To achieve the desired results with this LoRa, it’s important to review my instructions before usage. Thank you very much for your patient reading.
Important matters stated three times: Do not directly overlay ID-photo LoRa and character LoRa in the prompt section. Do not directly overlay ID-photo LoRa and character LoRa in the prompt section. Do not directly overlay ID-photo LoRa and character LoRa in the prompt section.
The optimal approach for using ID-photo LoRa is as follows: Begin by using weights to apply the adetailer plugin for facial replacement.
(1) You already have a facial LoRa with relatively high similarity, but do not place it in the prompt section. The LoRa in the prompt section should only feature the ID-photo, with a weight of 1 (ensure a high weight setting). (2) Utilize majicmixrealistic v6 as the base model, using DPM++ 2M SDE Karras sampling method, with 30 or more steps, CFGscale7. Remember to activate high-definition repair for generating a good base image for the ID-photo. (3) For frontal prompts: identification photo, 1 girl, long hair, bangs, looking at the viewer, (white shirt, blue tie, red background: 1.3), happy, smile. Negative prompts can be self-created. Background color can be changed to white/blue/grey. (4) Open the adetailer plugin, choose the face model, and input your facial LoRa into the Adetailer prompt in the format , with a recommended weight of 0.75 or higher.
Fine-Tuning Optimization Plan for ID Photo Lora:
(1) Distinguishing Background and Clothing Colors: If you encounter situations where the background and the person’s clothing are always of the same color, it is recommended to use the anti-staining tool, the ‘cutoff’ plugin. For example, if I want to achieve a photo with “white shirt, blue tie, red background,” input (white shirt, blue tie, red background) in the Target tokens and set the weight to 2.
(2) Adjusting Facial Shape of Subject against Background: If you find that after using ‘adetailer’ for face swapping, the ‘facial features match but not the face shape,’ I recommend you download the ‘Muggle Lora’ facial shape adjustment tool from https://www.esheep.com/model/1569. This Lora tool won’t interfere with the composition of the image. It can be combined with the ID photo Lora and added to the frontal prompts. The weight range is from -1.5 (oval face) to 1.5 (round face).
(3) Generating a Female Image from a Male Subject: I can clearly tell you that this is a common issue with SD (Style-DALI). You can try using a base model that is more skilled at generating male subjects. This is not just a Lora-specific problem; it’s widely known that SD’s training data is skewed towards females. When generating a male base image, remember to use relevant tags like (1 boy, male focus: 1.3) and fine-tune the weights.
What to Do If You Only Have One Photo? Please use Photoshop or apps like Meitu Xiuxiu; avoid using SD. The original intention of using Lora is not for simple face swapping with a single image. If you insist on using SD’s ROOP plugin, facial similarity cannot be guaranteed due to issues with the ROOP model, not Lora’s model.”
Visit Official Website