file across every endpoint
What the link must return: The .safetensors file itself: a Hugging Face …/resolve/… link, not the …/blob/… page
Content filter: None from the platform; rights and consent stay with you

Pipeline: One set of weights, three stagesSettle the character and the style on stills, where a draw takes seconds. Confirm the movement with a three-second 480p clip here. Send only the shots with fast action on to Singularity. The loras array travels unchanged through all three, so nothing has to be re-tuned between stages.
Frame or weights: Decide which one owns the pictureOn image-to-video your opening frame and a strong LoRA compete for the same pixels. If the LoRA's world is the point, write the scene on text-to-video or reference-to-video and let the weights lead. If your frame is the point, stay on image-to-video with the LoRA turned down and the prompt about what happens next.
Ecosystem: The H3 LoRA scene is under two months old, so vet every fileMiniMax published the open weights on 3 August 2026, so every H3 LoRA you can download is new, and quality swings from polished to half-trained. A good model card names the base, the trigger word and a working strength. A file that says none of that deserves one cheap test before it goes anywhere near a batch.

Getting LoRAs right on MiniMax H3 LoRA

Borrow the words the LoRA was trained on

A LoRA learns from captioned clips, and it answers best to the vocabulary of those captions. Open the model card, copy the trigger word and the phrasing of its sample prompts, then write your scene in that dialect. A camera LoRA captioned with 'orbit shot' responds more reliably to that phrase than to any paraphrase of it.

Pin the file so next month's clip matches this one

A link that runs through /resolve/main/ follows whatever the author uploads next, so the same request can quietly change its look. Replace main with the commit hash from the repository history and the address points at one frozen file for good. Keep that pinned link in your configuration, next to the scale it was tuned at.

Test looks on stills, test movement on clips
Character and style LoRAs show up on a single frame, so MiniMax H3 Image LoRA answers 'is this the right look' for the price of a still. Motion and camera LoRAs cannot be judged that way, because a frame has no movement in it. Run those here, at 480p and three seconds, the smallest clip this family renders.
Debug a stack by taking slots away
When a clip with three LoRAs goes wrong, hold the seed and the prompt and re-run with one entry removed at a time. The run that recovers names the file at fault in at most three short requests, which beats nudging three scales at once and guessing which nudge helped.
No review queue means your pipeline is the review
Nothing on our side checks the weights you load or the clip you get back, which is exactly what makes uncensored LoRAs usable here. It also means the checks your business depends on belong in your own pipeline: that everyone shown is an adult, that you hold the rights to any likeness, and that no one is being misrepresented.

Who uses MiniMax H3 LoRA

Studios with a house style
Turn a look into a versioned file
Pick or train one style LoRA, pin it by commit and route every clip through it. The look stops living in a prompt that each editor rewrites and starts living in a file you can version, roll back and hand to a new team member in one line of configuration.
Platforms letting users bring LoRAs
Check the link before you forward it
The weights are fetched when the job runs, so a bad link from a user comes back later as a failed task. Check it in your own app first: the address answers, it serves a .safetensors file, and its size sits within a limit you set to keep first-call times predictable.
Pitch and storyboard teams
An animatic in the client's style
Load the style LoRA the campaign will use, render each storyboard panel as a three-second 480p clip and cut them together. The client sees motion, pacing and the final look in one pass, long before a finished sequence is paid for, and the approved panels re-render at 1080p with the same array.
People training their own H3 LoRAs
A hosted test bench for every checkpoint
Push each checkpoint to a URL, run the same short prompt battery against each one with a fixed seed, and compare the clips side by side. There is no GPU to rent and no local install to keep in step with the base model, and uncensored checkpoints run through the same fields as everything else.

