exportanything/comfyui-diffusiongemmapromptbuilder

This project is a set of five custom nodes for ComfyUI, a visual, node-based tool for building AI image and video generation workflows. Rather than generating images or video itself, it uses an AI model called DiffusionGemma purely as a smart prompt writer, taking in text, an image, and a video as context, and producing a carefully constructed prompt for a separate video generation model called LTX to actually use. The five nodes work together in a chain: one loads the DiffusionGemma model itself, one combines your text description with an optional reference image and video into a single context object, one stores your target settings like video profile, duration, and aspect ratio, one asks DiffusionGemma to generate the final structured prompt based on all of that, and one splits that result apart into the individual pieces an LTX video workflow needs, such as the positive prompt, negative prompt, and resolution. The project ships a full example workflow for character motion transfer, where a reference image defines who appears in the video and a separate source video defines the motion, pose, and camera movement to apply. Using it requires downloading several large model files separately from Hugging Face, including the DiffusionGemma prompt model, the LTX diffusion model, supporting LoRA files, video and audio decoders, and a text encoder, each placed into a specific ComfyUI models folder. Setting this up requires a ComfyUI installation with a CUDA-capable GPU and several other community node packs installed through ComfyUI's manager. The project includes a verification script to confirm the specialized compressed model format it relies on is working correctly before attempting a real generation. It is released under the MIT license.