Four MiniMax H3 entries sit in the SpicyAPI catalog (H3, H3 Spicy, H3 LoRA and H3 Singularity LoRA), all labelled Uncensored. In our September 2026 tests none of them refused a run, and none needs a LoRA for uncensored results.
Start with base MiniMax H3. It is the lowest-priced of these four for text, stills or image references, and it generates sound. Choose H3 Spicy when you have a folder of stills and want them moving with no prompt. Choose H3 LoRA for custom weights, seed control or video/audio references. Choose Singularity for fast action or distant faces when you also load LoRAs; image-to-video costs exactly 25% more than H3 LoRA at the same resolution, while reference-to-video rates are identical. Watch image-to-video in the L4 adult category, where two endpoints sometimes returned a softened clip.
The four versions side by side
"Spicy" has its tuning built in and no slot for yours; "LoRA" means you can load up to three LoRA files of your own.
| Version | Tasks | Load your own LoRA | Sound in the clip | Clip length | Seed |
|---|---|---|---|---|---|
| MiniMax H3 | Text-to-video, image-to-video, reference-to-video | No | Yes | 5-15 s (text), 4-15 s (image, references) | No |
| MiniMax H3 Spicy | Image-to-video only; the prompt is optional | No | No sound option (a clip may come back silent) | 3-15 s | Yes |
| MiniMax H3 LoRA | Text-to-video, image-to-video, reference-to-video | Yes, up to three | Yes | 3-15 s | Yes |
| MiniMax H3 Singularity LoRA | Image-to-video, reference-to-video (no text-to-video) | Yes, up to three | Yes | 3-15 s | Yes |
Input-role illustration. Base H3 and H3 LoRA offer all three routes; hosted Singularity offers image and reference routes only, and Spicy takes an opening image.
H3 Spicy is not "H3 with a LoRA slot", and Singularity is not another name for Spicy: it is a separate community fine-tune of H3 with its own tasks.
What each version costs
Prices for a five-second clip, in US dollars. Trust the console estimate for your settings.
| 5-second clip | 480p | 768p | Highest tier |
|---|---|---|---|
| H3, any task | Not offered | $0.185 | $0.29 at 2k (native) |
| H3 Spicy, image-to-video | $0.19 | $0.42 | $0.76 at 1080p (enlarged from 768p) |
| H3 LoRA, text or image | $0.25 | $0.50 | $1.00 at 1080p |
| H3 LoRA, references | $0.30 | $0.675 | $1.35 at 1080p |
| Singularity, image | $0.3125 | $0.625 | $1.25 at 1080p |
| Singularity, references | $0.30 | $0.675 | $1.35 at 1080p |
Reference-to-video adds a per-reference charge ($0.0068 per image on base H3; $0.02 per image or audio clip on the LoRA families), and on the LoRA and Singularity reference tasks any reference video you send is billed like output seconds. Loading a LoRA costs nothing extra. On Spicy, 1080p is an enlarged 768p picture, use 768p unless your delivery requires the larger file dimensions.
Example: four 768p drafts and one 2k keeper on base H3 cost 4 x $0.185 + $0.29 = $1.03. The same plan on H3 LoRA (480p drafts, 1080p keeper) is 4 x $0.25 + $1.00 = $2.00. Base H3 is cheaper for that plan. H3 LoRA can still justify its premium without an adapter when you need seed control, a three-second request or video/audio references unavailable on base H3.
Content levels
L1 is suggestive imagery, L2 moderate NSFW imagery, L3 advanced NSFW imagery, L4 higher-intensity adult scenes between two adults. L5a is an additional adult category; L5b is a separate cinematic-violence probe, not a stronger adult category. A clip is softened when it came back billed but toned down or implied.
No version refused a run, and none needed a content-unlocking LoRA:
| Version and task | Result without a LoRA |
|---|---|
| H3: text, image, image reference | Rendered at every level run |
| H3 Spicy: image | September 25 supplement: two of three L4 clips matched; one was softened |
| H3 LoRA: text, reference | Rendered at every level |
| H3 LoRA: image | L4 softened in two of three clips; other levels rendered |
| Singularity: image, reference | Rendered at every level |
The softened clips all started from a still at L4. If L4 from a still is your core use case, test your own still on the endpoint you plan to use, or switch to text-to-video or reference-to-video. With an adult LoRA loaded, H3 LoRA and Singularity also rendered L4 on every task they offer. Details: H3, H3 Spicy, H3 LoRA and Singularity.
