This node is specifically designed for Video Style Transfer using the Wan2.1-T2V architecture and TeleStyle custom weights. Unlike the original repository, this implementation strips away all heavy image-editing components (Qwen weights) to focus purely on video generation with speed/quality.

Requirements

GPU VRAM: 6GB minimum
Disk Space: ~6GB for models and weights

https://aistudynow.com/how-to-fix-slow-style-transfer-in-comfyui-run-telestyle-on-6gb-vram/

Youtube Tutorial: https://www.youtube.com/watch?v=yHbaFDF083o

✨ Key Features

High Performance:
- Acceleration: Built-in support for Flash Attention 2 and SageAttention for faster inference.
- Fast Mode: Optimized memory management with aggressive cache cleanup to prevent conflicts between CPU offloading and GPU processing.
Simplified Workflow: No need for complex external text encoding nodes. The model uses pre-computed stylistic embeddings (prompt_embeds.pth) for maximum efficiency.

📦 Installation

Navigate to your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes/

Clone this repository:

git clone https://github.com/aistudynow/Comfyui-tetestyle-image-video.git

Install dependencies:

pip install -r requirements.txt

Note: For SageAttention support, you may need to install sageattention manually.

📂 Model Setup

This node requires specific weights placed in the ComfyUI/models/telestyle_models/ directory.

The weights are downloaded automatically at the first run

Directory Structure:

ComfyUI/
└── models/
    └── telestyle_models/
        ├── weights/
        │   ├── dit.ckpt            # Main Video Transformer weights
        │   └── prompt_embeds.pth   # Pre-computed style embeddings
        └── Wan2.1-T2V-1.3B-Diffusers/
        │   ├── transformer_config.json
        │   ├── vae/
        │   │   │   ├── diffusion_pytorch_model.safetensors
        │   │   │   └── config.json
        │   └── scheduler/
        │       └── scheduler_config.json

Where to get weights:

https://huggingface.co/Danzelus/TeleStyle_comfy/tree/main

🚀 Usage

FP8 Qwen Image Workflow (Comfy Native)

Import:

workflow/TeleStyle_Qwen_FP8_Image.json

Expected model files:

models/unet/Qwen-Image-Edit-2509_fp8_e4m3fn.safetensors
models/clip/qwen_2.5_vl_7b.safetensors
models/vae/qwen_image_vae.safetensors
models/loras/diffsynth_Qwen-Image-Edit-2509-telestyle.safetensors
models/loras/diffsynth_Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors

1. TeleStyle Model Loader

This node loads the necessary model components.

| Parameter | Description | |-----------|-------------| | dtype | Choose between bf16 (best quality), fp16 |

2. TeleStyle Video Transfer

The main inference node.

| Parameter | Description | |-----------|-------------| | model | Connect the output from the Loader | | video_frames | Input video batch (from Load Video or VHS_LoadVideo) | | style_image | A reference image to guide the style transfer | | steps | Inference steps (default: 12) | | cfg | Guidance scale (default: 1) | | time_schedule | Backward-compatibility selector for old workflows (fixed) | | scheduler | Choose your sampler (FlowMatchEuler, UniPC, DPM++) | | fast_mode | Keep True for speed. Set to False for low-VRAM offloading (slower) | | acceleration | sdpa - PyTorch SDPA (safest default)<br>flash_attn - Force Flash SDP kernel when available<br>sage_attn - Use ComfyUI SageAttention backend (start ComfyUI with --use-sage-attention) |

3. TeleStyle Image Transfer (Official, Recommended)

These nodes wrap the official TeleStyle image architecture (Qwen Image Edit + TeleStyle LoRA):

TeleStyle Image Model Loader (Official)
TeleStyle Image Transfer (Official)

Install extra dependencies once:

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

| Parameter | Description | |-----------|-------------| | model | Connect from TeleStyle Image Model Loader (Official) | | content_image | Input content image | | style_image | Style reference image | | prompt | Style transfer instruction text | | steps | Inference steps (official lightning default: 4) | | seed | Random seed | | min_edge | Resize target for the shorter edge (multiple of 16) |

TeleStyle Image Model Loader (Official) extra performance options:

| Parameter | Description | |-----------|-------------| | attention_mode | Default attention backend for this loaded pipeline (sdpa, flash_attn, sage_attn) | | auto_benchmark_attention | Benchmarks flash_attn vs sdpa and auto-picks fastest stable mode for this load | | cache_policy | reuse_cached keeps model in memory for faster next runs, force_reload clears and reloads | | vram_cleanup_before_load | Optional Comfy cleanup before load: none, soft_empty_cache, unload_all_models | | compile_dit | Enable torch.compile on Qwen DiT (first run slower, later runs can be faster) | | enable_tf32 | Enable TF32 matmul/cudnn speedups on supported NVIDIA GPUs | | enable_vram_management | Enable DiffSynth offload/VRAM management for low-memory GPUs (slower but safer) | | vram_limit_gb | Optional VRAM budget hint for DiffSynth (0 = auto/default behavior) | | clear_result_cache | Clears cached image outputs for this node run |

Official loader mode:

HuggingFace-only pipeline for stability.
Local model-file selection has been removed from this node.

Fast defaults:

dtype=fp16
attention_mode=sdpa (stable default)
auto_benchmark_attention=True (if you want auto-pick for your GPU)
cache_policy=reuse_cached
compile_dit=False (set True only after a first successful run)
enable_vram_management=False
steps=4
min_edge=512

Performance notes:

Official image inference now has an in-memory result cache for repeated identical runs (same content/style/prompt/seed/steps/size).
auto_benchmark_attention adds a short warmup benchmark on load; reload is slower once, then cached runs are faster.
If flash_attn gives invalid/black output, the node now retries once automatically with sdpa.
If PyTorch raises CUDAMallocAsyncAllocator internal assert at high resolution, the node auto-recovers once (soft cache cleanup + TF32 off + sdpa retry).
If a run gets stuck or behaves wrong, set cache_policy=force_reload once to clear runtime cache and reload.
vram_cleanup_before_load=soft_empty_cache or unload_all_models is for recovery/OOM situations, not for every run.
attention_mode=sage_attn uses ComfyUI global SageAttention backend and requires launching ComfyUI with --use-sage-attention.
enable_vram_management=True is useful for low VRAM but can be much slower because of extra CPU/GPU movement.
If allocator asserts continue, launch ComfyUI with --disable-cuda-malloc to avoid CUDA async allocator issues.
HuggingFace Space demos often run on much larger GPUs (for example H100-class), so local consumer GPUs can be significantly slower at the same resolution.
Keep only one TeleStyle Image Model Loader (Official) in the graph. Changing loader options creates a new model config; this build now auto-clears old TeleStyle image caches to reduce OOM risk.
Qwen Image Edit uses separate components (DiT + text encoder + VAE). A single merged DiT+text-encoder safetensors file is not the standard format for this official DiffSynth pipeline.

To-Do List

[x] Initial release
[ ] More samplers
[ ] Consistency for very long videos

Guys, I’d really appreciate any support right now. I’m in a tough spot:

📜 Credits

This project is an unofficial implementation based on the amazing work by the original authors. Please refer to their repository for the original research and model weights.