Z image turbo?
Do you think this idea could be used also for z image turbo? Thank you for your contribution.
This will probably work with Z-Image as well. However, once Z-Image-Omini is released, it will likely enable the same level of control as ControlNet.
Yeah, but I highly doubt they will ever release it, at this point... Can you please give me just a simple idea of the steps involved in the training process? What classes and elements inside the pipeline should I study to try to implement it for z image turbo? It doesn't touch specifically the weights, but the attention scores, is this right? Thanks for your patience.
Indeed, we don't know when Omni will be released... Your understanding is correct. ControlNet-LLLite is a completely add-on (monkey patch) network and does not modify the DiT core. See the following source code in sd-scripts:
https://github.com/kohya-ss/sd-scripts/blob/feat-anima-lllite/networks/control_net_lllite_anima.py
https://github.com/kohya-ss/sd-scripts/blob/feat-anima-lllite/anima_train_control_net_lllite.py
https://github.com/kohya-ss/sd-scripts/blob/feat-anima-lllite/anima_minimal_inference_control_net_lllite.py
Please note that the feat-lllite-anima branch is scheduled to be merged soon.
thank you very much!