MiniMax H3 is one of the most capable video models running inside the VirtuaVixen AI Studio, and we run it in two distinct flavors: MiniMax H3 Image to Video (I2V) and MiniMax H3 Reference to Video (R2V). Both are fully self-hosted on our own GPU fleet, both generate native synchronized audio alongside the picture, and neither one requires you to own a graphics card, install ComfyUI, or download a single model file. You pick a mode, drop in an image, type a prompt, and the studio does the rest at your usual per-tier token cost.

This guide explains what each mode actually does, how to use them, when to pick one over the other, and — for the technically curious — exactly which models, LoRAs, and ComfyUI nodes are firing under the hood.

What MiniMax H3 Is on Our Studio

MiniMax H3 is a video diffusion model that produces short, high-motion clips with sound baked in. That native-audio capability is the headline feature: instead of generating a silent clip and bolting audio on afterward, H3 decodes a separate audio track in the same pass as the video, so lip movement, ambience, and moans line up with the picture instead of drifting out of sync.

Because we host it on our own fleet, you get the uncensored, NSFW-tuned version of the pipeline with no rate limits from a third-party API and no local setup. Both modes output 3:4 portrait clips — the ideal shape for phone viewing and for our Shorts feed — and both are priced per tier in tokens, so heavier tiers pay less per clip.

MiniMax H3 Image to Video (I2V), Explained

Image to Video takes a single still image and animates it. You give the model one starting frame — a photo you shot, an AI image you generated in the studio, or a picture of one of our AI pornstars — plus a text prompt describing the motion you want, and H3 brings that frame to life with matching audio.

I2V is the mode to reach for when you already have the exact frame you love and simply want it to move. The composition, lighting, and subject in your source image are preserved as the literal first frame of the video, and the model animates outward from there. It is built for speed and for faithful motion on tight, close-up, and explicit framing.

How to Use Image to Video

1. Open the AI Studio and select the MiniMax H3 Image to Video workflow.
2. Upload or pick your starting image (any still — your own upload, a studio generation, or a pornstar frame).
3. Write a prompt describing the action and motion you want to see.
4. Generate. The clip renders in the 3:4 portrait format with synchronized audio, ready to download or post to Shorts.

MiniMax H3 Reference to Video (R2V), Explained

Reference to Video works differently. Instead of animating one frame, it takes two reference images of a subject plus a text prompt, and it generates a brand-new scene that keeps that subject looking consistent. Your references are not the first frame — they are identity anchors. The model reads who the person is from the two images, then builds an entirely new shot, pose, and setting from your prompt while keeping the face and body recognizably the same.

This is character consistency in the truest sense: you can put the same AI pornstar into a scenario she was never photographed in, and R2V will render her there. Output is a fixed clip of roughly ten seconds in the same 3:4 portrait format, again with native audio.

How to Use Reference to Video

1. Open the AI Studio and select the MiniMax H3 Reference to Video workflow.
2. Upload two reference images of the same subject — different angles or expressions help the model lock identity.
3. Write a prompt describing the completely new scene you want her placed into.
4. Generate. R2V builds the new ~10-second scene from scratch while preserving your subject, with synchronized audio.

I2V vs R2V: Which One Should You Pick?

The choice comes down to one question: do you want to animate a frame, or place a character in a new scene?

• Pick Image to Video when you already have the perfect still and want it to move exactly as composed. It preserves your source frame precisely and, thanks to a turbo distillation LoRA (more on that below), it is the faster of the two. Great for close-ups and explicit detail where the framing is already right.
• Pick Reference to Video when you want the same person in a different scene, pose, or location that no single image of them covers. You trade the “exact frame” control of I2V for the freedom to prompt an entirely new shot while keeping identity intact.

A common workflow is to combine them: generate or choose a strong still, use I2V to animate it, and use R2V whenever you need that same character in fresh scenarios for a longer set.

Under the Hood: The Exact MiniMax H3 Stack

Both modes run as ComfyUI graphs on our fleet. They share the same dual-VAE audio-plus-video architecture and the same sampling backbone, but they diverge sharply in how they handle the subject — and that difference is the whole reason we ship two separate workflows.

