Software Engineering Courses (SECourses)•15mo ago

https://x.com/kohya_tech/status/1847581169933832521

Kohya Tech (@kohya_tech) on X

Differential Output Preservation(仮)、FLUX.1で学習してみた。prior_loss_weightは10。一応効果ありそうだけど、diffがだいぶ学習不足なだけかもしれない……。

Twitter

•

10/19/24, 10:10 AM

FFurkan Gözükara SECourses https://x.com/kohya_tech/status/1847581169933832521

Maxivi•10/19/24, 10:33 AM

no we have to wait for bmaltais to merge to gui

MMaxivi no we have to wait for bmaltais to merge to gui

Furkan Gözükara SECoursesOP•10/19/24, 10:37 AM

yep

Furkan Gözükara SECoursesOP•10/19/24, 10:38 AM

the videos waiting me to edit :/

Furkan Gözükara SECoursesOP•10/19/24, 10:38 AM

FFurkan Gözükara SECourses yep

Maxivi•10/19/24, 11:00 AM

kohya updated

Maxivi•10/19/24, 11:00 AM

now we wait for bmalt

snowmonkey•10/19/24, 11:51 AM

I'm looking for ComfyUI workflow with Flux Schnell. Anyone got ?

Ssnowmonkey I'm looking for ComfyUI workflow with Flux Schnell. Anyone got ?

druhl•10/19/24, 2:37 PM

https://comfyanonymous.github.io/ComfyUI_examples/flux/

ComfyUI_examples

Flux Examples

Examples of ComfyUI workflows

Ddruhl https://comfyanonymous.github.io/ComfyUI_examples/flux/

snowmonkey•10/19/24, 4:33 PM

Thanks!

snowmonkey•10/19/24, 5:06 PM

Does anyone know a good tutorial to use runpod-worker-comfy ? I just want to deploy an API to generate images https://github.com/blib-la/runpod-worker-comfy

GitHub

GitHub - blib-la/runpod-worker-comfy: ComfyUI as a serverless API o...

ComfyUI as a serverless API on RunPod. Contribute to blib-la/runpod-worker-comfy development by creating an account on GitHub.

dsienra•10/19/24, 5:53 PM

Update https://huggingface.co/nyanko7/flux-dev-de-distill Lora training, removing the class did not help with class bleeding, the result was almost the same, and maybe worse because difficulted the prompting on inference. I'm doing another test with just one class and some regularization images, idk if kohya new code was implemented yet, hope this will help too.

nyanko7/flux-dev-de-distill · Hugging Face

Pelolisu•10/19/24, 6:37 PM

New technique I'm introducing for identity preservation.

PPelolisu New technique I'm introducing for identity preservation.

Furkan Gözükara SECoursesOP•10/19/24, 6:43 PM

wow this is good

PPelolisu New technique I'm introducing for identity preservation.

art-ai_usa-girl•10/19/24, 7:35 PM

more information please

Aart-ai_usa-girl more information please

Pelolisu•10/19/24, 7:52 PM

this feature is part of new webui in development, qwen-flux-wan-such This feature is part of a new webui under development, when finished, Dr. will make a tutorial on how to use it. For now I'm in the final stages and presenting some previews here. This feature only uses the image-to-image diffusion pipeline and a controlnet to refine.

FFurkan Gözükara SECourses in the upcoming video i have explained

tazztone•10/19/24, 8:43 PM

i am hyped

Furkan Gözükara SECoursesOP•10/19/24, 9:02 PM

editing video

Furkan Gözükara SECoursesOP•10/19/24, 10:06 PM

so much more to edit

FFurkan Gözükara SECourses so much more to edit 😄

Pelolisu•10/19/24, 10:19 PM

The Neverending history haha

PPelolisu The Neverending history haha

Furkan Gözükara SECoursesOP•10/19/24, 10:21 PM

FFurkan Gözükara SECourses well flux has internal captioning mechanism

Nick44•10/20/24, 5:01 AM

whats the best way to character flux caption? replicate / use joy captioning or simply only use an instance prompt + class prompt ?

