Is MiniMax H3 LoRA uncensored?

On our API, yes. We add no content filter to the prompt, the LoRA weights, or the output. With no LoRA loaded, text-to-video and reference-to-video passed every rung of our content ladder on 27 September 2026 and image-to-video softened only L4 Intense, and with a community adult LoRA loaded it rendered our L4 Intense probe on all three endpoints on 26 September.

Can I run a civitai or Hugging Face LoRA through the API?

Yes, as long as it was trained for MiniMax H3 and you can point at a direct HTTPS link to the .safetensors file. Nothing is uploaded to us and nothing about the file is reviewed; it's fetched when the request runs.

Does loading a LoRA cost extra?

No. You pay per second of output at the resolution you choose, whether you load zero, one, two or three LoRAs.

Why did my LoRA seem to do nothing?

The most common reason is that it was trained for a different base video model. It loads without an error and the clip renders like a plain MiniMax H3 output. Confirm a LoRA actually loads by running one short 480p clip with its trigger word, if it has one, at the start of the prompt.

How is this different from MiniMax H3?

Base MiniMax H3 on this site runs MiniMax's hosted model with no custom weights, tops out at a 2k tier and costs less per second. MiniMax H3 LoRA runs the open weights with your own LoRAs, adds 540p and 1080p tiers, allows clips as short as 3 seconds, and adds reference video and audio input, but its top resolution is 1080p.

How is this different from MiniMax H3 Singularity LoRA?

Singularity is a community fine-tune of the same open weights, aimed at crisper fast motion and cleaner faces. It only offers image-to-video and reference-to-video, at a slightly higher per-second rate, with the same three-LoRA cap.

How much does MiniMax H3 LoRA cost?

Text-to-video and image-to-video: $0.05/s at 480p, $0.075/s at 540p, $0.10/s at 768p, $0.20/s at 1080p. Reference-to-video: $0.06/s at 480p up to $0.27/s at 1080p, plus $0.02 for each reference image or audio clip; reference video seconds count toward the 15-second output cap. Prices from our public catalogue on 27 September 2026.

MiniMax H3 LoRA Review: Your Own LoRAs, No Content Filter

Pick MiniMax H3 LoRA when you need your own weights on H3 video: three LoRAs a call, no review step, and every rung we ran came back uncensored.

Create in Studio Open the Playground

Spicy Index

79.6/ 100

Preliminary

Freedom Score

98.3/ 100

Content levels

Tested 6 of 6 probe types · 60 runs

Mild · The lightest prompts in the test
Moderate · Where filtered models start to refuse
Strong · Mid-scale; must be tested before a model is ranked
Very strong · Where most models start to hold back
Maximum · The top of the scale, scored in two halves: class A and class B

Price: $0.50 per 5-second clip
Median time: 77 s
Success rate: Collecting data

MiniMax H3 LoRA is the open-weights MiniMax H3 with a loras field: point up to three .safetensors links at it, each with its own strength, and they load with no upload, no allow-list and no review of what they contain. On 27 September 2026 we ran all three endpoints with no LoRA loaded through our full content ladder: nothing was refused, and image-to-video pulled back at one rung, L4 Intense, where two of three runs implied the scene instead of showing it. On 26 September, with a community adult LoRA loaded, all three endpoints (text-to-video, image-to-video, reference-to-video) rendered our L4 Intense probe exactly as written; image-to-video also held L3 Strong.

SpicyAPI Engineering
Tested Sep 27, 2026
Methodology v3.3 →
We sell these models; scores ignore price →

What MiniMax H3 LoRA made in our tests

Straight from our test runs. Copy a prompt or open it in the playground to try it yourself.

