sd-scripts
sd-scripts copied to clipboard
kohya-ss
Fix `Unexpected key(s) in state_dict: "text_model.embeddings.position_ids".` problem aroused by loading **text_model**.
AuraFlow is a new completely open source text2img model: https://huggingface.co/fal/AuraFlow It would be great when we could use kohya-ss to also train LoRAs and finetunes of AuraFlow.
There is a new paper, "u-μP: The Unit-Scaled Maximal Update Parametrization" at https://arxiv.org/abs/2407.17465 that promises selection of optimal hyperparameters, like the learning rate, even when using very small datatypes like...
When training batch size 4 on H100 the speed is 1.27 second / it When training batch size 4 on 2x H100 the speed is 2.05 second / it So...
Hi, I really, really need weighted captions and I see this in sdxl_train_network: ``` # Get the text embedding for conditioning # TODO support weighted captions # if args.weighted_captions: #...
I recently merged 2 finetuned models, on both which I am able to train LoRas (with Kohya for example). But it's not possible to train LoRas on merged model. My...
When training with a fixed seed, samples are different every time due to this block in `train_util.sample_image_inference()`, specifically the `else` branch: https://github.com/kohya-ss/sd-scripts/blob/25f961bc779bc79aef440813e3e8e92244ac5739/library/train_util.py#L5264-L5270 I understand that this was added so each...
I was wondering why the sampling during training produced such bad results and after many hours of debugging, I found the issue. The `clip_sample` value in the scheduler configuration [must...
