TeleStyle Image Model Loader (Official)

The name undersells this node. "TeleStyle Image Model Loader" sounds like a small utility that loads a LoRA. What it actually does is spin up a full Qwen-Image-Edit-2509 pipeline - transformer, Qwen2.5-VL text encoder, VAE - inside DiffSynth, bolt on TeleStyle's two LoRAs, and hand you a ready-to-run model object. It's the front half of this pack's "official" image style-transfer path, and its only job is to feed TeleStyle Image Transfer (Official).

What you're actually loading

TeleStyle is Tele-AI's content-preserving style transfer: give it a content image and a style reference, and it re-renders the content in the reference's style without melting the subject. It's built on Qwen-Image-Edit; TeleStyle bends that into style transfer with a LoRA, plus a Lightning 4-step LoRA so you're not stuck with 30 denoising steps.

This is the "official" route - the same DiffSynth pipeline the research code runs, unlike the pack's Wan 2.1 video nodes, which use precomputed style embeddings and skip the big text encoder. The distinction matters: the "6GB VRAM" sticker was written for the video path. The image path here is heavier - a multi-billion-parameter DiT plus a 7B text encoder - so on a consumer card you'll lean on offloading and run slower than the H100-class HuggingFace demos. Set expectations early.

How it works

Hit Queue and the node builds a cache key from your options. If it matches a previously loaded pipeline (the default cache_policy=reuse_cached), you get the cached model back instantly. Otherwise it does two things: downloads the LoRA weights - diffsynth_Qwen-Image-Edit-2509-telestyle.safetensors and the Lightning 4-step LoRA - from the Tele-AI/TeleStyle repo into ComfyUI/models/telestyle_models/weights/, then tells DiffSynth to fetch the base Qwen/Qwen-Image-Edit-2509 (transformer, text encoder, VAE) from HuggingFace. That first run is a big download - tens of gigabytes for base + LoRAs - fully automatic, which reads as "it's broken" if you're not watching the console. It isn't. Give it time.

The inputs that matter

Ten inputs, most fine at defaults. The ones you'll actually touch:

• dtype - fp16 (default) or bf16 (best quality, per the README). Start with fp16; reach for bf16 if quality bugs you.
• attention_mode - sdpa (safe default), flash_attn (forced Flash SDP kernel), or sage_attn (requires launching ComfyUI with --use-sage-attention). flash_attn is the usual speed bump but can produce black output on some setups - the node auto-retries with sdpa.
• auto_benchmark_attention - benchmarks flash vs sdpa at load and keeps the faster one. Adds a one-time warmup.
• cache_policy - reuse_cached (default) or force_reload. force_reload is the reset button: run it once when the pipeline is stuck or misbehaving after you changed options.
• vram_cleanup_before_load - none / soft_empty_cache / unload_all_models. Recovery tools for OOM situations, not for every run.
• enable_vram_management + vram_limit_gb - the low-VRAM levers. Slower, but they let a modest GPU breathe.
• compile_dit - torch.compile on the DiT; first run slower, later runs faster.
• clear_result_cache - wipes the inference node's cached image outputs.

What comes out

One output: model, type TELE_STYLE_IMAGE_MODEL - it wires only into TeleStyle Image Transfer (Official). Keep exactly one loader in the graph; changing loader options auto-clears old image caches to cut OOM risk.

Installing it

Here's the catch: the pack's requirements.txt does not include DiffSynth. The loader needs it, so your first run of the official image path throws a RuntimeError that helpfully prints the exact install command:

pip install git+https://github.com/modelscope/DiffSynth-Studio.git@11315d7 transformers==4.57.3 accelerate==1.2.1

That error is also the #1 community failure mode for this pack. People install DiffSynth with the wrong Python - on Windows Easy Install builds that's ComfyUI's bundled Python, not your system pip - and the error never goes away. The pins matter: that DiffSynth commit expects transformers 4.57.3 and accelerate 1.2.1. Full install:

cd ComfyUI/custom_nodes
git clone https://github.com/aistudynow/Comfyui-tetestyle-image-video.git
pip install -r requirements.txt
# then the DiffSynth line above, with ComfyUI's Python

ComfyUI Manager can find the pack by its title, but it only runs the requirements.txt - not that git+https extra. Check it manually.

When things go wrong

• The DiffSynth RuntimeError - read the message; it prints the exact fix. Run it with the Python that launched ComfyUI.
• Black output with flash_attn - it auto-retries with sdpa; or just set sdpa from the start.
• OOM / allocator asserts - vram_cleanup_before_load=soft_empty_cache once, or flip on enable_vram_management. If asserts persist, launch ComfyUI with --disable-cuda-malloc.
• Slow first load - that's the download. And slow runs on a small card are the price of the official path, not a bug.

As of mid-2026 a TeleStyle V2 exists in the community; this pack runs the V1 LoRAs, so check whether V2 matters to you first.