Text to video
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.
Try this model
Text to video
gulliv3r, 1939 hand-drawn cartoon. Two lovers lean on a rooftop railing at dusk above a glowing city, her head on his shoulder, wind moving her hair, then he twirls her once. Audio: scratchy big-band record, distant car horns.
Try this model
Image to video
From this exact frame: he turns away from the window, walks toward the camera tying the belt of his robe, and stops with a half smile. Slow handheld push-in. Audio: morning city hum through the glass, bare feet on carpet.
Try this model
Image to video
vh5tape, worn VHS tape. From this exact frame: she leans into the microphone and sings a slow line with her eyes closed, then opens them and smiles as the crowd claps softly off-camera. Audio: a torch song through tape hiss, soft applause.
Try this model
Reference to video
Picture 1 is the woman, Picture 2 is the hotel room. She walks into the room at night, drops her coat on the bed and turns to the mirror, pushing her hair back. Slow dolly. Audio: door closing, heels on carpet, a low bass line from the next room.
Try this model
Reference to video
Replace the dancer from Video 1 with the woman from Picture 1: her face, her dark hair and her black long-sleeve outfit. She copies the fast floor spin from Video 1 in a bright white photo studio with a seamless backdrop, then rises and smiles at the camera. Only the motion comes from Video 1; the person comes from Picture 1. Audio: an upbeat drum loop, sneakers squeaking on the floor.
Try this model

Make your own with MiniMax H3 LoRA

Create in Studio Open the Playground

The verdict

Pick it when you already have LoRAs, whether that's a character, a style or an adult fine-tune, and want them on video without asking anyone's permission. Skip it if you have no LoRA to load: base MiniMax H3 runs the same content ladder for less per second and adds a 2k tier.

How it scores. This page's score card carries the current Spicy Index and Freedom Score, recalculated as we add runs; we don't repeat the numbers here because they move as testing continues. See How far it goes for what those extra rounds actually tested.

Try MiniMax H3 LoRA →

Key facts

Content ladder, no LoRA: on 27 Sep 2026 text-to-video and reference-to-video passed every rung, from L1 Suggestive to both L5 Extreme probe types; image-to-video softened L4 Intense in two of three runs. Zero refusals.
Content ladder, with an adult LoRA: text-to-video, image-to-video and reference-to-video all rendered our L4 Intense probe on 26 Sep 2026, one run per endpoint; image-to-video also held L3 Strong.
LoRA input: up to 3 entries, each a direct HTTPS link to a .safetensors file plus a strength from 0 to 4; no separate charge for loading one.
Price: $0.05/s at 480p on text-to-video and image-to-video ($0.25 for 5 s), rising to $0.20/s at 1080p; reference-to-video starts at $0.06/s plus $0.02 per reference image or audio clip.
Limits: 3 to 15 second clips, 480p to 1080p, stereo sound in the same pass; weights must be trained for MiniMax H3 or they load and do almost nothing.
Sources: the full parameter list is on the MiniMax H3 LoRA model page, scoring rules are on our review methodology page, and our wider ladder run across six LoRA families is summarised in How far it goes.

What we liked and what held it back

What we liked

No review step for the weights you load. A loras entry is a link and a number; the file is fetched when the request runs and nothing about it is inspected first.
Text and reference input passed every rung with no LoRA loaded. L1 Suggestive through both L5 Extreme probe types (restraint/impact and gore) all rendered on 27 September, across 3 runs per rung.
An adult LoRA rendered on all three endpoints. Text, image and reference input all held our L4 Intense probe with a community LoRA loaded, on 26 September.
Style LoRAs are obviously real. The VHS-tape LoRA in our diner example visibly adds tape noise and colour bleed; the 1939 cartoon LoRA turns the same kind of prompt into rubber-hose animation. Neither is a filter you could get from the base model alone.
Picture and sound in one pass. Every clip returns as an MP4 with a stereo soundtrack generated alongside the picture, described on its own Audio: line in the prompt.
Reference-to-video stacks with LoRAs. Up to nine images, three clips and three audio clips sit under the same three-LoRA cap, so a character LoRA and reference stills of that person can reinforce each other.

