AI Influencer Master Class - 🌶️SPICY🌶️ H3 REF2V Xtra HD (Patreon)

Want your character to look, move and sound exactly like your references in explicit scenes, without writing a single H3 prompt by hand? Drop in a photo and a voice clip, type one line, and this workflow writes the full MiniMax H3 prompt for you.

It's the next step in the AI Influencer Masterclass: the same character from your images, now talking in the voice you give it, with the NSFW LoRAs already loaded.

What's new vs the FL2VA version

The FL2VA workflow animated between a first and last frame. This one runs in reference mode: you give H3 pictures, clips and voice clips, and it pulls the character, look, motion and voice from them. The prompt now writes itself.

Media Loader: drop pictures, videos and audio onto one node. Everything is prewired, so load only what you need and empty slots are skipped.

Auto prompting: the Universal Writer looks at and listens to your references, then writes the full MiniMax H3 prompt for you.

Two pass Xtra HD: a low res pass, a latent upscale, then a high res pass.

Optional Sharpen Pipeline: off by default, for pushing a finished video further.

NSFW version differences

Same workflow as the SFW version, with two extra LoRAs in the Power Lora Loader, both at 0.75:

HMNSFW AIO V2

Worship It v0.6

Links are in the Model Links note. Toggle either one off in the Power Lora Loader if a scene doesn't need it.

If the writer tones down or refuses a scene (for example putting clothes on a nude reference), switch writer_model to Qwen3.5-9B Uncensored. Check the captions output first: if the captions describe the image correctly but the prompt changes it, it's the writer.

Before you start

Update ComfyUI. The hybrid model and the nvfp4 text encoder need a recent build with native MiniMax H3 support.

Install missing nodes through ComfyUI Manager (Install Missing Custom Nodes), then restart.

Download the models from the Model Links note inside the workflow. The auto prompt models download themselves on first run.

Quick start

Drop your files onto the Fantastic H3 Media Loader.

Set the video shape and size on LOW RES PASS and HIGH RES PASS (aspect ratio and megapixels), and the length in seconds on VID LENGTH. VID LENGTH also sets the writer's duration for you.

On the Universal Writer, check the task is Ref2VA, set its aspect picker to match the passes, and type a short idea in the prompt box.

Check the badge on each reference square (see below).

Queue. The writer builds the prompt and sends it to both passes.

The first run takes longer while the captioner, writer and MiniMax prompt guide download.

The reference strip (most important part)

Every reference gets a square on the Universal Writer. Its badge decides what the captioner looks at:

pic: the image is a shot or frame. It describes framing, setting and lighting, not the person.

subj: the image is a person or thing to keep recognisable. It describes hair, build and clothing. Use this for your character.

vid: a clip, or a batch of frames read as one.

aud: a voice, music or sound.

Click a badge to change it. Drag squares to reorder. Click the number to switch a reference off.

If the prompt describes your character wrong (wrong hair, wrong outfit), the image is almost always badged pic when it should be subj.

Custom instructions (instr)

Click the instr band under a square to change what gets asked about that reference.

Unticked: your line is added to the normal question.

Ticked "Ask this instead of the role's own question": your line replaces it.

Two are preloaded for you: the first picture is told to lock the frame and person in place, and the first audio slot is told to transcribe the words.

Getting the exact words from your audio

By default the writer only describes a voice (tone, pitch, speed). Its built in audio question literally tells the model not to transcribe. Adding a line on top does nothing, because the model follows the original instruction.

The fix is already set on the first audio slot. To set it on another one:

Click the instr band under the aud square.

Tick Ask this instead of the role's own question.

Paste: Transcribe exactly what is said, word for word, in quotes. Then describe the speaker's voice timbre, apparent age and gender, delivery, speaking rate and accent, and the background sound.

Keep caption_length on detailed so the transcript isn't cut short.

Still missing from the final prompt? Wire the writer's captions output to a text preview. If the words are there but not in the prompt, add She says exactly the words from the reference audio. to your idea.

The two models

The writer splits the work between two models:

caption_model looks at and listens to your references. Only the Qwen2.5-Omni models can hear audio. The workflow uses 7B, which describes better than the 3B.

writer_model turns those descriptions into the H3 prompt. It never hears the audio, it only reads the captioner's notes. The workflow uses Qwen3.8-27B Q4_K_M (about 19 GB of VRAM).

Smaller card? Pick by VRAM from the dropdown: Qwen3.5-4B for 8 GB, the 9B options for around 8 to 12 GB, Mistral Small 24B for 16 GB. If scenes get refused or toned down, switch to Qwen3.5-9B Uncensored.

Keep keep_model_loaded off so both models leave your VRAM before H3 starts.

The system prompt

The writer uses MiniMax's official prompt writing guide as its system prompt. Open guide folder on the writer shows it.

Typing in the system_prompt box replaces the guide completely. Leave it empty to use the official one. To add your own rules on top, edit the bottom of VIDEO_PROMPT_WRITING_GUIDE_ref_en.md in that folder. It never gets overwritten once it exists.

Self check warnings

After writing, the node checks its own prompt:

"Picture connected but never cited": the image reaches H3 but the prompt never says what it's for. Add to your idea: Define <Subject 1> as the woman in <Picture 1>.

"detailed_description is X words, the guide suggests 350 to 500": your idea was too thin. Add 2 or 3 actions, a camera move and an ending, or ask for it directly: Write the detailed_description in 350 to 500 words across 2 or 3 shots.

Raising n_ctx does not make prompts longer. It's only how much text the model can hold, and the node already raises it on its own for Ref2VA.

Limits and fixes

The writer takes 12 references max, so it's wired for 6 pictures, 3 videos and 3 audio clips.

Error about llama.dll: on the Rewriter Options node, set gguf_runtime to llama.cpp. It's already set in this workflow.

Changing the turbo LoRA: change the steps to match, and set the Sigma Shift values the new LoRA's author recommends.

Optional: Sharpen Pipeline

Off by default. It reloads your finished video and regenerates it sharper with the same subject, motion and audio. Run the base workflow once first, set FINAL OUTPUT MEGAPIXELS, then select the group and press Ctrl+B. The note inside the workflow covers the details.