Load EvTexture Model

The Two Checkpoints Behind EvTexture

Everything in this pack eventually funnels into EVTTextureUpscaleVideo, and that node needs a model - which is what EVTLoadEvTextureModel provides. It's the loader, the same role Load Checkpoint plays in any generation workflow: point it at a weights file, get back a model object the graph can consume. Nothing runs at load time, so this node is basically free; the VRAM bill arrives at the upscaler.

The weights are the whole game

Here's the part the README could be louder about: the pack does not ship any weights. You download them yourself from the upstream research repo, DachunKai/EvTexture (Releases, Google Drive, or OneDrive), and drop them in your upscale_models folder. Two official checkpoints exist, both trained for 4x super-resolution with bicubic downsampling:

• EvTexture_REDS_BIx4.pth - trained on the REDS dataset
• EvTexture_Vimeo90K_BIx4.pth - trained on Vimeo-90K

The pack author's own guidance: these models are at their best on low-resolution frames (roughly 64–256 px) packed with detail, and the Vimeo90K version shows fewer artefacts on larger images. So your default should be Vimeo90K, and your expectations for feeding it 1080p should be modest.

The inputs and output

• model_name - a dropdown populated from whatever sits in ComfyUI/models/upscale_models (same folder your ESRGAN models live in, so don't be surprised to see them listed). Pick the .pth.

Output: model (EVTEXTURE_MODEL) - feeds the upscale node's model input. That's the entire contract.

How it loads

Internally it builds the EvTexture network - the architecture is vendored straight from the upstream basicsr code, SPyNet flow network included - then does torch.load(..., weights_only=True)["params_ema"] and loads it strictly. What that means for you: no separate SPyNet file to hunt down. The official checkpoints carry the full state dict, and the strict load will complain loudly (rather than silently) if you hand it the wrong file.

Install

The pack installs the usual way:

cd ComfyUI/custom_nodes
git clone https://github.com/tocubed/ComfyUI-EvTexture

(Or via ComfyUI Manager, searching "ComfyUI-EvTexture" - same result, plus restart.) The only extra step for this node is the manual model download above. The pack's own dependencies are just torch and torchvision, which you already have.

Common issues

Most "why is my load broken" reports reduce to one of two things: the file isn't actually in upscale_models (double-check the exact folder path), or you grabbed a mismatched .pth and the strict load refused it. Note the model list doesn't refresh itself mid-session - restart ComfyUI after dropping a checkpoint in. And set expectations: this is a small, single-commit wrapper around a research model, so don't expect a curated model catalog or auto-updating weights. You get the two checkpoints, you feed them the right input sizes, and that's the whole deal.