What held it back

A LoRA for another base model does nothing you'd notice. It loads without an error and the output looks like a plain H3 render. We didn't see a warning anywhere in the response.
A strong concept LoRA can steal the opening frame. On image-to-video, a concept LoRA at full strength combined with a prompt describing a different scene replaced our starting picture; describing what happens next in the same frame and starting the LoRA at 0.6 to 0.8 kept it.
Large LoRAs slow the first call. An 86 MB LoRA added noticeable time to a 5-second 480p clip in our testing, since the file has to be fetched before generation starts.
Reference clips eat the duration budget. Reference video seconds count toward the 15-second output cap, so a long reference clip leaves little room for new motion.
Image-to-video pulls back at L4 Intense without a LoRA. Two of three runs implied the scene instead of showing it on 27 September; the other two endpoints rendered it every time.
Style LoRAs fade when a photo leads. From a photographic opening frame or reference, a style LoRA barely changes the look; the VHS and cartoon LoRAs show clearly on text-to-video.

What the clips look like

The gallery below holds published, safe-for-work examples from the MiniMax H3 LoRA model page, shown as rendered. Our content-ladder test clips aren't shown; we report those results in text only.

A style LoRA changes the frame, not just the color grade. In the diner clip, a plain prompt plus a VHS-tape LoRA at strength 1 produces tape noise, chroma bleed and a washed-out palette that a text prompt alone wouldn't get you. The 1939 cartoon clip runs the same idea in the opposite direction: a different LoRA turns the same kind of scene into hand-drawn animation with a scratchy big-band soundtrack.

Realism People LoRA on both image-to-video and reference-to-video. At strength 0.8 it keeps skin and fabric believable as a subject moves, without overriding the opening frame from a still made on MiniMax H3 Image LoRA, or the two reference stills that pin a woman and a hotel room in the reference-to-video clip.

Reference video carries motion, not identity, once you say so. The spin clip takes a still of a performer and a five-second reference clip of someone else's floor spin; the prompt states plainly that the person comes from the picture and only the motion comes from the clip. Leaving that sentence out the first time brought the wrong dancer along with the move.

Prompts that worked

Put a LoRA's trigger word, if it has one, at the very start of the prompt, then write the shot as a short timeline, then finish with an Audio: line. Restating what the LoRA already does tends to fight it rather than reinforce it.

1. Style LoRA with its trigger word (text-to-video, 480p, 5 s). The prompt behind our diner clip, with the VHS-tape LoRA at strength 1.

text

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.

Run this prompt on MiniMax H3 LoRA →

2. What happens next, not a new scene (image-to-video). Keeps a strong concept LoRA from pulling the clip away from the opening frame.

text

From this exact frame: he turns away from the window, walks toward the camera tying the belt of his robe, and stops with a half smile. Slow handheld push-in. Audio: morning city hum through the glass, bare feet on carpet.

Run this prompt on MiniMax H3 LoRA →

3. Naming each reference by number (reference-to-video). Say plainly which input supplies the person and which supplies the motion, or the model can bring both from the clip.

text

Replace the dancer from Video 1 with the woman from Picture 1: her face, her dark hair and her black long-sleeve outfit. She copies the fast floor spin from Video 1 in a bright white photo studio with a seamless backdrop, then rises and smiles at the camera. Only the motion comes from Video 1; the person comes from Picture 1. Audio: an upbeat drum loop, sneakers squeaking on the floor.

Run this prompt on MiniMax H3 LoRA →

How far it goes

We ran two separate tests. On 27 September 2026 we probed all three endpoints with no LoRA loaded: 3 runs per probe type across L1 Suggestive, L2 Mature themes, L3 Strong, L4 Intense and both L5 Extreme probe types (restraint/impact and gore). Text-to-video and reference-to-video rendered every rung as written. Image-to-video did the same except at L4 Intense, where two of three runs implied the scene instead of showing it. That was 54 runs on this family, part of 216 clips and stills across six LoRA families that day, and none was refused.

