This model accurately captures Lae'zel from Baldur's Gate 3, trained on 35 carefully curated, long-captioned images including screenshots and renders for versatility.

Launching v2.0 - LoRa

Now available as a Rank 8 LoRa trained with an improved dataset and hyperparameters for better likeness and prompt adherence over the v1.0 LoKr version but with similar or slightly worse compatibility with other LoRas. The v2.0 LoKr version is in development and should improve even further.

⚠️ Important: If using the v1.0 LoKr model, please read the "ComfyUI Setup" section below before use.

Usage

Trigger Word: l43z3l elf

Important:

1. Do not include "woman" for likeness, like l43z3l elf woman

2. Include "she" or "her" in your prompt to ensure female generation, otherwise the model may generate a male Lae'zel spinoff.

Usage Tips

• Start with weight 1.0 and adjust as needed

• The more mundane or realistic the prompt is, the more likely you will need to increase the strength to 1.1 or even 1.2.

• For baldur's gate like prompts (e.g. mentioning armor) you might need to decrease the strength to 0.9 or even 0.8 in order to avoid a too strong 3d render like output

• When combining with other models, first consider increasing this model's strength up to 1.4 if the likeness isn't quite there yet. If still not quite there, try then lowering the other loras strength.

• This is because the model has a weaker interaction vs most other loras, depending on how realistic and/or overfitted they are.

ComfyUI Setup for LoKr Model (v1.0 Only)

Important: The following setup instructions are only required for the v1.0 LoKr model. The v2.0 LoRa model works out of the box with ComfyUI.

For some reason, LoKr models for Wan 2.1 are not yet supported in ComfyUI and you need to do a small code change. This change should not impact any other LoRas or LoKr models you use for Wan or other base models.

Installation Steps:

1. Navigate to your ComfyUI folder: comfy/lora.py

2. Find the model_lora_keys_unet function

3. Replace this section:

for k in sdk:
    if k.startswith("diffusion_model."):
        if k.endswith(".weight"):
            key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
            key_map["lora_unet_{}".format(key_lora)] = k
            key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
        else:
            key_map["{}".format(k)] = k #generic lora format for not .weight without any weird key names

With this:

for k in sdk:
    if k.startswith("diffusion_model."):
        if k.endswith(".weight"):
            key_lora = k[len("diffusion_model."):-len(".weight")].replace("attn.o", "attn.to.out.0").replace("ffn.0", "ffn_net_0_proj").replace("ffn.2", "ffn_net_2").replace(".", "_")
            key_map["lora_unet_{}".format(key_lora)] = k
            key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
            key_map["lycoris_{}".format(key_lora)] = k #SimpleTuner lycoris format
        else:
            key_map["{}".format(k)] = k #generic lora format for not .weight without any weird key names

Note: This modification should not impact using other LoRa or LoKr models for any other base models.

Other UIs: Setup instructions currently only available for ComfyUI. Please share your experiences with other interfaces in the comments.

Feedback

Share your results and report any issues in the comments.

❤️ Last but not least, if you enjoy this model, please give it a like and share your generations in the comments! Your support helps other creators discover it.