Model Training ‐ Comparison ‐ [Checkpoint]
Models | Logs | Graphs | Configs
So, we can train the model on a different checkpoints. It is possible to train the model on one checkpoint and generate images on another checkpoint. So, you may think that if we use the same checkpoint for training and generation then it will lead to the most quaility results but this is not how it works, sadly. Ok, you may also think that training on the most basic checkpoint (Stable Diffusion 1.5) will give the best results but it doesn't work this way too. Anyway, we'll try to find the best checkpoint.
Compared values:
DLR(step)
Loss(epoch)
And graphs won't help us this time.
Instead of generating progression grids and grids with random prompts, I took 3 prompts and made grids with different models and checkpoints. These prompts matches appearance that is presented in the dataset so it's easier for the checkpoint to produce a good result.
Pink frame means that training checkpoint matches generation checkpoint.
It's pretty hard to compare everything at once so we'll divide checkpoints by groups.
Let's start with Stable Diffusion 1.5 models. By looking at the results you may think that this is a clear winner. But it never matches the character. It superficially captures facial features that makes you think everything is OK but it's not. We get completely different character that looks a little similar. If that's your goal then basic SD1.5 checkpoint is your choice.
Next is Realistic Vision checkpoints. The furthest generation we use, the less similarity we get. RV2.0 model is a clear winner.
Reliberate 2 model shows similar to RV2.0 results but has quality drops on some checkpoints.
RealCartoon3D model works good on even fewer checkpoints.
Photon is a little bit better but still worse than RV2.0.
Both EpicRealism models are worse too.
Well, everything else is the same. It either has quality drops on more checkpoints either misses character similarity.
To avoid comparing a million of results for GR = ∞ I took only the most promising or interesting checkpoints.
Stable Diffusion 1.5 models, once again, shows the most colorful and vivid results and is the most compatible with all the checkpoints. But it doesn't match the character.
Everything else is, once again, worse than RV2.0. But 28D28 model is quite promising.
Good results still requires a good prompt and a good amount of luck. But I think it's much way easier to get closer to a good and similar to the character result using models trained on Realistic Vision 2.0 checkpoint. Stable Diffusion 1.5 gives the most quality results but doesn't match the character appearance. You can try it for training if you are a lucky person and has been praying for gods.
