Somehow got ZiB LoRAs applied to ZiT in 1.121, cannot reperform and wonder what it was.

[alex-ie]

I've been playing with new SDUI of 1.121, tried to select LoRA claimed to be for ZiT on steps=3, got different picture. Restarted with LoRAs for ZiB folder (ZiT model gguf) and got different pictures again for couple of LoRAs (same seed, prompt, steps, dimensions - all). Then after restart it stopped working - very similar picture (thought different sha hash). Tried with ZiB + ZiB LoRA - different picture -OK.

Now left with couple of images and json files saying they are ZiT+ZIB_LoRA and wonder how they got produced. Any ideas?

[wbruna]

A few things to keep in mind with the new UI:

• it works like the prompt syntax: the LoRAs there are a list of requests, not necessarily what will end up into the final image. They also get stored as the prompt syntax in the image gallery
• if you switch server models or LoRA directories without refreshing the UI, the available LoRA list isn't necessarily refreshed (in the same way as the prompt isn't refreshed). So it could send requests for LoRAs that aren't available anymore
• the list of LoRAs that actually got embedded on the image is sent back by the server, and registered separately on the gallery.

As an example:

• I've made an image in Krea2 with the krea2_turbo_4step_rank_64_lora LoRA;
• Then I've switched the server to Z-Image Turbo, but kept the Krea2 LoRA selected;
• Then I've selected the Turbo_Booster_v1-fp16 LoRA, and made another image.

The result:

We see:

• the prompt flower <lora:krea2_turbo_4step_rank_64_lora.safetensors:1> <lora:Turbo_Booster_v1-fp16:0.14> with the requested LoRAs (including the nonsensic Krea2 LoRA for a ZiT model)
• the model and LoRA list below reporting what actually went into the image: the model z_image_turbo-Q8_0 and the LoRA <lora:Turbo_Booster_v1-fp16:0.14> (the Krea2 LoRA was discarded)

So if you want to reproduce an image, follow the model and LoRA data below the prompt.

Also, keep in mind ZiT and ZiB have the same architecture, so LoRAs are interchangeable; but they don't necessarily give the same results between the models.