Minimax H3 LLM Vision and Reasoning. NSFW/SFW (REF2VA)

This is Minimax H3 workflow, that has LLM agent to write a prompt for you. It automatically takes context from your reference images (vision), refines your request and formats it for Minimax H3, which makes prompting much easier.

Main features:
- User experience-oriented design
- LLM-based prompting (using Ollama). Write a simple prompt and let LLM do the tedious work for you
- Vision capabilities. LLM can see your references and reason through their contents, no need to describe your references by hand
- Upscale and interpolate

- NSFW support. LLM understands NSFW concepts and can prompt almost anything


Other features:
- Advanced reference resolution control. Choose size of reference images/videos more precisely to balance between quality and generation speed
- Trim reference Audio/Video inside workflow
- LoRa support
- Mid-generation video viewer. Skip failed generations before they are completed


Video example with Loona in office contains very short and generalized prompt to show the power of LLM vision (imbedded workflow uses SFW system prompt)


Model Preview Override tiny_vae model (taeh3_madebyollin.safetensors)
Download it here and put into models/vae_approx folder:
https://github.com/madebyollin/taehv/blob/62f7591f59dfbb4c3c02b7a621d180a9eeaba26c/safetensors/taeh3.safetensors


You should be able to run this workflow with any LLM that can analyze images and runs with Ollama, but I can't guarantee that smaller ones will be able to follow instructions well enough. If you already have LLM running with Ollama, then skip to 7-th clause of LLM setup below, and adjust options according to your model specs.



LLM setup (Ollama, Qwen3.8-27B Heretic GGUF):

1. Create directory for models (for example Qwen3.8-27B-Heretic-RVN)

2. Download LLM with vision inside your directory
Go to this link and search for quants, that fit your setup (push button below, to show all files)
RVN-Q[YOUR_QUANT]_K_M-multilingual-vision.gguf
https://huggingface.co/0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF/tree/main

If you have 32Gb VRAM, you can use this Q5: RVN-Q5_K_M-multilingual-vision.gguf
If you have less, then consider overhead of a few Gb (don't pick those which barely fit on your VRAM).

LLM context will require additional VRAM (higher->more memory for thinking)

• At 32768 context, the buffer takes roughly 2 to 3 GB of VRAM.

• At 65536 context, it takes 4 to 6 GB.

• At 131072 (128k) context, it takes 8 to 12 GB.

Workflow's default is 49152.
You can read more about overhead on the main page of this repository.

3. Download vision adapter from the same repo (same file for all setups) inside your directory
mmproj-Qwen3.8-27B-Q8_0.gguf

4. Install Ollama
https://ollama.com/download/windows

5. Create a new file Modelfile (no extension) inside your model directory with two files, open it with your text editor and write:

FROM ./RVN-Q[YOUR_QUANT]_K_M-multilingual-vision.gguf

ADAPTER ./mmproj-Qwen3.8-27B-Q8_0.gguf

Replace [YOUR_QUANT] to quant of model that you downloaded and make sure that filenames are exactly the same.

6. Add LLM to Ollama.
Click on the address bar at the top of the Windows File Explorer window.

Delete whatever text is there, type cmd, and press Enter.

Type command (you can change "Qwen3.8-27B-Heretic-RVN" to other name, this name we will choose from a node):
ollama create Qwen3.8-27B-Heretic-RVN -f Modelfile

and then press enter. Wait for completion.

7. Make sure your ollama is opened and active (icon should appear in system tray)

Open workflow and go into this subgraph

Select your newly created model at the right node (if it's not there then try restart comfyui)
You can change context value at the left depending on your VRAM