AI Influencer Master Class - MiniMax H3 Spicy Version (Patreon)

The free version of this workflow is the highest quality MiniMax H3 setup I've built. This is that same pipeline tuned for spicy content, and it writes your prompts for you. Drop in a photo, hit run, and you get a full quality H3 video on turbo speed with spicy LoRAs stacked on top.

This one is exclusive to AI Influencer Masterclass members. Below is a full walkthrough of every group so you know exactly what to touch and what to leave alone.

📺 Video walkthrough of the base workflow: https://youtu.be/Y602t6RK8Q0

📥 Setup

Download the workflow JSON attached to this post and drag it into ComfyUI.

Open the Model Links note on the canvas. Every model and LoRA has a link, and the folder tree at the bottom shows where each file goes.

After downloading the LoRAs, open the Power Lora Loader and reselect each one from the dropdown so the names match your files. Some of them download with a different filename than the one in my workflow.

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

Install any missing custom nodes through the Manager.

🎛️ USER INPUTS / CONTROL PANEL

Everything you touch on a normal run lives here.

FIRST FRAME is your starting photo. The auto prompter reads this image too, so pick a photo with a clear pose and good lighting. The better the photo, the better the prompt it writes.

LAST FRAME is bypassed by default. Turn it on only if you want the video to end on a specific image.

POSITIVE PROMPT is filled in automatically by the auto prompter. You don't type here unless you switch auto prompting off (covered below).

VID LENGTH is in seconds. The math node underneath converts it to a frame count H3 accepts, so you never deal with frame counts yourself.

LOW RES PASS and HIGH RES PASS set the size of your video. Low Res is where the first generation happens and it's small on purpose so it runs fast. High Res is the size the video gets upscaled to. Set the aspect ratio on both to match your photo, and make sure both nodes use the same aspect ratio or the upscale will stretch your video.

FINAL OUTPUT MEGAPIXELS only matters if you turn on the sharpen pipeline. On a normal run it does nothing.

Seed controls the video generation. Keep it fixed while you're dialing in a shot, randomize it when you want variety.

🤖 AUTO NSFW PROMPTER

This is the big addition over the free version.

A small vision model (Huihui Qwen3 VL 4B) looks at your FIRST FRAME and writes a complete MiniMax H3 prompt in the official format. It reads the pose, where her hands are, and how close things are, then writes a scene that starts from that exact position and builds naturally from there. The trigger words get added to the front automatically and the whole thing goes into POSITIVE PROMPT.

Checking the prompt: after a run, open the Preview Any node to read what it wrote.

Getting a different prompt from the same photo: change the seed on the Prompt Enhancer node and run again. This seed is separate from the video seed in the control panel.

Writing your own prompt instead:

Select every node in this group and press Ctrl+B to bypass it. This also stops the 4B model from loading.

On the POSITIVE PROMPT node, drag the incoming link off its input so the box becomes editable.

Type your prompt and start it with the trigger words from the String Concat node (hmmotion, mvmt, fixed camera,) so the LoRAs still activate.

Changing what kind of scenes it writes: the large text box on the Prompt Enhancer node is the system prompt, the full set of rules the model follows.

Leave the format rules alone. The opening line and the three fields (integrated_multimodal_description, overall_soundscape, non_diegetic_music) are the official H3 structure. Remove them and H3 gets a prompt it can't read properly.

Edit the scene rules to change the acts it describes, the pacing, camera framing, sound, and anything you want it to avoid.

Edit the last paragraph to set the type of scene you want it to go for.

Keep your rules short and direct. Small models follow clear instructions better than long ones.

Paste the original text somewhere safe before editing so you can always go back.

🧩 Models

Standard loaders for the hybrid model, the main text encoder, and the video and audio VAEs. The hybrid model is a merge of the official first frame and reference checkpoints, which is what lets the sharpen pipeline reuse the same model later. Nothing to change here once your files are selected.

💊 LoRAs + Shift + Attention

Turbo LoRA is what lets this run on so few steps. You can swap it for another turbo LoRA, but look up the video and audio shift values that LoRA's author recommends and set the Sigma Shift node to match. Swapping the LoRA without changing the shift gives worse results.

Power Lora Loader holds the four NSFW LoRAs:

HMNSFW AIO handles general NSFW motion and anatomy. Trigger word hmmotion.

Mystic XXX adds explicit detail and range.

Worship It covers body focused scenes.

Cumshot handles finishes.

They're set to strengths that play well together, so run them as is first. If a result looks overcooked, lower the strength on whichever LoRA matches what went wrong. You can also toggle any of them off with the switch next to its name if a scene doesn't need it.

Sigma Shift holds the shift values for the turbo LoRA. Only touch it if you swap the turbo LoRA.

Attention backend stays as is.

🎬 Sampling (first pass)

This is the first generation at your small Low Res size. It runs 6 steps even though the turbo LoRA is built for 4. You can drop it to 4 for a little extra speed, but 6 gives a noticeably cleaner result and costs very little at this size.

Use the live preview. The preview node shows the video forming while it samples. The upscale and sharpen passes make a good generation look better, they won't rescue a bad one. If the motion is off or the scene drifted, cancel, change the seed, and run again. Rerolling here is cheap. Rerolling after the upscale isn't.

🔍 H3 Latent Upscaler

H3 keeps video and audio together in one latent. This group splits them, upscales only the video half with the H3 latent upscaler model, then joins them back together.

The upscaler is a trained model, so it grows the video without decoding to pixels and re-encoding, which on H3 is slow. The workflow uses the fp32 version. If you have the bf16 version, pick it in the upscaler node and nothing else changes.

This saves time, not VRAM. The next pass still runs at your full High Res size.

✨ Two-stage sampling upscaling (second pass)

The second sampler resamples the upscaled video at High Res to clean it up and add detail.

There are three sigma nodes labeled 3 steps, 4 steps, and 5 steps. Only one connects to the sampler at a time and 4 is the default. To change it, drag the sigmas link off the sampler and plug in the one you want. Fewer steps is faster, more steps is cleaner.

After that it decodes the video and audio and the Video Combine saves your finished file.

Don't change the save settings on this Video Combine. Keep save output on and keep the format as mp4. The sharpen pipeline loads this exact file.

🔪 SHARPEN LORA + EXTRA SAMPLE PASS (optional)

Off by default. Use it when you want to push a finished video further.

It reloads the video you just saved and runs it back through the hybrid model as a reference to video pass with a sharpness LoRA and a ref2v turbo LoRA. Your finished video becomes the guide and the model regenerates it with crisper detail while keeping the motion and audio the same.

Run the base workflow at least once so there's a saved video to load.

Set FINAL OUTPUT MEGAPIXELS in the control panel. This sets the size of the sharpened result. Higher is sharper but costs more VRAM and time.

Select every node in the group and press Ctrl+B to turn it on.

Run again.

You get two outputs: the sharpened final video saved to your output folder, and a side by side before and after preview so you can judge whether it was worth it.

🛠️ Troubleshooting

The prompt ignores my photo: make sure FIRST FRAME has an image loaded and the Prompt Enhancer's image input is connected.

Video looks stretched after the upscale: your Low Res and High Res aspect ratios don't match each other, or don't match your photo.

Sharpen pipeline errors or loads the wrong video: the base Video Combine save settings were changed, or the base workflow hasn't been run yet.

LoRA not found errors: reselect each LoRA in the Power Lora Loader so the names match your downloaded files.

Out of memory on the second pass: lower the HIGH RES PASS megapixels. The latent upscaler doesn't reduce VRAM for this pass.

Questions or issues, drop them in the Masterclass channel on Discord. 🙂