Image to Video — Models, LoRAs, and Nodes

The I2V graph loads its base model with a UNETLoader pointed at PinkCherry_fl2va_MiniMax_H3_int8_convrot-beta-0.6.safetensors — an int8 build of H3 chosen for a strong quality-to-speed ratio — with the text encoder brought in via CLIPLoader. On top of the base model it stacks three LoRAs, each applied through a LoraLoaderModelOnly node:

• minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors at strength 1.0 — the turbo 8-step distillation LoRA. This is what makes I2V fast: it collapses the sampling schedule down to roughly eight steps instead of the usual dozens, so clips render quickly without falling apart.
• moawxx_000002000.safetensors at strength 0.8 — a motion and style LoRA that shapes how the animation actually moves.
• SynthPussy_H3_closeups_v1-step00008300.safetensors at strength 0.8 — an NSFW anatomy LoRA tuned specifically for tight and close-up framing, keeping explicit detail coherent where generic models tend to smear.

The heart of the graph is the MiniMaxH3ImageToVideo node, which turns the starting frame plus prompt into latent video. Input sizing is handled by ImageScaleToTotalPixels and GetImageSize, and a ResolutionSelector set to 3:4 portrait at roughly 0.4 megapixels (kept to a multiple of 32) defines the output shape. Sampling uses KSamplerSelect with the res_multistep sampler feeding a SamplerCustomAdvanced, driven by BasicScheduler, BasicGuider, and RandomNoise. An EasyCache node speeds up repeated passes.

Decoding is where the audio comes from. The graph carries two VAEs: minimax_h3_video_vae_fp16.safetensors decodes the picture, while minimax_h3_audio_vae_fp32.safetensors decodes the soundtrack through a VAEDecodeAudio node — that dual-decode is exactly how H3 produces synchronized native audio in a single generation. Finally an HDR Effects pass (SuperBeasts.AI) grades the image, and VHS_VideoCombine muxes picture and audio into the finished file.

Reference to Video — Models and Nodes

R2V starts from a different base model entirely: UNETLoader loads minimax_h3_ref2va_pruned_bf16.safetensors, a bf16 reference-tuned build. The critical detail here is that R2V runs with no LoRAs at all. It does not need the turbo LoRA, the motion LoRA, or the anatomy LoRA, because the dedicated reference model already does the identity work itself — that is its entire job.

Instead of a single starting frame, R2V takes two reference images through two separate LoadImage nodes. Those two images and your text prompt are assembled by a MiniMaxH3PromptBuilder node, which packages identity plus intent into a single conditioning bundle. That bundle feeds the MiniMaxH3ReferenceToVideo node, which generates a completely new prompted scene while holding the subject from the references consistent. This is why R2V can invent a shot the person was never photographed in and still keep them recognizable.

Everything downstream mirrors I2V so the two modes stay visually consistent: the same ResolutionSelector at 3:4 portrait ~0.4 MP, the same KSamplerSelect with res_multistep into SamplerCustomAdvanced plus BasicScheduler and BasicGuider, and the same dual VAE — video fp16 for picture and audio fp32 through VAEDecodeAudio for sound. An HDR Effects (SuperBeasts.AI) pass grades the result, and the clip is assembled with CreateVideo and written out by SaveVideo at a fixed length of about ten seconds.

The Key Technical Contrast

The cleanest way to understand the two modes is by how each solves the subject problem. I2V animates an existing frame and leans on a turbo 8-step LoRA (plus motion and NSFW-anatomy LoRAs) to do it fast, using your image as the literal first frame. R2V invents a new frame and uses no LoRAs at all, relying instead on a purpose-built reference model and the MiniMaxH3PromptBuilder to carry identity across a scene the model has never seen. One preserves a shot; the other preserves a person.

Tips and Use Cases

• For I2V, start with a clean, well-lit source frame. Because your image becomes the literal first frame, its composition and lighting carry straight into the clip — the SynthPussy closeup LoRA rewards tight, explicit framing especially well.
• For R2V, give it two varied references. Two different angles or expressions of the same subject give the reference model more to lock onto, which tightens identity consistency in the new scene.
• Write motion, not just description. Both modes respond to verbs. Describe what is happening and how it moves, not only what the frame looks like.
• Use the portrait shape to your advantage. Both modes output 3:4 vertical clips built for phone viewing and drop straight into the Shorts feed.
• Mix modes for a series. Lock a character with R2V across several scenes, then use I2V to animate your favorite stills at speed.

Both MiniMax H3 workflows are live now in the VirtuaVixen AI Studio — native audio, real character consistency, and zero local setup. Pick your starting image, choose I2V for speed or R2V for a fresh scene, and generate.