This is insanely good! Thank you for your commitment. Works like a charm even with Videos. And comparing it to generations without the lora it feels like a game-changer too bad the sample/schedular combination takes away the speed-boost I got used to. Nevermind. The Quality makes up for it.

Can you please post some comparisons of with and without using this Lora?

Please upload your others, this is brilliant work.

Is this Wan2.1Text2Image only a ComfyUi thing? Can I do it on WanGP2? This is my first time learning about this, very interesting.

This is really interesting, training loras for single images on a video model. I've been using Wan a lot, and it's become very clear to me that Wan understands how to make a coherent image better than any image generation model I've ever used.

It has weird transitions in animation sometimes, but it seems like all the training on video has helped it to understand visuals so much more deeply than training on still images. Like, of course it would know how to make a human body better if it also had to know it well enough to change the viewing angle at every single degree of rotation, and the bend of a limb at every single degree of bend, also viewed from every angle.

Wan actually feels like it knows what arms and legs do, and when it makes horrors, it's still using those parts with a higher level of coherence and logic. It seems like when it gets it wrong, it's merely getting confused about what you want, rather than being confused about the object it's showing, unlike models like Flux or SDXL, which you can tell can't really understand anything about the pose it tries to do, or how arms/legs work with that pose.

Anyway, this is an interesting idea, and I might start trying to use it this way.

Broken workflow. Even with all the pre-requisites installed, the link between image saver and scheduler fails, even though they're connected properly. Even tried changing image saver versions. Tried to fix manually, but ferris wheel image did not match at all with the same settings.

nice work

我想知道这个应该怎么训练，我有大量的数据集，但是我不知道怎么训练，如果你能告诉我怎么训练我将非常感激！！！

This is actually pretty amazing - I had no idea Wan could generate such excellent hi-res stills, but this works really well. Thanks :)

Awesome stuff! Did you share your Lora training config for want2v anywhere yet?

Very cool, thx for sharing! Could text2image also work with Wan character Loras? Would love to try it with your Aloy Lora

The word "phone" in your Trigger sentence adds a phone to the image. It is more effective to have a meaningless collection of letters as a Trigger word.

Where do you get the bong_tangent scheduler? And the res_2s sampler? Mine gives an error why I try to use it, and it's not selectable in the list. It's only in your workflow.

Is this LoRA so unassuming? In the field of image generation, it's even like the quality of the next version of the model, truly stunning

can you do a inpainting at full resolution workflow with this? Pleeeease thank you :D

With Wan 2.1 for t2i I was able to get consistently good complex interactions and motion, and with this Lora I was able to get back natural looking images. Both the people it invents and the camera angles it choses don't feel like they were made with AI. I can't believe both of these problems got resolved at one time.

This is a must have if you want the highest quality image!

Yes it does work for video generation and it's pretty awesome! I used it with Wan 2.2