StreamingT2VLoaderEnhanceModel

StreamingT2VLoaderEnhanceModel loads the third stage of this pack's pipeline in isolation: the damo/Video-to-Video model from modelscope that takes your long 256×256 video and re-renders it at a larger, sharper resolution (default 1280×720). The msxl_model output feeds StreamingT2VRunEnhanceStep, which is where the actual enhancement happens.

Here's the context you need. StreamingT2V extends video at 256×256 because that's what the autoregressive model was trained at. It's fast and cheap, but 256px looks like a webcam. So the pack's third stage runs a separate video-to-video model over the whole long clip, upscaling and cleaning it in one pass. That model is Alibaba's modelscope video-to-video pipeline (damo/Video-to-Video), and this loader is the part that stands it up.

How it works

The loader calls the modelscope video-to-video pipeline with a fixed config: target size 1280×720, padding on, revision v1.1.0. It's downloaded automatically from modelscope on first run - no manual model download, no files to place in models/checkpoints. You pick a device (cuda or cpu, default cuda) and you're done.

The single output, msxl_model, is the enhancer object. Wire it into StreamingT2VRunEnhanceStep, along with the path to your low-res video, and that node handles the frame-by-frame upscale pass.

Inputs

There's exactly one input worth mentioning: device, defaulting to cuda. If you're short on VRAM you can try cpu, but this is a video transformer doing a full-length pass - CPU will be slow, and this stage is the last thing you run, so give it the GPU if you can.

Installing it

The standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/chaojie/ComfyUI_StreamingT2V

or ComfyUI Manager → search StreamingT2V → install, then restart. The pack's requirements.txt is a heavy one (diffusers, pytorch-lightning, modelscope, xformers, bitsandbytes, imageio[ffmpeg]...), and note that this stage specifically leans on the modelscope SDK - if your environment struggles with modelscope imports, this loader is where it'll show up. First run also downloads the enhance model, so expect a stall on the first Queue click.

The take

This is one of the two nodes (with StreamingT2VLoaderStreamModel) you'd load on their own if you're running the full staged pipeline. The clean way to think of it: you don't need the enhancer resident while the long extension runs, and you don't need the extension model resident while enhancing. Loading them separately keeps your VRAM budget honest. The enhancement quality is decent for its era - don't expect modern Wan-level cleanups - but for a 2024 research port, going from 256px to 720p in one pass without visible seams is the genuinely impressive part of this pack.