New to this. First time training a chroma Lora and it came out pretty good in the sampler (ai-toolkit), and I’ve been experimenting with a simple workflow based on the default Comfy chroma template, adding only my Lora to the graph (picrel). Results have been good but inconsistent. I’ve been using a fixed seed so that I have reproducibility between all experiments. The settings in picrel repeatably make a VERY good representation of the subject. Like so good I think I could use it to train with… HOWEVER:

>If I change only the seed, results are often inconsistent in both facial retention, body shape, and other details.

>Even with the “perfect seed” if I make what I think are minor changes to the prompt (eg to change her position etc) I get similarly inconsistent results.

>If I remove the inconsistent typo (“She's in a relaxed pose with her right arm on her hip. Her She”) I lose some facial retention and get weird arm artifacts.

I noticed during my training that by the final epoch some samples were spot on and some were meh. Does all this point to my LORA being shit? Should I go back and retrain until all 10 training samples are perfect? I’m using the default training prompts from ai-toolkit. Is my prompting shit? It’s just cobbled together crap I found.

Given the workflow I’ve made, if my LORA is actually good what consistency should I be expecting? What other variables are at play in terms of locking in a body shape? The body from the current settings is something I’d like to bake in somehow (re-training with the generated images?) I don’t know yet how to explain the wild inconsistency. Any advice appreciated, sorry I can’t post the actual pix or lora. Using a runpod L40S.