Uncensored MiniMax H3 Max: Prompt Guide, Use Cases, GGUF Quants and LoRAs

H3 and H3-Max at a glance

Developer

MiniMax (Nanonoble Pte. Ltd.)

Released

July 31, 2026; weights August 3, 2026

Parameters

33B dense Omni Transformer

Clip length

4–15 seconds, 24 FPS

Audio

Native 32 kHz stereo

H3 resolutions

768P, 2K via regeneration

H3-Max resolutions

480P and 768P

Open source?

H3 base weights yes; H3-Max no

Inputs

Text, up to 9 images, 3 videos, 3 audio

Uncensored?

No; community builds only

What is H3-Max?

MiniMax, operated by Nanonoble Pte. Ltd., released H3 on July 31, 2026 and opened its base weights on August 3. H3 is the newest model in the company's video line: a 33B-parameter dense, single-stream Omni Transformer that reads text, images, video and audio and outputs 4 to 15 second clips at a default 768px short side, 24 FPS, with native 32 kHz stereo sound, in 11 stable languages. Two checkpoints are public: H3-Base-FL2VA for first, last or first-and-last frames, and H3-Base-Ref2VA for up to 9 images, 3 videos and 3 audio files (12 files at most). H3-Max is a separate tier post-trained together with fal.ai that trades the 2K path for speed, rendering at 480P or 768P only.

Why people look for an uncensored H3-Max

People search for an uncensored H3 because the model card says suggestive and pornographic content may be intercepted, and MiniMax's acceptable use policy forbids bypassing or weakening its guardrails. Neither H3 nor H3-Max has an uncensored mode, and the API applies the same rules. Unofficial NSFW fine-tunes exist on Hugging Face, but they are not MiniMax releases. Spicy AI does not host H3-Max; its image-to-video tool runs H3 Turbo without an extra moderation layer, for consenting adults only, and that is the honest alternative this page points to.

H3 vs H3-Max: which tier to choose

Pick by output resolution and by whether you need the weights, not by name. H3 is the quality tier: it renders 768P directly, reaches 2K through a second regeneration pass, and its base weights are open, so you can test it on your own GPU before committing. H3-Max is the throughput tier that fal.ai helped post-train: it is not open source, runs only through the API, tops out at 768P and adds a lighter 480P option for previews, social drafts and high-volume batches. MiniMax has not published a side-by-side quality comparison and we have not found an independent one, so run the same prompt on both before choosing.

What is open source and what stays API-only

"Open-weight" does not mean the whole pipeline is on Hugging Face. MiniMax published the H3-Base checkpoints with their processor, tokenizer, text encoder, visual and audio VAEs, diffusers format and inference code. Two parts stay closed: H3-Context-IR, the multimodal instruction model that interprets and rewrites prompts, and H3-Regenerate-2K, which re-renders a 768p clip at 2K using the original context. In practice a local machine produces stable 768p audio-video, and a true 2K result is a hybrid of a local or API base pass plus the API-only regenerator. H3-Max is closed entirely.

• Open: H3-Base-FL2VA and H3-Base-Ref2VA weights, VAEs, Qwen3-VL-32B text encoder, inference code.
• Closed: H3-Context-IR prompt rewriting, H3-Regenerate-2K upscaling, and every H3-Max checkpoint.
• Repository size: about 499 GB, because both checkpoints are included.

What people make with H3

The clearest signal of how people use the model is which add-ons they download. The most popular community LoRAs target speed, character replacement, orbiting camera moves, realistic skin and lighting, and physical motion, which maps onto a handful of recurring projects. Native stereo audio is the other big draw: because sound is generated with the picture, short talking, singing and ambient clips come out without a separate audio step. The list below groups the creative trends we see across model cards and LoRA pages; it describes what tools exist for, not a ranking of results.

• Photo-to-clip portraits: a single still becomes a 5–10 second shot with a smile, a turn or a hair flip, plus matching room tone.
• Character swaps: the Character Swap LoRA replaces the subject of a reference clip with a new face or outfit while keeping the motion.
• 360 product and fashion spins: the 360 Orbit LoRA circles a person or object, a popular format for outfits, figurines and sneakers.
• Multi-reference storyboards: Ref2VA accepts up to 9 images, so creators feed a character sheet, a location and a prop and get one consistent scene.
• First-and-last-frame transitions: FL2VA morphs between two stills, used for before/after reveals and outfit changes.
• Music and voice clips: lip-synced singing or spoken lines in one of 11 supported languages, with stereo ambience.

How the community rates H3

Adoption numbers are the most reliable public measure of reception. By mid-October 2026 the official Hugging Face repository had about 6,215 likes and roughly 4.6 million downloads in 30 days, while the Comfy-Org ComfyUI package passed 25 million, which suggests most people run it inside ComfyUI rather than with raw scripts. 36Kr's launch report cited an arena Elo of 1130 and 16 ecosystem partners at release, with support in SGLang, vLLM, diffusers and ComfyUI. The trade-offs discussed most often follow directly from the model card: the 499 GB download, clips capped at 15 seconds, and 2K that needs the closed regenerator. We do not quote individual reviews here; test it on your own prompts.

Hugging Face figures are 30-day platform counts and change daily; they measure adoption, not quality.

How to write prompts for H3 video

The model reads prompts like a short shot description, and MiniMax's GitHub repository even ships a prompt skill for structuring them. The most reliable pattern is one shot, one main action, written in plain sentences in this order: subject, action, setting, camera, light, sound. Keep it to two to four sentences for a 5-second clip; long lists of adjectives dilute the motion. When you upload a photo, describe only what should change, because the image already defines the person and the scene. Name the sound explicitly, since audio is generated with the picture.

