Quite similar to this.
Assume there are five classes (flower object), i.e. [daisy
, dandelion
, roses
, sunflowers
, tulips
]) and using dreambooth technique, would it be possible to generate new view from each of this object with their corresponding prompt (with a single model, not re-training on each object separately)…
a photo of daisy
a photo of dandelion
a photo of roses
a photo of sunflowers
a photo of tulips
And training with
--instance_data_dir="$instance_daisy, $instance_dandelion, ..."
--class_prompt="$prompt_daisy, $prompt_dandelion ..."\
It looks very usual cases, have anyone tried so far like this in dreambooth? Closes thing I’ve found is this Research project multi subject dreambooth by klopsahlong · Pull Request #1948 · huggingface/diffusers · GitHub but not optimal.