On 26 September 2026, with a community adult LoRA loaded at default strength, we ran one probe per endpoint: text-to-video, image-to-video and reference-to-video all rendered our L4 Intense probe as written, with no refusal and nothing swapped for a safer scene; image-to-video also held L3 Strong. The image-to-video endpoint held the opening frame once the prompt described what happens next in that frame rather than a different scene; a strong concept LoRA at its default strength, combined with a prompt that contradicted the picture, was what moved the clip away from the frame in an earlier attempt.

Not tested yet: the upper rungs with a LoRA other than the one adult fine-tune we used; anime or illustrated LoRAs at L4 Intense or L5 Extreme; a LoRA stacked with reference audio; clips longer than the ones in our probes.

The page label is Uncensored: we add no content filter on our side, on this endpoint or on any LoRA weights you point at. Adults only, in every still, prompt and output. Never depict a real, identifiable person without their consent.

LoRA compatibility

Three community LoRAs, all trained for MiniMax H3, loaded and visibly changed the output in our testing.

VHS tape style LoRA, strength 1: works on text-to-video. Tape noise, colour bleed and a washed-out palette are obvious in the diner scene. On image-to-video from a photographic still, the same LoRA at 0.8 barely showed: the photo set the look.
1939 hand-drawn cartoon LoRA, strength 1: works on text-to-video. The same kind of prompt that would otherwise render photoreal turns into rubber-hose animation with painted backgrounds.
Realism People LoRA, strength 0.8: works. Keeps skin and fabric convincing as a subject moves, on both an image-to-video still and a reference-to-video shot built from two reference images.

A LoRA trained for a different base model loads without an error but does almost nothing we could see; if a clip looks like a plain, un-styled H3 render after loading a LoRA, that's the most likely reason, not a broken request. Concept and anatomy LoRAs are the ones most likely to overpower a scene at full strength; the style LoRAs above were fine at 1, while the realism LoRA held cleanly on top of an existing frame at 0.8. Style LoRAs show most on text-to-video. When a photo is the opening frame or the reference, expect the photo's look to win: in a same-seed check on 27 September, the 1939 cartoon LoRA left image-to-video and reference-to-video clips from a photo looking photographic even at strength 2.5.

What the community says

MiniMax H3 LoRA runs the same open-weights model people already discuss for its lack of filtering: "Minimax has NO filters of any kind". On the LoRA side specifically, feedback on the community repositories we used in our own testing runs from straightforward praise, "Just wanted to say this is a great lora, well done", to setup trouble that has nothing to do with content policy: one user reported a LoRA producing distorted output in their ComfyUI workflow while "Other Loras work fine" in the same setup, a reminder that a LoRA that works on our endpoint can still misbehave in a different pipeline, and vice versa.

Alternatives

No LoRA to load? MiniMax H3 runs the same base weights for less per second, adds a 2k tier, and passed the same content ladder in our testing. Our H3 review.
Want the uncensored tune without any weights of your own? MiniMax H3 Spicy animates a still with no prompt required and no loras field to manage. Our H3 Spicy review.
Crisper fast motion and faces? MiniMax H3 Singularity LoRA is a community fine-tune of the same base with the same three-LoRA cap, on image-to-video and reference-to-video only, from $0.0625/s at 480p. Our comparison.
A different base model for your LoRAs? Wan 2.2 LoRA image-to-video starts at $0.024/s at 480p, splits high-noise and low-noise LoRA slots, and has no sound output.
An uncensored build of that same Wan base? Wan 2.2 Spicy LoRA adds a video-extend endpoint on top of image-to-video.

How far MiniMax H3 LoRA goes

Each level, on each task, based on confirmed results. A level earns points only when the content actually appears, and a single run earns 60% at most.

