These are subgraphs that you can add that will enhance your H3 Prompt using Ollama and the LLM of your choice. The use the same guidance that Minimax released but I added some stuff about cum texture words to use. I'm had best luck with Qwen uncensored 3.8 27B or 3.6 35B parameters. Extract the zip contents to your subgraphs folder which is a sibling folder to your workflows folder, so default\users\subgraph and then you can search for enhance h3 and you'll see them. configure the model you want to use.
Known Issues
It accepts only one image so i use image stitch to combine them then resize to 1MB and pass that to ollama with thick red borders between the images. You need to tell it that the model is getting the entire image in each reference and it is being shown to you stitched together with red borders so do not refer to them as part of that stitched image, but rather picture 1,2,3 starting from left to right. If doing t2v you need to bypass the image resize node inside the subgraph or it will error. If anyone can improve on this to accept multiple images without the stitch let me know.
They require the comfyui-ollama node, and I customized it more to allow the 'use internet' option, but author never merged it to his codebase when I submitted a PR, so you will have to live without that option. There is a memory management node from KJnodes since you often need to unload a model from your GPU so Ollama can use it.
Virtually always requires a little manual tweaking, but better than doing it all yourself. You can connect it to a CoachBate Preview and Edit Text node any port, bypass ollama subgraph and then you can edit it in that node . That CoachBate node is nice because it saves your generated dynamic prompts with the workflow so when you open it up again from embedded metadata in your mp4, you can see what prompt your video actually used, not just what your ollama prompt was. Search the ComfyUI manager for CoachBate and you'll find my node pack, but it isn't required to use this Ollama subgraph.
There are custom nodes that do a better job at this, but this is faster for sure.