guaneec
guaneec
I think the constant is defined in the [model config file from CompVis/stable-diffusion](https://github.com/CompVis/stable-diffusion/blob/main/configs/stable-diffusion/v1-inference.yaml#L17).
The main difference is where the blending occurs - this mixes the text encoder outputs (77x768), and the new one mixes denoiser outputs (4xH/8xW/8). When I last tested, both approaches...
Also [alternating words](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#alternating-words) is a thing, which has a similar effect.
May I suggest that you also add a standard error indicator? From my limited tests, the loss coming from the same single image can vary wildly even when the prompt...
Fixed. Also rolled back dropout off-by-one fix because that messes up layer names of old nets.
Added missing dropout layer fix by @aria1th
Hmm, first negative report after the change. How many HNs have you trained? Are you sure it's not just an unlucky run? This PR actually makes 3 changes: 1. Remove...
I could make it one. I didn't do it originally because the code change is quite light for now, but I guess that could change as features get added.
LoRA in general is a reparameterization a matrix. [cloneofsimo/lora](https://github.com/cloneofsimo/lora) applies it on Pivot Tuning, which finds the TI and then finetunes the UNet and the text encoder. Low-rank decomposition can...
I'm continuing the development of this as [an extension](https://github.com/guaneec/custom-diffusion-webui).