how to use stand-in lora in native workflow with wan?

[jasoncow007]

will you provide a workflow for native comfyui

[KBRASK]

Yes, developing the official ComfyUI integration is currently our top priority.

Stand-In has already been included in a workflow shared by the community member Kijai. You can get an early preview of Stand-In by replacing the preprocessing node in that workflow with ours.

[jasoncow007]

Yes, developing the official ComfyUI integration is currently our top priority.

Stand-In has already been included in a workflow shared by the community member Kijai. You can get an early preview of Stand-In by replacing the preprocessing node in that workflow with ours.

kijai的采样器很怪异，不同的模型出来的画质不一样，使用vace接入control和stand-in使得画质很难保证，不接入姿态控制还行，只是wan2.1的动态效果差点意思，当年同样使用gguf q6的模型发现在原生工作流里跑出来的视频画质原来可以这么好，难免会嫌弃kj的采样器有点暴殄天物

[Djanghost]

Can't wait the native integration ! 🙌

[jasoncow007]

Yes, developing the official ComfyUI integration is currently our top priority.

Stand-In has already been included in a workflow shared by the community member Kijai. You can get an early preview of Stand-In by replacing the preprocessing node in that workflow with ours.

When the proportion of the face in the video is small, it still causes distortion and lack of resemblance. Is there any good way to improve this?

[KBRASK]

Yes, developing the official ComfyUI integration is currently our top priority.
Stand-In has already been included in a workflow shared by the community member Kijai. You can get an early preview of Stand-In by replacing the preprocessing node in that workflow with ours.

When the proportion of the face in the video is small, it still causes distortion and lack of resemblance. Is there any good way to improve this?

Because our current version's training dataset is focused on mid-to-close range shots, bad cases do occasionally occur when generating distant scenes. Our next version, which will use an expanded dataset, will address this issue.

[makisekurisu-jp]

After testing, I found that it’s not compatible with ComfyUI’s official diffusion model loader or City96’s UNet GGUF loader. At the moment, it only works with Kijai’s custom nodes, and lacks native integration.

[KBRASK]

After testing, I found that it’s not compatible with ComfyUI’s official diffusion model loader or City96’s UNet GGUF loader. At the moment, it only works with Kijai’s custom nodes, and lacks native integration.经过测试，我发现它与 ComfyUI 官方的扩散模型加载器和 City96 的 UNet GGUF 加载器不兼容。目前，它仅支持 Kijai 的自定义节点，并且缺乏原生集成。

Stand-In is not a traditional diffusion LoRA; integrating it requires in-depth modifications to the pipeline, the forward process of the DiT blocks, and the internal implementation of the attention module. Therefore, we need to develop custom nodes, and we are accelerating the development. Please be patient.

[jasoncow007]

Has there been any progress on version wan2.2? @KBRASK
fun vace ， it seems that Introducing references will affect dynamics

[CriuslvCN]

Yes, developing the official ComfyUI integration is currently our top priority.
Stand-In has already been included in a workflow shared by the community member Kijai. You can get an early preview of Stand-In by replacing the preprocessing node in that workflow with ours.

kijai的采样器很怪异，不同的模型出来的画质不一样，使用vace接入control和stand-in使得画质很难保证，不接入姿态控制还行，只是wan2.1的动态效果差点意思，当年同样使用gguf q6的模型发现在原生工作流里跑出来的视频画质原来可以这么好，难免会嫌弃kj的采样器有点暴殄天物

严重同意以上观点，话说，兄弟，这个支持官方采样器和官方的节点吗？这个节点到底有用吗？