← sources

How to create a high-quality dataset to train a LoRA with limited references?

source: reddit (reddit:comfyui)published 2026-10-29 16:13ingested 2026-10-30 02:40

retained text

I have a tilted facial portrait, a full-body reference, a detail reference showing the quality I'm aiming for, and an example of what the final dataset should look like.

I know my reference images aren't ideal for dataset creation. I'm looking for a way to work around their limitations while maintaining facial identity, body consistency, at a high level of detail.

I've tried dataset a generation workflow, but it tends to reproduce the face's original angle rather than create useful variations. The output also has an obvious AI-generated look compared to my reference.

What's the best workflow for this? Would face swapping be the best approach, or are there better alternatives?

I'd really appreciate someone willing to walk me through the process in detail so I can understand what I'm doing and why, as I'm still learning.

Thanks in advance to everyone who takes the time to help!

document fetched 2026-10-30 00:44 · canonical https://reddit.com/r/comfyui/comments/1xiocm5/how_to_create_a_highquality_dataset_to_train_a