A model trained on 146 images of both the best Ai generated images created with the V1 model and more images from both Avatar movies. All images were handpicked and further enhanced to choose the best and sharpest images for the training. You may use any VAE of your choice. Default is no VAE. Add Your Face: To add yourself you can easily train your face with Dreambooth using either the V1.5, V1 or V1-Alt as the base model. Alternatively you can try doing inpainting or img2img using your images, tho the results may vary. Differences: V1.5 was trained using a different model apart from the base SD1.5 for better and more coherency between the images. V1 and V1-Alt were trained using the base SD1.5 model. Overall V1.5 creates better and more creative images than both those models but in some cases the image quality for V1 is more varied, mainly on the body stripes. Note: Currently this model still doesn't have a tail. The ai has issues learning it. By V2 i will make sure the Ai is capable of creating the tail. Example prompt: Positive: <code>`Avatar Style, masterpiece, best quality, ultra-detailed, cartoon style woman, cute art style</code> Negative: <code>`weird face, deformed body, deformed hands and fingers, (saturation, colors, multiple people, more than five fingers, multiple faces:1.2)`</code> Euler_a: <code>20 steps</code> cfg scale: <code>4.5</code> Seed: <code>883802183</code> Dimensions: <code>512x512</code> The model works good even at higher res like 768 or 896. Alt Negative: <code>(multiple people, more than five fingers, multiple faces:1.2) deformed hands and fingers, cropped, saturation</code> Recommended to use 20 - 35 steps and HIGHLY recommend to use cfg scale 3.5 - 4.5 Also my ETA noise seed delta is blank. My images were created with no ETA noise, as default the UI makes it 31337 in the Sampler Parameters, if you can't replicate the images of the examples is prob your ETA noise delta or im using ControlNet with ClipVison for creating a different art style. Usage: You can do <code>boy, girl, old man, old woman, man, woman, teen boy</code> Best to use Clip Skip 1, but you can also do Clip 2. Keyword: <code>Avatar Style</code> or <code>AVTR</code> Note: The model was trained for <code>AVTR</code> as the keyword but it can work with <code>Avatar Style</code> it's best you alternate between them and see which one you like best. Upscaling: For my upscaling method here is a video of me generating images, upscaling, fixing the eyes and using ControlNet to have different art styles. <a target="_blank" rel="ugc" href="https://drive.google.com/file/d/1GYq4eHRrCc-jZNM90LOUj1Qf09F2aCln/view?usp=sharing">https://drive.google.com/file/d/1GYq4eHRrCc-jZNM90LOUj1Qf09F2aCln/view?usp=sharing</a> You may need these: ControlNet with Tile resample model <a target="_blank" rel="ugc" href="https://github.com/Mikubill/sd-webui-controlnet.git">https://github.com/Mikubill/sd-webui-controlnet.git</a> <a target="_blank" rel="ugc" href="https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11f1e_sd15_tile.pth">https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11f1e_sd15_tile.pth</a> UltimateUpscale for Automatic1111 <a target="_blank" rel="ugc" href="https://github.com/Coyote-A/ultimate-upscale-for-automatic1111.git">https://github.com/Coyote-A/ultimate-upscale-for-automatic1111.git</a> The ESRGAN model i use <a target="_blank" rel="ugc" href="https://drive.google.com/file/d/1GL0OjGPsMwSmCr_FmkqtRjZu2IM0s019/view">https://drive.google.com/file/d/1GL0OjGPsMwSmCr_FmkqtRjZu2IM0s019/view</a> Last minute info: I'm still working on the Metkayian model but it's giving me a few issues, but so far i have been able to create some amazing images with it. No release date yet tho. Foreseeable Future: "Why so Blue" NOTE: This model can produce NSFW content if you so wish but this model was trained for mainly SFW in mind. Q: Will you do a Safetensors version? A: No. The main reason is because if i did then people would not be able to train with it using Dreambooth because it is not a CKPT file. The second reason, i don't know how and the Dreambooth version i use can't export a safetensors version (i'm using an old Dreambooth version to train). Besides if you wanted to convert it back to CKPT to train with it, it loses quality from my testing on other models. I understand it loads faster and it's safer, but when it comes to security all my models are malware free. I have no interest and intention of uploading any malware or any use for that matter. I get the many concerns but if you are curious or still doubtful you can do a pickle scan yourself.

Avatar Style

Great overall, no particular preference.

avatarStyle_v1.ckpt

avatarStyle_v1.yaml

V1.5 Inpainting

V1.5

V0.4

V1-Alt

No particular improvement over V1 just an alternative, tho V1 might produce better full bodies in some cases.

avatarStyle_v1Alt.yaml

avatarStyle_v1Alt.ckpt

train settings included , trained using lora unet 0+ text encoder 0.5for 10k stepsthen lora off and text encoder off with a 0.000002 learning rate202 pic datasetIf you enjoyed the model I provided, I would greatly appreciate your consideration in making a donation to support my college education.<a target="_blank" rel="ugc" href="https://www.patreon.com/user?u=87243198">https://www.patreon.com/user?u=87243198</a>

Files with SHA256: fa47cd954edea66bd31d26066e6196d97a98d20b87fc1fd7eb344b6fc295a56b