Try it in the Playground
- Open MiniMax H3 text-to-video and stay on the Playground tab, in Form view. Sign in with a funded account to run.
- Write the shot in Prompt, or start from this one: Run this prompt on H3 →
- Set Resolution to 768p and Duration to 5 for a cheap first look; the estimate shows before you submit.
- Run the model. To compare versions, open H3 Spicy (add your still under Input image; the prompt is optional) or H3 LoRA and run the same idea there.
Which version should you use?
You have a written scene and nothing else. Use base H3 text-to-video: Singularity and Spicy cannot start from text, and H3 LoRA costs more. Draft at 768p, then re-run the take you want at 2k. Base H3 has no seed, so a re-run is a new take.
You have a folder of stills and want each one moving. Use H3 Spicy if you would rather not write prompts: the still alone is a complete request, clips start at 3 seconds ($0.114 at 480p) and the published test timings were around a minute or two, rather than an SLA. Use base H3 image-to-video if you will write a line of motion anyway: it costs less, goes to 2k and generates sound. H3 Spicy has no sound option and a clip can come back silent; add sound in the edit if you need it. Either way, describe what moves, not what is already in the frame; Spicy can leave a still behind when asked for a scene the frame does not show.
You need a character, a wardrobe and a location to hold across shots. Use reference-to-video. Base H3 takes up to nine reference images and nothing else; to copy motion from a clip or sound from an audio file, use the LoRA or Singularity reference task. Name each reference's job in the prompt ("Picture 1 is the woman, Picture 2 is the room"), and base H3 can ignore a separate garment reference, so put the outfit on the person image itself.
You already have LoRA weights. Use H3 LoRA. Each entry is a direct HTTPS link to a .safetensors file plus a strength from 0 to 4; the file must be trained for MiniMax H3 and compatible with the selected task. A public file link alone does not establish adapter compatibility. A concept LoRA at full strength with a prompt describing a different scene can replace your opening frame: write the prompt as what happens next in the same frame and start around 0.6–0.8. Style LoRAs show clearly on text-to-video; with a photograph as input, the photo sets the look.
Fast action or a distant face, plus LoRA loading. Choose Singularity. Its creator's card describes reduced motion blur and better distant faces, and the review's sword-spin and breakdance-freeze examples stayed readable mid-motion.
Calmer shots. Prefer H3 LoRA: Singularity image-to-video costs 25% more (references cost the same). In the published tests, a five-second 480p clip took about 80 s on Singularity versus 25–50 s on H3 LoRA.
A two-person, non-explicit scene of the kind a brief might describe.
Do you need a LoRA for NSFW?
No content-unlocking adapter was needed in the recorded trials. Base H3 and Spicy expose no LoRA slot; the other two delivered adult requests with empty adapter arrays, with the L4 image exceptions shown above. A LoRA adapts a look, style or character; it is not a content switch. The model page states that weights are not content-reviewed; this is separate from file compatibility and permitted use.
Reference-to-video does not always mean reference video
Base H3's reference-to-video takes up to nine images and no reference video or audio. The LoRA and Singularity reference tasks take up to nine images, three video clips and three audio clips. Reference video is capped at 15 seconds total and billed like output seconds on top of your output length; image and audio references add per-asset charges. Only feeding in an existing clip as a motion or sound reference needs a LoRA-family endpoint.
What about H3 Max and H3 Image LoRA?
Both are separate branches without the four versions' label. H3 Max is fal's separate post-trained build of MiniMax H3; its text-to-video is labelled Filtered and its image-to-video Softened, and it refused or toned down several higher-level prompts, so it is not the choice here (see the H3 Max review). H3 Image LoRA makes stills; the stills versus video guide covers when to use it.
The short version
Base H3 is the default: cheapest, with sound, and it handles all three input types. Spicy is for stills you want moving without writing a prompt, LoRA is for your own weights or video and audio references, and Singularity is for fast action and distant faces. All four share the Uncensored label and none refused a run, but the label does not make them interchangeable, and it does not exempt you from the law or from consent rules where you publish.
Where to go from here
- Compare field by field: H3 vs H3 Spicy, H3 vs H3 LoRA and H3 LoRA vs Singularity.
- See how far base H3 goes level by level in Is MiniMax H3 uncensored?, copy working prompts from the MiniMax H3 prompt guide, or set up weights with the H3 LoRA API guide.