Freedom points by level and task
Level	Image to video	Reference to video	Text to video
MildThe lightest prompts in the test	3/3 delivered20.0 / 20 pts	3/3 delivered20.0 / 20 pts	3/3 delivered20.0 / 20 pts
ModerateWhere filtered models start to refuse	3/3 delivered20.0 / 20 pts	3/3 delivered20.0 / 20 pts	3/3 delivered20.0 / 20 pts
StrongMid-scale; must be tested before a model is ranked	5/5 delivered20.0 / 20 pts	3/3 delivered20.0 / 20 pts	3/3 delivered20.0 / 20 pts
Very strongWhere most models start to hold back	2/4 delivered15.0 / 20 pts	4/4 delivered20.0 / 20 pts	5/5 delivered20.0 / 20 pts
MaximumThe top of the scale, scored in two halves: class A and class B	Class A: 3/3 deliveredClass B: 3/3 delivered20.0 / 20 pts	Class A: 3/3 deliveredClass B: 3/3 delivered20.0 / 20 pts	Class A: 3/3 deliveredClass B: 3/3 delivered20.0 / 20 pts
Task score	95.0 / 95.0Lowest: pulls the average downTested Sep 27, 2026	100.0 / 100.0Tested Sep 27, 2026	100.0 / 100.0Tested Sep 27, 2026

Tested Sep 26, 2026 to Sep 27, 2026

This site stays safe for work: results are written up as text, any image shown is pixel-censored, and test prompts are never published. We never test anyone who looks underage, real people’s likenesses or non-consent.

See every test record →See the uncensored leaderboard →

Run your own prompt and see how far it goes

Create in Studio Open the Playground

How MiniMax H3 LoRA earns its Spicy Index

Each dimension below links to its full workings: the table row every point came from and why points were lost.

Public leaderboards 45% of quality: Not yet tested
Blind output test 35% of quality: 68.7
Arena quality 20% of quality: Not yet tested
Capability 10.5% weight: 87.8
Freedom Score 30% weight: 98.3

Only our blind test covers this model so far, and its agreement with public human-vote leaderboards is limited on this board. The preliminary Index is ranked by its current score; independent evidence is still pending.

Open the full test report →LoRA support75.0 / 100

Capability, item by item

Points from the public scoring table. Upscaled tiers and limits we set ourselves don’t count.

Native max resolution1080p: 25.0 / 25
Longest single clip15 s: 9.0 / 20
Native audioYes: 15.0 / 15
Reference inputs9 reference images · 6 video or audio: 15.0 / 15
Frame controlFirst and last frame: 10.0 / 10
Task coverageImage to video, Reference to video, Text to video, Video extension: 10.0 / 10
Customization (LoRA)LoRA supported: 3.8 / 5

What MiniMax H3 LoRA can do on SpicyAPI

Straight from the live model specification on SpicyAPI. Some creators offer more in their own apps; this is what you can call through our API today.

Capabilities by endpoint
Feature	Image to video	Reference to video	Text to video	Video extension
Resolutions	480p · 540p · 768p · 1080p	480p · 540p · 768p · 1080p	480p · 540p · 768p · 1080p	480p · 540p · 768p · 1080p
Clip length	3–15 s	3–15 s	3–15 s	3–15 s
Aspect ratios	—	16:9 · 9:16 · 1:1 · 4:3 · 3:4 · 21:9 · 9:21	16:9 · 9:16 · 1:1 · 4:3 · 3:4 · 21:9 · 9:21	—
Native audio	Yes	Yes	Yes	Yes
Reference inputs	—	15	—	—
Seed control	Yes	Yes	Yes	Yes
Negative prompt	No	No	No	No

See every parameter and code sample →

How long it really takes

Measured on real jobs over the last 30 days, from request to finished result. Times include our own test runs.

Median time: 77 s
90th percentile: 4.4 min
Wait per second of video: 10.6 s per output second
Success rate: Collecting data

A time appears once at least 10 completed jobs are in. Success rates count customer jobs only and stay hidden for every model until we have a full 30 days of traffic.

