I started creating Loras locally, so needed a dataset creator.
Uses Klein 9_B with any good facial image to create 20 different images.
也有中文说明
Operates fairly well on Low VRAM
run 2 or 3 times on a random seed and select the good ones.
Instagram: https://www.instagram.com/synth.studio.models/
Buy me a☕ https://ko-fi.com/lonecatone
This represents many of hours of work. If you enjoy it, please 👍like, 💬 comment , and feel free to ⚡tip 😉
Description
Modified for Klein_9b
Fixed save node broken by ComfyUI
FAQ
Comments (23)
amazing. Thank you so much!.
1 question: What is the best image resolution to use as input?
2 question: Is it necessary to generate NSFW for the dataset (for training Lora) if it is planned to be used in the future? If necessary, could you please suggest which lora can be used for this purpose so that the face remains unchanged?
3 question: Have you considered using a pre-generated body as a reference for the entire dataset to preserve the body? I saw this in one of the workflows.
The face remains unchanged without a LoRA with this workflow. That is the concept.
In my experience, the fastest way to mess up a dataset and get frustrated is trying to train a body. Train the face, then prompt accordingly. I hear constantly that there is no consistency. If you must train, then train naked with teh body in different positions, but solo or POV.
I always train at different resolutions between 1028 and 2096 with relatively similar images in three different resolutions so the model knopws what to do with the space. Never lock your self in to one resolution. have you seen the streched out bodies? That's why.
@lonecatone23 thank you!
@lonecatone23 I also think it would be a good idea to automatically save prompts in txt files under generated images.
@cheyokey576 It save in the metadata automatically
@lonecatone23 By the way, in my case, photos without Consistence Edit Lora preserve the face much better. Using Lora also seems to make the quality of the photos worse.
@cheyokey576 Noted
You have to check WF for when GGUF model is used Widget To String at GGUF loader, it causes error when saving at the very end dataset is not saved only previews remain visible. With diffusion model loader it works fine.
that's inyteresting. There is no gguf name attached to flux 9b models 🤔.
I fixed it. Thanks
@lonecatone23 Widget To String node next to GGUF loader would become red and wf would crash, data set was finished and visible but would not be saved in Output, only happened with GGUF model used . If it is fixed I will download it again and this was the error (I run latest comfy all updated to stable version):
NameError: Widget not found: 246.gguf_name File "C:\ComfyNEWBEgining\ComfyUI\execution.py", line 525, in execute output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\ComfyNEWBEgining\ComfyUI\execution.py", line 334, in get_output_data return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "C:\ComfyNEWBEgining\ComfyUI\execution.py", line 308, in asyncmap_node_over_list await process_inputs(input_dict, i) File "C:\ComfyNEWBEgining\ComfyUI\execution.py", line 296, in process_inputs result = f(**inputs) File "C:\ComfyNEWBEgining\ComfyUI\custom_nodes\ComfyUI-KJNodes\nodes\nodes.py", line 990, in get_widget_value raise NameError(f"Widget not found: {node_id}.{name}")
@erdelman73267 fixed already.
@lonecatone23 thanks, doing test run now.
This works great. Does anyone know if there is the same but for ZImage Turbo? We can use this workflow but does a "consist" lora exist for ZIT too?
No. Z-Image does not do I2I like that. Also, you don't need the consist LoRA.
Hi i am getting a lot of grain in my generations, what do you recommend?
Try turning off image consist LoRA. and increasing steps. Unfortunately, that Klein_9b
Leaving everything the way you uploaded the workflow comes out a complete mess. out of the grid maybe 1 image is decent
Then use a better image and change the settings to suite you. I've had zero issues with it.
@lonecatone23 there is nothing wrong with my image. and you dont have any instructions what to change
@2018cfh are you kidding me? I have instructions in two languages and this workflow is free!!!
If you do not understand the basics, you can walk thorugh one of my guides on my page.
nevermind. your "instructions" are what models to use . load an image and run multiple times. thx for answering anyway.
does it have captioner?
No, but you can plug it in where the prompt goes