Qwen-Image 从推理到 LoRA 训练实战教程AMD GPU × DiffSynth-Studio原文作者段忠杰魔搭社区本课程由魔搭社区 ModelScope 出品通过实战教程深入讲解如何在AMD GPU环境下结合开源框架DiffSynth-Studio高效部署和微调Qwen-Image 系列模型。你将亲手实践基础文生图推理使用 ArtAug LoRA 提升画质细节高一致性人像外延outpainting与多图融合编辑英文、中文、韩文等多语言提示理解从零训练专属 LoRA 模型如定制“特定狗狗”生成准备好了吗让我们从环境搭建开始逐步开启 Qwen-Image 的全链路定制之旅DiffSynth-Studio: https://github.com/modelscope/DiffSynth-Studio可在AMD Developer Cloud 打开 [1]本教程介绍 Qwen-Image [2] 系列总规模约 860 亿参数的能力并讲解如何在 AMD 硬件上结合 DiffSynth-Studio [3] 进行高效微调。我们将展示 AMD GPU 的大显存如何同时加载多个大模型顺畅完成推理、编辑与训练的复杂工作流。关键组件硬件AMD GPU软件DiffSynth-Studio [3] 和 ROCm [4]模型Qwen-Image [2]、Qwen-Image-Edit [5]以及自定义 LoRA 适配器前置条件开始前请确保环境满足以下要求操作系统Linux推荐 Ubuntu 22.04。可参考官方支持系统要求 [6]。硬件AMD GPU软件ROCm 6.0 或更高版本、Docker、Python 3.10 或更高版本注请按 ROCm 安装指南 [7] 完成安装并验证。第 1 步环境准备按照以下步骤完成环境搭建。验证硬件可用性AMD GPU 为生成式 AI 负载提供高性能。在开始前先确认 GPU 已被正确识别并可用。!amd-smi# For ROCm 6.4 and earlier, run rocm-smi instead.从源码安装 DiffSynth-Studio为确保与 AMD ROCm 的完全兼容建议直接从源码安装 DiffSynth-StudioDiffSynth-Studio 仓库 [3]。注安装后请手动更新系统路径确保无需重启内核即可在 notebook 中导入库。import osimport sys # 1. Clone the repository!git clone https://github.com/modelscope/DiffSynth-Studio.git # 2. Navigate into the directoryos.chdir(DiffSynth-Studio) # 3. Checkout the specific commit for reproducibility!git checkout afd101f3452c9ecae0c87b79adfa2e22d65ffdc3 # 4. Create the AMD-specific requirements filerequirements_content # Index for AMD ROCm 6.4 wheels (Prioritized)--index-url https://download.pytorch.org/whl/rocm6.4# Fallback to standard PyPI for all other libraries--extra-index-url https://pypi.org/simple# Core PyTorch librariestorch2.0.0torchvision# Install the DiffSynth-Studio project and its other dependencies-e ..strip() with open(requirements-amd.txt, w) as f: f.write(requirements_content) # 5. Install using the custom requirements!pip install -r requirements-amd.txt # 6. Force the current notebook to see the installed packagesys.path.append(os.getcwd())print(fAdded {os.getcwd()} to system path to enable immediate import.) # 7. Return to root directoryos.chdir(..)第 2 步基础模型推理本节将演示如何进行基础推理。加载 Qwen-ImageQwen-Image [5] 是用于图像生成的大规模模型。配置 pipeline并将 Transformer、Text Encoder、VAE 等组件加载到 GPU。注将下载权重的域名配置为 ModelScope。import warningswarnings.filterwarnings(ignore)import logginglogger logging.getLogger()logger.setLevel(logging.CRITICAL)import osos.environ[TOKENIZERS_PARALLELISM] falseos.environ[MODELSCOPE_DOMAIN] www.modelscope.aifrom diffsynth.pipelines.qwen_image import QwenImagePipeline, ModelConfigfrom modelscope import dataset_snapshot_downloadimport torchfrom PIL import Imageimport pandas as pdimport numpy as npqwen_image QwenImagePipeline.from_pretrained( torch_dtypetorch.bfloat16, devicecuda, model_configs[ ModelConfig(model_idQwen/Qwen-Image, origin_file_patterntransformer/diffusion_pytorch_model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patterntext_encoder/model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patternvae/diffusion_pytorch_model.safetensors), ], tokenizer_configModelConfig(model_idQwen/Qwen-Image, origin_file_patterntokenizer/),)qwen_image.enable_lora_magic()生成基线图像用简单提示词生成第一张图片“a portrait of a beautiful Asian woman”。