Is it able to control the special item (or at least 90% similar one) exists in output result?

[godlockin]

I prepare hundreds of test images on an item (export from a 3D model in different angle and direction) and a special token for 'sks', after hundreds of epochs training on stable-diffusion-v1.5, the output generated by model may only 30~40% similar with the "special" item.

e.g. i prepared the photos for a table, white, half height, but the output ones are brown one, black one, human height, too large, too small ...

Is it able to ensure the result is quite similar with the original ones I injected? or it's totally a false proposition to do this attempt?

If yes, what I can do next in professional perspective?

1. to pick other model as base model for further attempt?
2. prepare much more feed images up-to thousand level?
3. more steps/epoch with less learn rate? Or no, try other technology e.g. face swap .etc ?