Hello everyone. I am Dr. Furkan Gözükara. PhD Computer Engineer. SECourses is a dedicated YouTube channel for the following topics : Tech, AI, News, Science, Robotics, Singularity, ComfyUI, SwarmUI, ML, Artificial Intelligence, Humanoid Robots, Wan 2.2, FLUX, Krea, Qwen Image, VLMs, Stable Diffusion
In the end, it turned out to be 180 images. I'm using 100 epochs, and right now it's at 50% of the training process at a speed of 6.6s per step. We'll see how the final result turns out
What would be the best way to train a model for interior design? Let me explain. I need to apply a specific product to a given room. For example, I want a Renaissance-style room with this product on the wall and this product on the floor. I understand that in this case, it makes sense to train a highly structured model with all the individual products and, at the same time, the same products applied to different rooms. I assume that captions are important in this scenario. What would be your approach to training a model for this application?
Hey @Furkan Gözükara SECourses, I'm enjoying the 5090 videos you've done and had a question. Wondering if you're planning on doing a video comparing 3090 vs 5090 for fine tuning or lora training? Sorry if asked elsewhere discord didn't show if it had or hadn't.
Hi! Don't know which channel to write this on. I'm training a lora and I'm wondering if I can stop the training and start a training on the latest save like you can when training checkpoint?
I have finetuned flux-dev with 180 images of a character and 180 photos from various angles. Maybe it's not the best dataset because it didn’t have varied backgrounds. Almost all the images were just of the face without the body. I'm having quite a few issues with prompts where I request a stylization. I trained a batch with 100 epochs and have 4 safetensors—I’ve tested them all, and I’m experiencing the same issues across the board. When I ask for a stylized result, like a cartoon or anime style, it struggles to adhere to the prompt and often generates random things instead. What could be happening? Thanks a lot!
you may also try different prompt engineering, generate a lot of images and cherrypick from them secondary dataset to include in traning of second gen model with more stilyzed images in training data
I did imagine that. But I watched @Furkan Gözükara SECourses tutorial and saw that he had some incredibly stylized images... I think he also only trained with realistic images. That's where my doubt comes from
The grid test I am analyzing to find newer FLUX DEV training. This is not even including the first 10 different trainings i made 1176 images. I think we gonna have a better workflow than what we have with using more VRAM and more training time.
Out of interest, what causes the VRAM per GPU requirements to jump up so much? Just curious on what's going on in the background processes that requires VRAM requirement to go from 24gb VRAM per GPU from single training to 80gb VRAM per GPU for muli-traning
How does multy GPU trainign work anyways? Each GPU has a copy of a trained model and perform traning steps independantly, so how in the end we are not gettign two separate models?