Qwen TeleStyle Loader

This is the shopping node. Two dropdowns, one wire out, and everything scary - the multi-gigabyte weights, the file locations, the pipeline assembly - happens here so the Apply node only ever sees a finished pipeline. If this node is your entry point to TeleStyle, the honest headline is: it works, it just costs you disk space and a first-run download.

What it does

Pick an architecture (2509 or 2511) and a device (cuda, cpu, or auto - auto just means "cuda if it's there"), and out comes a TELESTYLE_MODELS socket. Wire that into the Apply node's models input and you're done loading forever.

The two architectures are the two Qwen-Image-Edit builds, named for the year and month of the build: 2509 is the original (September 2025) release, 2511 the later revision with better character consistency and less drift. If you're starting fresh, 2511 is the one to try; 2509 is the more battle-tested default and has the bigger LoRA ecosystem around it.

How it works

On first run, the Loader resolves five files, downloading them from HuggingFace into ComfyUI's own model folders if they're missing:

• models/diffusion_models/Qwen-Image-Edit-2509.safetensors - the ~20B transformer, pulled as one pre-merged file. The 2511 build has no pre-merged file yet, so it downloads the original shards and merges them once (a single merge that wants ~38GB of RAM free).
• models/text_encoders/qwen_2.5_vl_7b.safetensors - shared across both architectures.
• models/vae/qwen_image_vae.safetensors - shared across both.
• the TeleStyleV2 style LoRA plus a Lightning 4-step speed LoRA, per architecture.

Then it builds a diffusers pipeline directly from those flat files in bf16 and caches it, keyed by files-and-device, so your next queue reuses the built pipeline instead of re-reading 54GB of weights.

The deliberate design choice: no ComfyUI MODEL/CLIP/VAE objects are involved. An earlier version bridged comfy's live model objects, which parked the ~54GB checkpoint in memory twice - comfy's own resident copy plus the pipeline's independent copy - and reliably OOM-killed the machine when system RAM was smaller than VRAM. This version loads each file exactly once, straight into its own modules. A side benefit: the text encoder keeps its lm_head, which a normal comfy CLIP object never carries, and that's what makes the Apply node's captioning toggles possible at all.

Installing

ComfyUI Manager: search ComfyUI-Qwen-TeleStyle. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/Mister-Link/ComfyUI-Qwen-TeleStyle

then restart ComfyUI. One real gotcha: the pack ships no requirements.txt. It leans on torch, safetensors and huggingface_hub (all already in ComfyUI's venv) plus transformers and diffusers, which your install may or may not have; qwen-vl-utils is only needed for the Apply captions. If you hit an import error at startup, install into ComfyUI's environment:

pip install transformers diffusers qwen-vl-utils

Common issues

The first run downloads roughly 54GB total. The community's reaction to the underlying model said it best: "I thought it would be lightweight, but it's not." It isn't - budget the disk, and ideally a 24GB+ card. The 2511 one-time shard merge wants a lot of system RAM. And remember the cache: once built, that bf16 pipeline parks on your GPU until you run the pack's Unload node. That's exactly why this pack ships one.