• Subject first: "A woman in her thirties in a satin robe" beats "beautiful, stunning, 8k".
• One verb per shot: turns, laughs, walks toward the camera. Two big actions in 5 seconds usually blur together.
• Camera in film terms: slow push-in, handheld, orbit left, locked-off wide shot.
• Light with a source: candlelight from the left, neon through blinds, golden hour backlight.
• Sound as a line of its own: soft rain and distant traffic; she whispers "come here".
• Match length to action: 4–6 seconds for a gesture, 10–15 seconds only for a slow, continuous move.

H3 prompt templates you can copy

Replace the bracketed parts and keep the structure. Each template is written for image-to-video, so it assumes your photo already shows the adult subject; for text-to-video, add a one-line description of the person at the start. All of them also work as a starting point in the H3 Turbo generator above.

Step by step: animate a photo with H3 Turbo on Spicy AI

The browser workflow skips the 499 GB download and any GPU setup. It uses H3 Turbo, a faster H3 variant, not the H3-Max API, so expect a different look from official H3-Max output.

• 1. Open the image-to-video tool from the button at the top of this page; the H3 Turbo model is preselected.
• 2. Upload a clear photo of an adult subject that you own or have consent to use.
• 3. Paste one of the templates above and fill in action, camera, light and sound.
• 4. Generate a short clip first, check the motion, then adjust a single element per retry.
• 5. Download the result; if the action looks rushed, make it simpler rather than longer.

GGUF, INT8 and NVFP4 builds for local use

Of the 33B parameters, about 13B sit in AdaLN-related branches that can be precomputed and not loaded at inference, and the text encoder is the full Qwen3-VL-32B, so community packaging is what makes local runs practical. Download counts below are 30-day Hugging Face figures from October 2026; they show adoption, not quality. MiniMax publishes no official VRAM table for these quants, so read each model card and re-check video and audio quality after quantizing.

• Comfy-Org/MiniMax-H3: ComfyUI packaging, about 25.24 million downloads; the easiest baseline to compare against.
• unsloth/MiniMax-H3-GGUF: GGUF quants for lower VRAM and CPU offload, about 1.93 million downloads.
• Other GGUF routes (Abiray Pruned-GGUF, realrebelai, leejet, molbal) prune or quantize differently; test quality and speed yourself.
• INT8 and NVFP4 builds (Abiray, MATLOWAI, coolthor): NVFP4 targets RTX 50-series cards; older GPUs are better served by GGUF or INT8.

Turbo LoRA, Character Swap and 360 Orbit LoRAs

Speed is the most common reason to add a LoRA to the base model. lightx2v/Minimax-h3-Turbo is a few-step distilled build with about 1.5 million downloads, and larryvrh/MiniMax-H3-Turbo-Lora packs the same idea into a LoRA, with a ComfyUI-ready version by drbaph. After cutting steps you usually need to retune CFG and step count. For specific looks, akatz-ai's Character Swap LoRA replaces a subject using a reference image and mask, pablodawson's 360 Orbit LoRA produces a circling camera move, and fal published a Realism People LoRA for skin and lighting texture. Orbit LoRAs can fight camera directions written in the prompt, so use one or the other.

Uncensored MiniMax H3: community NSFW builds

Anyone looking for an uncensored MiniMax H3 will find several unofficial adult fine-tunes on Hugging Face, such as PinkCherry-NSFW_MiniMax-H3, AfterMidnight-MiniMax-H3-NSFW and a Sakura uncensored GGUF with about 13,000 downloads. None of them is a MiniMax release, and we found no public statement that MiniMax endorsed them; its usage policy prohibits circumventing safety guardrails. We do not link them or explain how to strip filters. If your goal is adult video from your own photos, a hosted tool is simpler than downloading and running a 499 GB model, and illegal content, minors and real people without consent remain off limits everywhere.

H3 vs Seedance 2.5 and Wan 3.0

H3 is the only one of the three with open weights, which matters if you need local runs, LoRAs or fine-tuning. Seedance 2.5 and Wan 3.0 are closed APIs that win on clip length and resolution. Pick H3 for local or customizable work; pick a hosted rival for long shots, many references or a fully managed service.

Seedance 2.5 and Wan 3.0 figures come from the approved research brief and their providers' public documentation.

Prompt tips for H3-Max-style results

1. Write one shot with one main action; split bigger ideas into several clips.
2. Need 2K? Only H3 offers it, and it requires the API regeneration step.
3. Start local tests with the Comfy-Org build before trying GGUF or INT8 quants.
4. Retune CFG and step count whenever you add a Turbo LoRA.
5. Describe every character as an adult (18+) and never upload photos of real people without consent.

Sources

• H3 model card (Hugging Face)
• H3 inference code and prompt skill (GitHub)
• MiniMax video generation guide
• Launch and open-source report (36Kr)
• Unsloth H3 GGUF (third party)
• lightx2v H3 Turbo (third party)

More uncensored AI Video models

• uncensored minimax h3
• Unlimited Seedance 2.5
• Uncensored Kling
• Uncensored Kling 2.6
• Uncensored Kling O1
• Uncensored Sora
• Uncensored Veo
• Uncensored Hailuo AI
• Uncensored Wan
• Uncensored Wan 3.0
• Uncensored Seedance
• Uncensored Seedance 2.5
• Uncensored Seedance 2.0
• Uncensored Seedance 2.0 Mini
• Uncensored HunyuanVideo
• Uncensored LTX Video
• Uncensored Runway
• Uncensored Pika
• Uncensored Luma Dream Machine
• Uncensored Vidu
• Uncensored PixVerse
• Uncensored EdenLive
• Uncensored Grok Imagine