SD1.5 模型:control_v11p_sd15_canny、t2iadapter_canny_sd1.5v2 SDXL 模型:control-lora-canny-rank128、control-lora-canny-rank256 ④效果预览: 2. HED 模糊线预处理器: ①介绍:从图像中提取边缘线,相对于 Canny 处理的锐利边缘线,HED 提供了边缘过渡,可以保留更多柔和的边缘细节,类似手绘效果(好处:通过线...
或: huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank128https://huggingface.co/stabilityai/control-lora/tree/main/control-LoRAs-rank256 stabilityai/control-lora at main 文件名: control-lora-canny-rank128.safetensors control-lora-depth-rank128.safetensors control-lora-canny-ran...
"https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank128/control-lora-canny-rank128.safetensors", "https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank128/control-lora-depth-rank128.safetensors", ...
创意无限,画质优良,超分修复更精准!LoRA支持,随心生成大图,支持视频,控制网络助力创作。高清修复v2,更强修复控制。自定义VAE、直接获取大模型,创作自由轻松!造梦神器,快来体验吧 - 飞桨AI Studio星河社区
depth): "sai_xl_depth_256lora.safetensors", resource_id(ResourceKind.controlnet, SDVersion.sd15, ControlMode.normal): None, resource_id(ResourceKind.controlnet, SDVersion.sd15, ControlMode.pose): "control_lora_rank128_v11p_sd15_openpose_fp16.safetensors", resource_id(ResourceKind.control...
The 128 and 256-rank LoRA perform very similarly. You can use the 128 variant if you want to conserve space. sai_xl_canny_128lora (weight 0.75) sai_xl_canny_256lora (weight: 0.75) T2I adapter t2i-adapter_xl_canny t2i-adapter_diffusers_xl_canny ...
rank=64 # lora的rank max_training_steps=2000 # 最大训练迭代步数 PYTHONPATH=./ deepspeed hydit/train_deepspeed.py \ --task-flag ${task_flag} \ --model ${model} \ --training_parts lora \ --rank ${rank} \ --resume-split \
8、Network Rank (Dimension”网络大小):强化训练细节,建议128〜192,128以上增加提升相对不明显。 9、Network Alpha (网络Alpha):建议96以上,弱化训练细节,有正则化效果,可与Dim同步增加。 10、让AI训练AI:首发训练采用Dadaptation,所有学习率均设为1。
resource_id(ResourceKind.controlnet,SDVersion.sdxl,ControlMode.pose): ["control-lora-openposexl2-rank","thibaud_xl_openpose"], resource_id(ResourceKind.controlnet,SDVersion.sd15,ControlMode.segmentation): ["control_v11p_sd15_seg","control_lora_rank128_v11p_sd15_seg"], ...
zero.GatheredParameters(param, modifier_rank=None) for deepspeed support. it's starting to feel like we need an autocast manager wrapper in train utils 👍 1 examples/instruct_pix2pix/train_instruct_pix2pix_sdxl.py Outdated Show resolved examples/instruct_pix2pix/train_instruct_pix2pix_sd...