From a new LoRA to a delivered clip on MiniMax H3 LoRA

Read the model card before you download anything
Confirm the LoRA was trained on MiniMax H3, then note three things: the trigger word, the strength its author recommends and two or three of its sample prompts. Those three facts decide how your first prompt should read.
Copy the pinned download link
On Hugging Face, take the …/resolve/… link to the .safetensors file and replace main with a commit hash. The address must serve the file without a login, because the request carries only a path and a scale, never credentials.
Choose the endpoint by what should own the picture
Words only: text-to-video. Your frame must survive: image-to-video with a gentle scale. A face, a room and a move from separate assets: reference-to-video. The loras array is identical on all three, so switching later costs nothing.
Run a three-rung strength ladder
Send the same prompt and seed three times at 480p for three seconds, with the new LoRA at 0.5, 0.8 and 1.1. Keep the rung that reads right, or split the difference between two. The whole ladder is nine seconds of the lowest tier.
Deliver the keeper and collect it by callback
Re-send the winning request at 768p or 1080p with the same seed and expect a close variation rather than a copy. Renders are asynchronous: pass a callback URL or poll the task until it finishes, and the MP4 arrives with its soundtrack inside.

MiniMax H3 LoRA prompts worth copying

A strength ladder for a style LoRA

Three runs of this prompt at 0.5, 0.8 and 1.1 with the VHS-tape LoRA loaded show where the tape look starts to swallow the scene. The trigger word leads because that is how the LoRA was captioned.

vh5tape, a birthday party in a wood-panelled living room in 1987, a man in a striped sweater carries a cake with lit candles toward the camera, friends clapping around him, handheld home-video camera. Audio: everyone singing, a dog barking once

Continue the frame instead of replacing it
On image-to-video the prompt only says what happens next inside the uploaded picture, so the room, the light and the person carry through while a gently scaled LoRA shapes the movement.
```
She sets the glass down on the windowsill, turns toward the camera and smiles, the city lights behind her staying exactly where they are, slow push-in. Audio: distant traffic and a soft exhale
```
Write in a camera LoRA's own vocabulary
Camera-move LoRAs answer to the phrases in their training captions. Put that phrase first, name the subject, and let the weights handle the path of the lens.
```
orbit shot, a dancer in a red dress spins on a rooftop at dusk, her skirt flaring outward, the city skyline turning slowly behind her. Audio: a brass band and wind
```

Workbench

Input

Output

Introduction

Choose a mode by starting material

Key features & highlights

Your LoRAs, loaded per request

Picture and sound together

Uncensored with custom weights

Request parameters

Pricing

Use cases

Style tests before a campaign

Short scenes for companion and creator platforms

Retro and animated shorts

Prompt guide & pro tips

Trigger first, then the shot, then the sound

One scale change per run

Draft cheap, deliver once

API quick start

Related models and next tasks

Your LoRAs on MiniMax H3, from the first test to the final cut

Getting LoRAs right on MiniMax H3 LoRA

Borrow the words the LoRA was trained on

Pin the file so next month's clip matches this one

Test looks on stills, test movement on clips

Debug a stack by taking slots away

No review queue means your pipeline is the review

Who uses MiniMax H3 LoRA

Turn a look into a versioned file

Check the link before you forward it

An animatic in the client's style

A hosted test bench for every checkpoint

From a new LoRA to a delivered clip on MiniMax H3 LoRA

Read the model card before you download anything

Copy the pinned download link

Choose the endpoint by what should own the picture

Run a three-rung strength ladder

Deliver the keeper and collect it by callback

MiniMax H3 LoRA prompts worth copying

A strength ladder for a style LoRA

Continue the frame instead of replacing it

Write in a camera LoRA's own vocabulary

More MiniMax H3 LoRA questions

Can I paste the Hugging Face link from my browser?

Can I load a LoRA I trained myself?

Which of the three H3 LoRA families should I start with?

How do I compare two LoRAs fairly?

Can I use the same LoRA to make the opening frame?

Can I load a LoRA of a real person?

All MiniMax models11

Video models7

Audio models3

Image models1

Related comparisons

LoRA guide