TeleStyle Model Loader

TeleStyleLoader is the boring half of this pack, and it's the half doing the heavy lifting. It's a one-input node that assembles the entire TeleStyle video pipeline in memory: the Wan 2.1 1.3B diffusion transformer, Wan's 3D causal VAE, a flow-matching scheduler config, and the pre-computed style embeddings that make the whole thing run without a text encoder. Everything you need comes out of a single wire.

The context matters here. TeleStyle is real research - Tele-AI's content-preserving style transfer report landed in January 2026, and the community immediately wanted a ComfyUI wrapper for it. There was one catch: the original is built on Qwen-Image-Edit, and "lightweight" in the paper meant something different on the other side of a 40 GB model download. Danzelus's port rips the Qwen half out and keeps only the Wan 1.3B video stylization branch. Net result: ~6 GB of weights on disk, a 6 GB VRAM floor, and a video-only stylizer that would otherwise need a box you don't have.

The one input that matters

dtype - fp16 or bf16, default bf16. That's the entire parameter list, and bf16 is the right default: it's the best-quality option and what the author tuned against. fp16 exists for cards where bf16 is poorly supported, and it trades a little quality for a little VRAM. Want to A/B them? That takes two loaders, one per dtype - they don't share state.

How it actually works

First run is where the magic (and the download) happens. The loader pulls weights from the HuggingFace repo Danzelus/TeleStyle_comfy into ComfyUI/models/telestyle_models/ - dit.ckpt (the transformer checkpoint), prompt_embeds.pth (the style embeddings), and the diffusers-format Wan 2.1-T2V-1.3B VAE and scheduler. It builds the transformer from the shipped config, loads the checkpoint (handling the transformer_state_dict wrapper and module. prefix it ships with), loads the VAE, and reads those style embeddings.

That last piece is the clever bit. Because the style lives in a pre-computed embedding file, you never touch a CLIP or UMT5 text-encoding node. The negative side of the embedding is just zeros - at the inference node's default cfg of 1 there's no negative pass to run anyway. Simpler graph, faster load, less VRAM.

Install

Same steps as the whole pack (ComfyUI Manager → search "ComfyUI-TeleStyle", or):

cd ComfyUI/custom_nodes
git clone https://github.com/neurodanzelus-cmd/ComfyUI-TeleStyle.git
pip install -r requirements.txt

Then restart ComfyUI. Dependencies are diffusers, transformers, omegaconf, opencv-python and friends - nothing exotic, nothing that should collide with a normal video setup.

Where people get burned

• The first run looks hung. There's no download progress bar in the ComfyUI UI; the ~6 GB pull happens inside the Loader with output going to the terminal only. On a slow connection that's a real wait. Watch the console, not the node.
• It hard-targets cuda. No GPU, no loader.
• The weights come from the author's personal HF account, not Tele-AI's. This is an unofficial port of an unofficial distribution, and custom nodes execute arbitrary Python. Check the repo before you trust it - the usual rule.
• It's a 1.3B model. This is the low-VRAM lane for quick stylization drafts, not a 14B-quality renderer. Set expectations accordingly before you judge the output.

Output

One TELE_STYLE_MODEL, which wires into exactly one place: the model input on TeleStyle Video Transfer. You can't preview it, split it, or do anything else with it. It's the engine, and the other node is the gas pedal.