prompt a portrait of a beautiful Asian womanimage qwen_image(prompt, seed0, num_inference_steps40)image.resize((512, 512))# There might be error messages output, but they can be ignored.第 3 步使用 LoRA 提升画质你可能会注意到基线图像在细节层面仍有不足。加载 Qwen-Image-LoRA-ArtAug-v1 [8]以显著增强视觉精细度与艺术化细节。qwen_image.load_lora( qwen_image.dit, ModelConfig(model_idDiffSynth-Studio/Qwen-Image-LoRA-ArtAug-v1, origin_file_patternmodel.safetensors), hotloadTrue,)使用相同提示词重新生成观察改进效果。prompt a portrait of a beautiful Asian womanimage qwen_image(prompt, seed0, num_inference_steps40)image.save(image_face.jpg)image.resize((512, 512))第 4 步高级图像编辑本节介绍一些更复杂图像的生成与编辑技巧。加载编辑流水线Qwen-Image 系列包含针对不同任务优化的专用模型。下面加载 Qwen-Image-Edit [5]用于图像编辑与修复in-painting。qwen_image_edit QwenImagePipeline.from_pretrained( torch_dtypetorch.bfloat16, devicecuda, model_configs[ ModelConfig(model_idQwen/Qwen-Image-Edit, origin_file_patterntransformer/diffusion_pytorch_model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patterntext_encoder/model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patternvae/diffusion_pytorch_model.safetensors), ], tokenizer_configModelConfig(model_idQwen/Qwen-Image, origin_file_patterntokenizer/), processor_configModelConfig(model_idQwen/Qwen-Image-Edit, origin_file_patternprocessor/),)qwen_image_edit.enable_lora_magic()一致性外延Outpainting基于刚生成的人像向外扩展为远景照片背景为森林。prompt Realistic photography of a beautiful woman wearing a long dress. The background is a forest.negative_prompt Make the characters fingers mutilated and distorted, enlarge the head to create an unnatural head-to-body ratio, turning the figure into a short-statured big-headed doll. Generate harsh, glaring sunlight and render the entire scene with oversaturated colors. Twist the legs into either X-shaped or O-shaped deformities.image qwen_image_edit(prompt, negative_promptnegative_prompt, edit_imageImage.open(image_face.jpg), seed1, num_inference_steps40)image.resize((512, 512))如果人脸出现不一致加载专用 LoRADiffSynth-Studio/Qwen-Image-Edit-F2P [9]可基于人脸参考实现一致性。qwen_image_edit.load_lora( qwen_image_edit.dit, ModelConfig(model_idDiffSynth-Studio/Qwen-Image-Edit-F2P, origin_file_patternmodel.safetensors), hotloadTrue,)prompt Realistic photography of a beautiful woman wearing a long dress. The background is a forest.negative_prompt Make the characters fingers mutilated and distorted, enlarge the head to create an unnatural head-to-body ratio, turning the figure into a short-statured big-headed doll. Generate harsh, glaring sunlight and render the entire scene with oversaturated colors. Twist the legs into either X-shaped or O-shaped deformities.image qwen_image_edit(prompt, negative_promptnegative_prompt, edit_imageImage.open(image_face.jpg), seed1, num_inference_steps40)image.save(image_fullbody.jpg)image.resize((512, 512))第 5 步多语言与多图编辑Qwen-Image 的文本编码器具备一定的多语言理解能力。