Works well. Thank you.

fantastic lora with the node, works as described

the overlay example definetily looks good, i'll give a try

does this work for other features or poses that arent cloth or clothing related? if so is there something I am missing to accomplish that

Working with new 2509?

niceeeeeeee, i've actually tried it today after being disappointed with qwen edit and man... THIS LORA WORKS! The images got instatly better, thanks a lot!

can you create a wrokflow that goes with swarm? these JSONs dont load properly into Swarm's comfy back end and i'm not a node person.

When using with Qwen-Image-Edit-Rapid-AIO, a grid appears on the generated image. What could be the problem?

What is the recommended prompt for this to work properly?

At the risk of using a worn-out phrase: This Lora is a game-changer! Had remarkable success with Qwen image edit 2509_Q5_K_M on an old GTX 1070; I'm finally able to rotate a character and keep consistent body and face. I don't know what voodoo you pulled with this lora but I notice excellent results on an 8 step lightning lora at just the 2nd iteration (25%).

Great Lora! It works perfect! Thank you!

Hello, this is great! Thank you for your work. What is VL Encode?

Hello, I'm debugging this LoRA using diffusers. I found that those layer names for image stream is NOT in diffusers format, hence they aren't loaded by diffusers actually (e.g., ...attn.to_q.lora.down.weight, the correct name might be ...attn.to_q.lora_A.weight).

However, only under this strange setting does this LoRA work like a charm. I manually re-named those layers to the diffusers format ( lora.down -> lora_A, lora.up -> lora_B) but got very bad results.

Any idea of what is going on here?

Is this face consistency lora or cloth consistency or what am I missing?

Works great, faces, clothes and bodies are all as they should be.

Wow. This is such a game changer. Literally super accurate and without problems even on lower VRAM such as 12Gb. Thank you.

For the past month I've been trying to figure out how best to generate datasets from reference images to train LoRa's, rotating the camera around the character etc.. and I've been struggling with character likeness, having to generate tons of batches to sort through for good seeds. Sometimes sorting through hundreds of images to end up with only half a dozen.

I Tried so many different LoRa's, tons of different prompts, testing different settings Shift/CFGNorm/CFG/Steps/Samplers/Lighting/NoLightning, many different complicated workflows with control nets, face recognition, dozens of fancy custom nodes. You name it I've tried it.

Someone mentioned this LoRa on a reddit thread and I figured I might as well give it a try and for some reason this simple little LoRa just works. With a simple workflow and basic prompts it maintains the likeness of the character better than literally anything else I've tried. I'm actually flabbergasted at how good this works.

This thing is incredible. Great job xiaozhijason

大佬，请问我使用了提到的工作流和这个lora后，生成的图永远是第一张参考图，这是为什么，检查工作流是完全复刻的呀

hi brother it's really urgent i can't know why in some picture if i enable cons edit v2 with a lets say 0.3 weight it's work ( close shot) and when the girl or guy is far away it doesn't work i need to disable it. by the way if you can check your wechat i've added you im professional and doing business i need to have you quick, thx.🙏

Are you planning on doing a 2511 version?