Training Looks Like a Blob? (Pls Help) :(
Hi Friends! So I'm new to all this and maybe doing something wrong (been working with ChatGPT to try and figure things out) - it's helped me get this far. I managed to do sequential training on Kohya starting with 30 face images then 40 body images (all captioned with matching titles) and using the RevAnimate model and a VAE - then saved this as a .safetensors file. Then, I went to A1111 to generate images of her (using her .safetensors model) and see how it turned out with the hopes of generating images of her to use for content - BUT, my output looks like a yellow blob in a dark background. It makes no sense, I've fiddled with all the settings and it doesn't seem to be getting any better. What am I doing wrong? I'm desperate.
11 Replies
Unknown User•2mo ago
Message Not Public
Sign In & Join Server To View
I appreciate you getting back to me on this! My settings are all st to default besides these:
Sampling method: DPM++ 3M SDE
Sampling steps: 50
Hires fix: 800×1200
Hires steps: 25
Denoising strength: 0.3
CFG Scale: 7
SD VAE: vae-ft-mse-840000-ema-pruned.ckpt
Of course, I'm also using my saved .safetensors model using the images I trained, and my positive and negative prompt which includes my instance and class token that I trained with as well. Yet, when I hit generate, here is an example of the type of images I'm getting which are just the worst 😬
Unknown User•2mo ago
Message Not Public
Sign In & Join Server To View
more sampling steps or more hires steps? and i have no idea how to tell whether my trained model is good or not. like perhaps its overfit? I used a decent variety of angles and poses all captioned well so I don't know how to train any better than that or what (if anything) I did wrong. :(
Unknown User•2mo ago
Message Not Public
Sign In & Join Server To View
I used Kohya and the revAnimated_v2RebirthVAE model and set my training images/captions to the proper path. This saved to a .safetensors file which I am now using as my model to generate images in A1111
Unknown User•2mo ago
Message Not Public
Sign In & Join Server To View
yes i use kohya, jupyterlab, and a1111 all on the same pod but they're on diff ports - this was a template I got through runpod that was alrdy prebuilt
upped the sampling steps from 50 to 100 and it looks just as bad, if not worse - lol i have no clue what i did wrong and honestly about to just give up altogether on this project. 😆 so difficultttttt sheesh. thank you so much for your help though I rlly appreciate you helping me try to troubleshoot ♥
Unknown User•2mo ago
Message Not Public
Sign In & Join Server To View