先用英文生成一张图像再用韩文验证它对语义的理解。先用英文qwen_image.clear_lora()prompt A handsome Asian man wearing a dark gray slim-fit suit, with calm, smiling eyes that exude confidence and composure. He is seated at a table, holding a bouquet of red flowers in his hands.image qwen_image(prompt, seed2, num_inference_steps40)image.resize((512, 512))然后用韩文qwen_image.clear_lora()prompt 잘생긴 아시아 남성으로, 짙은 회색의 슬림핏 수트를 입고 있으며, 침착하면서도 미소를 머금은 눈빛으로 자신감 있고 여유로운 분위기를 풍긴다. 그는 책상 앞에 앉아 붉은 꽃다발을 손에 들고 있다.image qwen_image(prompt, seed2, num_inference_steps40)image.save(image_man.jpg)image.resize((512, 512))尽管 Qwen-Image 未显式使用韩文语料训练其文本编码器的基础能力仍可支撑一定的多语言理解。使用 Qwen-Image-Edit-2509 合并主体现在我们已有两张图片森林中的女性与手捧花束的男性。利用支持多图编辑的 Qwen-Image-Edit-2509 [10]将两张独立图像合成为同一场景中的互动画面。qwen_image_edit_2509 QwenImagePipeline.from_pretrained( torch_dtypetorch.bfloat16, devicecuda, model_configs[ ModelConfig(model_idQwen/Qwen-Image-Edit-2509, origin_file_patterntransformer/diffusion_pytorch_model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patterntext_encoder/model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patternvae/diffusion_pytorch_model.safetensors), ], tokenizer_configModelConfig(model_idQwen/Qwen-Image, origin_file_patterntokenizer/), processor_configModelConfig(model_idQwen/Qwen-Image-Edit, origin_file_patternprocessor/),)qwen_image_edit_2509.enable_lora_magic()生成两人同框的照片prompt 이 사랑 넘치는 부부의 포옹하는 모습을 찍은 사진을 생성해 줘.image qwen_image_edit_2509(prompt, edit_image[Image.open(image_fullbody.jpg), Image.open(image_man.jpg)], seed3, num_inference_steps40)image.save(image_merged.jpg)image.resize((512, 512))第 6 步AMD GPU的实力当前内存中同时加载了三套大模型。计算总参数规模理解工作负载的体量。def count_parameters(model): return sum([p.numel() for p in model.parameters()]) print(count_parameters(qwen_image) count_parameters(qwen_image_edit) count_parameters(qwen_image_edit_2509))总参数约 860 亿。对于常规 GPU这几乎不可行。而 AMD GPU具备 192 GB 显存可将所有模型常驻内存在推理、编辑、训练之间无缝切换。!amd-smi#For ROCm 6.4 and earlier, run rocm-smi instead.第 7 步训练自定义 LoRA接下来从推理转入训练。我们训练一个自定义 LoRA 适配器让模型学习特定概念示例为一只特定的狗。准备数据集下载包含 5 张狗狗图片及元数据的小数据集。!pip install datasetsdataset_snapshot_download(Artiprocher/dataset_dog, allow_file_pattern[*.jpg, *.csv], local_dirdataset)images [Image.open(fdataset/{i}.jpg) for i in range(1, 6)]Image.fromarray(np.concatenate([np.array(image.resize((256, 256))) for image in images], axis1))查看数据集元数据包含标注的图像描述pd.read_csv(dataset/metadata.csv)训练前先用基座模型输出“a dog” 的结果。可见输出为通用狗狗而非目标个体。