Because there are versions 1 and 2, and this would be version 3, here we have two versions of this LoRa, one focused on face swap and the other on head swap. This would be the head swap version 3; I haven't made a face swap version 2 yet.

That's the goal; I've been working on it for over a month, training many, many versions. The difficult part is creating a perfect dataset, since few tools nowadays do that accurately. That's why I'm releasing versions gradually until I have a perfect dataset and perfect lora.

Yes, it's possible, I just need to assemble a perfect dataset for it. But you can try changing the CFG Norm because you probably use lighting, and that's why the colors change so much in the region that was changed. But I intend to do that. The big question is, who has a perfect dataset for this? I've been working on this LoRa for over a month and still haven't found a perfect dataset. If you use LoRa without lighting, the results will be much better. I posted two examples of this in the LoRa examples, with and without lighting.

@forfreelsd368 Yes, it was a mistake. I ended up updating the workflow after I saw it. That's what happens when you rush to upload the workflow before you have to go to market, haha!

@forfreelsd368 He probably won't follow the prompt because he was heavily conditioned by the caption I used for the command. What you could try is to decrease the strength of the LoRa or change the eye color later in a second pass. You could also do a second pass where, for example, in the first pass (0 to 2 steps) he uses the LoRa, and in the second (2 to 4 steps) he removes the LoRa. Then you can put another prompt and also input the input more times with your reference photo, because then the swap will have already occurred in the first 2 steps, I think.

@aungkhant0911 For version 3 you need to use workflow 3. Since these are tests, the idea is to try and improve the previous version, which always ends up changing some things in the workflow as well.

@forfreelsd368 I didn't understand

Face Swap V11 – This was the very first version of the workflow, created before I started training any LoRAs.

BFS – Workflow (Focus Head) Advanced – Based on the first version but now using a LoRA. It allows the use of masks and provides more control over blending.

Head Swap V2 Simple Workflow (With Lighting) – A simple workflow for the Head Swap V2 (Focus Head) LoRA, combined with the Lighting LoRA.

Head Swap V2 Simple Workflow (Without Lighting) – The same V2 (Focus Head) workflow, but this time without the Lighting LoRA (LightX2V acceleration).

Head Swap V3 Simple Workflow (With Lighting) – The simple workflow for Head Swap V3 (Focus Head), using the Lighting LoRA. In this version, the order of the input images has changed, and the prompt has also been updated.

Head Swap V3 Simple Workflow (Without Lighting) – The V3 (Focus Head) workflow without the Lighting LoRA. As with the “With Lighting” version, the input image order and prompt are different from previous versions.

@forfreelsd368 Ah yes, the layer style is not used in Simple Workflows.

I don't think it's ineffective; I used it to generate all the examples and it worked. It might be something you did wrong, or your input examples, or you might not be using the trigger I provided. Many things can cause qwen to not work. Besides that, you shouldn't use the V3 workflow in LoRa V2.

@luoye8888 That's strange; it must be a very specific case, because I haven't had that problem yet.

From what I gathered from one of the images you posted, you probably used lightx2v. Lightx2v makes everything worse, which is why the hair looks like that. It's not my Lora itself; everything you use for inference (models, Loras, sampler, scheduler) will alter the original result. Remember that when a model is trained, it's not trained with lightx2v, but in a pure way.

If you use a CFG Norm value below 1, it will remove the heavy coloring created by Lightx2v if you are using it, but this will alter the coloring. If you want the coloring closer to your reference image, set the value to 1.0.

If you notice in the example images I post, some have "CFG Norm" written at the top of the image.

Thank you so much, haha! It took me a long time to understand how this LoRa needed to be trained and to assemble the datasets I have today. I spent over a month just focusing on that and training many, many models. I spent a lot of money, haha, just to bring this here. In the end, I don't even end up using the models I train; I make them to help the community in some way, always focused on things that are more based on styles and concepts. But thank you very much for the feedback. Regarding the 10-minute per image time, try using the Lightning LoRa and exploring the schedulers and number of steps a bit to get a better result with it, even mixing other LoRas in the process. This is still unexplored for me with this LoRa.

@qek Honestly, the best way I can think of is for you to send a pose map, something like a DWPose or MediaFacePipe.

I think they trained these Loras without much concern; I don't think they're well-trained. The best they do is really speed up inference because the rest isn't worth it in terms of quality.

You can use whatever quantization you want, but the lower the precision, the worse the quality of the result can be.

@nikolaibloom805 q8 or original model, I need to run some tests to see what the best parameters are, which is why community experiments are so important, but almost no one posts their results, so it's difficult to improve without at least some community feedback.

