TeleStyleV2-Qwen-Image-Edit-2511-bf16

A ready-to-run, pre-fused checkpoint of Tele-AI/TeleStyleV2 for content-preserving image style transfer, built on Qwen/Qwen-Image-Edit-2511.

TeleStyleV2 ships only LoRA adapters. This repo has those adapters merged into the base transformer at scale 1.0, exactly as the reference app does — so no LoRA loading is needed:

• diffusers-TeleStyleV2-QIE-2511-Lora (the style-transfer LoRA)
• QIE-2511-Lightning-4steps (the Lightning / Distribution-Matching-Distillation LoRA → 4-step inference)

Everything else (vae/, text_encoder/ = Qwen2.5-VL-7B, processor/, tokenizer/, scheduler/) is the unmodified Qwen-Image-Edit-2511 base, so the repo is a drop-in self-contained snapshot.

What it does

Given a content image and a style image, it transfers the style of the style image onto the content image while preserving the content — across realistic/stylized pairs (R+R, R+S, S+R, S+S). It also serves plain instruction edits (content image only). Image 1 = content, Image 2 = style; there is no role token, the fused LoRA learned the order.

Usage (diffusers)

import torch
from diffusers import QwenImageEditPlusPipeline
from PIL import Image

pipe = QwenImageEditPlusPipeline.from_pretrained(
    "mlx-community/TeleStyleV2-Qwen-Image-Edit-2511-bf16", torch_dtype=torch.bfloat16
).to("cuda")

content = Image.open("content.png")
style = Image.open("style.png")
image = pipe(
    image=[content, style],
    prompt="Style Transfer the style of Figure 2 to Figure 1, and keep the content "
           "and characteristics of Figure 1.",
    num_inference_steps=4,        # Lightning/DMD
    true_cfg_scale=1.0,           # DMD: no CFG
    generator=torch.manual_seed(123),
).images[0]
image.save("out.png")

Usage (MLX / Apple Silicon)

Consumable by qwen-image-edit-swift (the MLXTeleStyle package, an MLXEngine imageEdit surface with a styleTransfer mode). The Swift port reads this diffusers snapshot directly. ~60 GB resident in bf16 (20B DiT + Qwen2.5-VL-7B + fp32 VAE); 4-step DMD inference.

License & credits

Apache-2.0. Built on Tele-AI/TeleStyleV2 (the LoRAs) and Qwen/Qwen-Image-Edit-2511 (the base). Merge performed with diffusers fuse_lora(scale=1.0). Please cite the TeleStyleV2 technical report (arXiv:2601.20175).