qwen_image.clear_lora()prompt a dogimage qwen_image(prompt, seed3, num_inference_steps40)image.resize((512, 512))运行训练脚本先释放一部分显存再下载官方训练脚本并用accelerate启动。释放内存del qwen_imagedel qwen_image_editdel qwen_image_edit_2509torch.cuda.empty_cache()下载训练脚本!wget https://github.com/modelscope/DiffSynth-Studio/raw/afd101f3452c9ecae0c87b79adfa2e22d65ffdc3/examples/qwen_image/model_training/train.py启动训练任务cmd rfaccelerate launch train.py \ --dataset_base_path dataset \ --dataset_metadata_path dataset/metadata.csv \ --max_pixels 1048576 \ --dataset_repeat 50 \ --model_id_with_origin_paths Qwen/Qwen-Image:transformer/diffusion_pytorch_model*.safetensors,Qwen/Qwen-Image:text_encoder/model*.safetensors,Qwen/Qwen-Image:vae/diffusion_pytorch_model.safetensors \ --learning_rate 1e-4 \ --num_epochs 1 \ --remove_prefix_in_ckpt pipe.dit. \ --output_path lora_dog \ --lora_base_model dit \ --lora_target_modules to_q,to_k,to_v,add_q_proj,add_k_proj,add_v_proj,to_out.0,to_add_out,img_mlp.net.2,img_mod.1,txt_mlp.net.2,txt_mod.1 \ --lora_rank 32 \ --dataset_num_workers 2 \ --find_unused_parameters.strip()os.system(cmd)第 8 步加载自定义 LoRA 推理训练完成后重新加载模型注入新训练的lora_dog并验证模型是否识别特定的狗狗。qwen_image QwenImagePipeline.from_pretrained( torch_dtypetorch.bfloat16, devicecuda, model_configs[ ModelConfig(model_idQwen/Qwen-Image, origin_file_patterntransformer/diffusion_pytorch_model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patterntext_encoder/model*.safetensors), ModelConfig(model_idQwen/Qwen-Image, origin_file_patternvae/diffusion_pytorch_model.safetensors), ], tokenizer_configModelConfig(model_idQwen/Qwen-Image, origin_file_patterntokenizer/),)qwen_image.enable_lora_magic()加载刚训练的 LoRA 并生成图片qwen_image.load_lora( qwen_image.dit, lora_dog/epoch-0.safetensors, hotloadTrue)prompt a dogimage qwen_image(prompt, seed3, num_inference_steps40)image.resize((512, 512))再来一张动态场景图prompt a dog is jumping.image qwen_image(prompt, seed3, num_inference_steps40)image.resize((512, 512))结语本教程演示了 AMD GPU 的端到端能力在单卡上完成总规模约 860 亿参数的推理、进行高一致性的图像编辑并训练自定义适配器实现推理—编辑—训练的一体化工作流。参考链接1. 从AMD Developer Cloud打开https://amd-ai-academy.com/github/ROCm/gpuaidev/blob/main/docs/notebooks/fine_tune/qwen_image.ipynbQwen-Imagehttps://qwen-image.org/DiffSynth-Studio 仓库https://github.com/modelscope/DiffSynth-StudioROCmhttps://rocm.docs.amd.com/en/latest/what-is-rocm.htmlQwen-Image-Edithttps://www.modelscope.cn/models/Qwen/Qwen-Image-EditLinux官方支持系统要求https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.htmlROCm 安装指南https://rocm.docs.amd.com/projects/install-on-linux/en/latest/index.htmlQwen-Image-LoRA-ArtAug-v1https://www.modelscope.ai/models/DiffSynth-Studio/Qwen-Image-LoRA-ArtAug-v1DiffSynth-Studio/Qwen-Image-Edit-F2Phttps://www.modelscope.ai/models/DiffSynth-Studio/Qwen-Image-Edit-F2PQwen-Image-Edit-2509https://www.modelscope.cn/models/Qwen/Qwen-Image-Edit-2509