What people say about MiniMax H3 LoRA

2 positive · 1 mixed · 0 negative

Minimax has NO filters of any kind
hacker-news· Aug 11, 2026· PositiveSource
Just wanted to say this is a great lora, well done
hugging-face· Sep 1, 2026· PositiveSource
Other Loras work fine. Any ideas what could have gone wrong with my setup?
hugging-face· Sep 3, 2026· MixedSource

Quotes are verbatim from public posts, linked to the original. They are not ratings and don’t feed any score.

Models to consider instead of MiniMax H3 LoRA

MiniMax
MiniMax H3
No LoRA to load?
Spicy Index
78.0−1.6
Freedom Score
70.7−27.6
Price
$0.185−63%/ 5 s clip
Compare side by side Try
MiniMax
MiniMax H3 Spicy
Want the uncensored tune without any weights of your own?
Spicy Index
64.3−15.3
Freedom Score
97.5−0.8
Price
$0.42−16%/ 5 s clip
Compare side by side Try
MiniMax
MiniMax H3 Singularity LoRA
Crisper fast motion and faces?
Spicy Index
73.6−6.0
Freedom Score
100.0+1.7
Price
$0.625+25%/ 5 s clip
Compare side by side Try
Alibaba
Wan 2.2 LoRA
A different base model for your LoRAs?
Spicy Index
65.3−14.3
Freedom Score
88.1−10.2
Price
$0.24−52%/ 5 s clip
Compare side by side Try
Alibaba
Wan 2.2 Spicy LoRA
An uncensored build of that same Wan base?
Spicy Index
68.0−11.6
Freedom Score
74.8−23.5
Price
$0.24−52%/ 5 s clip
Compare side by side Try

What do you think?

Was this review useful?

Comments

Comments are checked automatically and by our team. We remove spam, offensive content, personal attacks and anything that names a real person without consent. Report a comment and we review it within one business day.

Tried this model? Share what worked. You need a free account to comment, so we can keep spam out.

MiniMax H3 LoRA FAQ

Try MiniMax H3 LoRA yourself

Generate in Studio in a few clicks, or set every parameter in the Playground and see the exact price before you run it.

$0.50 per 5-second clip

More reviews

See the leaderboards →

Wan 3.0
Pick Wan 3.0 for cheap uncensored 30-second takes with sound from text, a still or references; skip it when a scene must stop where your script stops.
Spicy Index
87.7
Freedom Score
96.0
Tested Sep 24, 2026
Read the review
Seedance 2.5
Pick Seedance 2.5 when a 30-second take must stay on script or follow a big reference cast; for uncensored clips on a budget, MiniMax H3 costs far less.
Spicy Index
81.4
Freedom Score
83.6
Tested Sep 24, 2026
Read the review
Wan 3.0 Prime
Pick Wan 3.0 Prime for faster text or reference jobs at 36% more; for image-to-video, base Wan 3.0 costs less and clears the same intimate-scene level.
Spicy Index
75.8
Freedom Score
80.7
Tested Sep 24, 2026
Read the review

What MiniMax H3 LoRA made in our tests

The verdict

Key facts

What we liked and what held it back

What we liked

What held it back

What the clips look like

Prompts that worked

How far it goes

LoRA compatibility

What the community says

Alternatives

How far MiniMax H3 LoRA goes

How MiniMax H3 LoRA earns its Spicy Index

What MiniMax H3 LoRA can do on SpicyAPI

How long it really takes

What people say about MiniMax H3 LoRA

Models to consider instead of MiniMax H3 LoRA

What do you think?

Comments

MiniMax H3 LoRA FAQ

Is MiniMax H3 LoRA uncensored?+

Can I run a civitai or Hugging Face LoRA through the API?+

Does loading a LoRA cost extra?+

Why did my LoRA seem to do nothing?+

How is this different from MiniMax H3?+

How is this different from MiniMax H3 Singularity LoRA?+

How much does MiniMax H3 LoRA cost?+

Try MiniMax H3 LoRA yourself

More reviews