FL DiffVSR Load Model

FL DiffVSR Load Model doesn't upscale a single pixel. It's the loader half of the ComfyUI-FL-DiffVSR pack, and it exists for one job: get the Stream-DiffVSR model onto your machine and into VRAM, then hand a ready-to-use handle to the FL DiffVSR Upscale node. Think of it as the "load checkpoint" step for a diffusion video upscaler - the part that does the heavy setup so the other node can focus on frames.

It's the first thing that runs, and the first time it runs it does something unusual for a ComfyUI node: it downloads its own model. On first use it pulls a ~2GB snapshot of Jamichsu/Stream-DiffVSR from HuggingFace into ComfyUI/models/stream_diffvsr/, organized into unet, controlnet, vae, text_encoder, tokenizer, and scheduler subfolders. No manual placement, no model page to babysit. If the folders already have files, it skips the download entirely and just loads.

The inputs that matter

• precision - auto, fp32, fp16, bf16. Auto picks fp16 on a GPU and fp32 on CPU, which is the sane default. fp16 is the one you'll actually run with; bf16 is there for cards that prefer it.
• device - auto, cuda, cpu. Auto grabs CUDA if it's available. CPU technically works, but a full diffusion upscaler on CPU is a patience exercise - the README says NVIDIA GPU recommended, and it means it.
• enable_xformers - default true. Memory-efficient attention, and notably the pack only honors it when you're actually on CUDA. If it can't enable it (missing or mismatched xformers build), it logs and continues rather than failing.

The single output is model, a typed FL_DIFFVSR_MODEL that wires straight into the model socket on FL DiffVSR Upscale. There's nothing else to connect.

The downloads nobody warns you about

The ~2GB snapshot is only the headline. Two more pieces come down on that first run: the pack's bundled tokenizer is incomplete (missing merges.txt), so the wrapper silently falls back to openai/clip-vit-large-patch14 from HuggingFace, and the RAFT optical-flow model that drives the temporal coherence loads its own weights through torchvision. All three downloads happen on first queue, so budget the bandwidth and don't panic when the first run takes a while. Offline, that first run will fail - the model genuinely has to come from somewhere.

Install

ComfyUI Manager, search "FL DiffVSR", install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-DiffVSR.git
cd ComfyUI-FL-DiffVSR
pip install -r requirements.txt

The requirements are the standard torch / diffusers / transformers / xformers stack on Python 3.10+, and 8GB VRAM is the floor (16GB+ is the comfortable zone for longer clips).

Common issues

• Long first load. It's downloading a 2GB model plus two side downloads. Not hung - working.
• Precision changes don't take effect mid-graph. Change precision or device and the queue re-loads the whole pipeline in the new dtype; that's normal, not a crash.
• xformers warning on the console. On CUDA it means your xformers build is mismatched with torch; the node keeps going with standard attention, just slower. On CPU it's expected and harmless.