The quality also depends a lot on the inputs you are providing; if you provide very poor angles of the reference image, the result tends to be worse.

@nikolaibloom805 But were you able to solve it or not? I didn't understand.

I think it's the quantization of your model; it's at q4, which is much lower than what I used for my examples. The lower the quantization, the worse the quality will be.

One possibility would be to try the LightX2V with that scheduler and sampler, because you reduce the number of steps, or you could really look for another one, there are many that don't have a second stage.

@noorhamadani Yes, that's the problem with light2xv. The solution is to try and find another scheduler/sampler. Usually, all those with 2s will do a "second pass," which is a second stage. You could try, if you're using res_2s, halving the number of steps to see if it works well. For example, instead of 20, try 10.

So I don't know if you read the post description with the workflows, but if you use version 3 of this LoRa, you won't be using the older workflow versions, but rather version v3.

Version 3 doesn't have OpenPose, but you can try adding it, but I don't know if it will work since LoRa wasn't trained with an OpenPose like Ctrl_img. You would have to try decreasing the LoRa strength and adding OpenPose if it doesn't work with strength 1, but this is using version 3 of workflows, which is a simple workflow.

Face Swap V11 – This was the very first version of the workflow, created before I started training any LoRAs.

BFS – Workflow (Focus Head) Advanced – Based on the first version but now using a LoRA. It allows the use of masks and provides more control over blending.

Head Swap V2 Simple Workflow (With Lighting) – A simple workflow for the Head Swap V2 (Focus Head) LoRA, combined with the Lighting LoRA.

Head Swap V2 Simple Workflow (Without Lighting) – The same V2 (Focus Head) workflow, but this time without the Lighting LoRA (LightX2V acceleration).

Head Swap V3 Simple Workflow (With Lighting) – The simple workflow for Head Swap V3 (Focus Head), using the Lighting LoRA. In this version, the order of the input images has changed, and the prompt has also been updated.

Head Swap V3 Simple Workflow (Without Lighting) – The V3 (Focus Head) workflow without the Lighting LoRA. As with the “With Lighting” version, the input image order and prompt are different from previous versions.

⚠️ Note: The Face Swap V11 workflow requires Torch 2.8.0 to function correctly, as it uses a custom node that depends on it. Some users have reported that updating to this Torch version may cause ComfyUI to break in certain setups. The same issue may also occur with the BFS – Workflow (Focus Head) Advanced, depending on your configuration. Proceed with caution when updating.

It's individual to each person; qwen is specialized in editing, the others aren't, aside from the problems that sdxl can generate, which rarely happen with qwen. The issue of slowness, etc., is also individual to each person; if you don't have a good enough GPU, you'll have to stick with sdxl.

It could probably be that the resolution is too high; try generating in 1024x1024. Or it could be the CFG; if you're using a CFG higher than 1 and at the same time using a LoRa of Lightning, it could be many things.

You can try creating a mask around the face you want, cropping the box that surrounds that mask, and using the image from the box to do the swap. You can use this node here which makes everything easier: https://github.com/lquesada/ComfyUI-Inpaint-CropAndStitch

@LesterBangs13 Just install this node with the manager, then create a mask using ComfyUI's own load image. If you right-click on the image in the load image, you can open the canvas and create the mask. Then, pass the mask to the ✂️ Inpaint Crop (Improved) node, and it will crop it for you. You can then use this image to perform the swap. After the swap, use ✂️ Inpaint Stitch (Improved) to return the cropped image to its original state.

All the models I train are created with the purpose of testing things. Training is not simply taking datasets and throwing them into a tool and that’s it. I experiment with different approaches. That’s why in V3 you first provide the body image and then the face image — it’s not about difficulty. The real difficulty is training; as a user, you only need to use what someone else has already trained.

Everything is written in the information provided. There is even a workflow made specifically for V3 so you don’t run into this issue.

Another thing: there are two models here — you just need to read. One is for face swap, which is the Focus Face, and the other is for head swap, which is the Focus Head. One replaces only the face, the other replaces the entire head.

Depending on which models you are using with 16GB of VRAM, you may be experiencing VRAM shortage issues or something similar.

So, in the dataset I got from Face Swap, there weren't that many samples with a perfect expression match. That would be the SOTA for face/head swapping, which is why it ends up not following your additional instruction to follow the expression. What you could do is a second pass after the swap to try to adjust the facial expression.

@newbieAiUser yes, without LoRA or with low intensity