Ddsienra Update https://huggingface.co/nyanko7/flux-dev-de-distill Lora training, removin...

druhl•10/20/24, 5:06 AM

My experience of training with de_distill:

1) It adds a sharpness to the image. The images are more noisier, better realism, and harsh reality kind. TBH, that is not always a good thing. If you like or do not like an image, it is subjective>
2) The ability to modify CFG can give you vastly varying results for the same seed. Its CFG scale is much, much tamer than dev2pro and produces linear effects (as you would expect) if you increase or decrease it.
3) The additional time it adds to inference is very frustrating. Original dev said 60+ steps, and he is right. I got good results at Step 70. You can get the generation much faster on flux_dev with 25 to 42 steps. Adding steps adds time to an already slower generation speed.
4) You need to use extra settings during generations like thresholding, which adds additional complexity to the traditional step-cfg system.

druhl•10/20/24, 5:08 AM

Dreambooth on flux_dev:

druhl•10/20/24, 5:08 AM

Dreambooth on flux_dev_dedistill:

NNick44 whats the best way to character flux caption? replicate / use joy captioning or ...

Nick44•10/20/24, 6:19 AM

also lets say im training to make a consistent yet diverse race (ie different robots as opposed to a consistent singular character robot like Optimus Prime) for a flux lora / DB . its best to only use the instance prompt: "ohwx" + no captions + instance : "robot"

ar.1•10/20/24, 9:52 AM

what do i fill for class prompt and instance prompt for finetuning on photography images for realism style, not like a single character

Vinch•10/20/24, 10:04 AM

We all know from Dr. Furkan Gözükara's work that it typically takes 15 images, 3000 steps, and a learning rate of 4e-6 to train a model to capture a face. However, training something as complex as an alternate futuristic reality is an entirely different challenge. The dataset size, steps, and learning rate required for this are much more demanding.

In my case, with a dataset of 500 images for stage 1 and 1000 images for stage 2, I initially started with a low learning rate (4e-6), fearing overfitting. However, after running into issues like underfitting and wasting valuable resources, I learned an important lesson: starting with a high learning rate (like 4e-5) allows you to capture the core of the concept much faster, even if it overfits. After that, you can restart with a lower learning rate to fine-tune the details.

In fact, starting with a low learning rate to avoid overfitting ended up costing me both time and money. It takes longer to realize you're underfitting, and by the time you do, you've already invested significant resources. Allowing the model to overfit initially and then restarting with a lower LR is much more cost-effective and efficient.

That said, I’m still open to learning if this insight can be further optimized. I tend to favor restarting rather than fine-tuning because of my perfectionist tendencies, but there may be a balance that I’m missing. For now, the strategy of beginning with a high LR and adjusting downwards has proven to be the most effective approach for me.

VVinch We all know from **Dr. Furkan Gözükara's work** that it typically takes 15 image...

druhl•10/20/24, 10:45 AM

Makes sense.

druhl•10/20/24, 10:46 AM

With overfit, you have a bad product. But at least you have a product.

NNick44 whats the best way to character flux caption? replicate / use joy captioning or ...

Furkan Gözükara SECoursesOP•10/20/24, 11:23 AM

prompt + class prompt

NNick44 also lets say im training to make a consistent yet diverse race (ie different ro...

Furkan Gözükara SECoursesOP•10/20/24, 11:24 AM

yes but try to find a class prompt that will match better

Aar.1 what do i fill for class prompt and instance prompt for finetuning on photograph...

Furkan Gözükara SECoursesOP•10/20/24, 11:25 AM

you can test different class prompts on dev model and see which one generates more resembaling images to yours

VVinch We all know from **Dr. Furkan Gözükara's work** that it typically takes 15 image...

Furkan Gözükara SECoursesOP•10/20/24, 11:26 AM

it is true every dataset is different, but i tested my LR on both 15 and 256 images worked very good. are you doing fine tuning or lora? your LR difference is huge 10 times

FFurkan Gözükara SECourses you can test different class prompts on dev model and see which one generates mo...