Comfyui-Lance-Aio Slashes Vram To 12Gb For Bytedance Multimodal Magic

The Comfyui-Lance-Aio release brings ByteDance’s Lance-3B model into ComfyUI with custom nodes that handle seven multimodal tasks. It can generate images and video from text, edit both media types, and perform image and video understanding. The implementation faithfully ports the official code but adds dynamic VRAM streaming so users with limited GPU memory can still run the full model.

Developer SteveImmanuel created this node pack to fill a gap, noting that existing Lance support in ComfyUI was missing or required too much VRAM. He leveraged ComfyUI’s dynamic VRAM feature to load and offload model parts on the fly, dropping the memory requirement from 40GB to as low as 12GB. All seven tasks were tested and confirmed working on an Ubuntu machine with an RTX 5070 and 12GB VRAM.

One model, seven tasks, less VRAM

This tool is aimed at users with prosumer GPUs like the RTX 5070 who want to run a unified image and video model locally. Small agencies can benefit from having text-to-video, editing, and understanding tasks in one ComfyUI workflow without relying on cloud services. Privacy-conscious professionals also gain a capable on-premises multimodal engine that stays entirely offline.

Developer notes and limitations

SteveImmanuel notes that image-related tasks could theoretically run on 8GB VRAM, but he has not confirmed the minimum requirement due to lacking hardware. The codebase requires transformers version 4.50.3 to below 5 because the official Lance model is incompatible with newer libraries. He used AI to understand the original code structure but manually wrote and reviewed every line of the custom nodes.

"I have tested running all of them in my GPU with 12GB VRAM and confirmed all works well." — Source: Reddit