{"id":2698,"date":"2025-02-15T14:21:22","date_gmt":"2025-02-15T19:21:22","guid":{"rendered":"http:\/\/blog.veronika-zemanova.info\/?p=2698"},"modified":"2025-07-25T21:14:09","modified_gmt":"2025-07-26T01:14:09","slug":"creating-veronika-ai-images-using-tensorart","status":"publish","type":"post","link":"https:\/\/blog.veronika-zemanova.info\/?p=2698","title":{"rendered":"Creating Veronika AI Images using TensorArt"},"content":{"rendered":"\n

The Veronika Models have been uploaded to TensorArt for now.\nThey might be uploaded to CivitAI eventually. The main reason for using\nTensorArt is that they offer free image generation for registered users. Once\nregistered you will receive enough daily credits to allow you to generate a\ndozen or so images per day. CivitAI does not have a similar program from what I\ncan tell. Of course, TensorArt does offer subscription options to allow you to\ncreate a virtually unlimited, higher resolution set of images.<\/p>\n\n\n\n

Here is the URL for TensorArt.<\/p>\n\n\n\n

https:\/\/tensor.art\/<\/a><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Click the arrow down next to the ‘Create’ button and then the Creation<\/strong> button.<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

In the left tab which will pop up, select “Text2img<\/strong>“:<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

When you run the Creation process for the first time, it will default to the Stable Diffusion (SDXL) model. The Veronika models are based on Flux and not Stable Diffusion models. You can read up on the differences here: https:\/\/getimg.ai\/blog\/flux-1-vs-stable-diffusion-ai-text-to-image-models-comparison<\/a><\/p>\n\n\n\n

The Model needs to be changed to Flux. Do this by selecting “Basic Model – XL”. This will bring up the Model selection page:<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

Pick Basic model “Flux.1 – dev-fp8<\/strong>” from\nthe list. This was the model used to create the LoRAs, so must be selected to\nget an accurate rendering. After that is selected, you will see that the Basic\nModel has been updated to Flux. <\/p>\n\n\n\n

The next step is to add the appropriate LoRA you want to\nuse. Simply select the ?Add LoRA? button.<\/p>\n\n\n

\n
\"\"<\/figure>\n<\/div>\n\n\n

For those unfamiliar with some of the AI generation tools, LoRAs (Low-Rank Adaptations) are extensively used in text-to-image models to generate images with specific elements like: distinct characters, unique styles, or particular objects. It essentially steers the image creation into a desired direction based on the attributes trained into the LoRA. In creating the Veronika LoRAs I wanted to be as accurate as possible in providing the initial training data (ie. images) to create them. For the LoRA to accurately represent a real person, you want to provide the training with as many good quality images from as many different angles of the person so that when an image is subsequently generated using the resulting LoRA, it is able to draw on the learned images for reference and not fill in missing gaps with predicted pixels when creating a new image. I happen to have quite a sizable catalogue of unreleased images of Veronika in my private collection which I was able to draw on to put together what I thought were the best images to use to train a Veronika LoRA model. <\/p>\n\n\n\n

After selecting the Add LoRA<\/strong> button, you will have to\nperform a search to find the Veronika LoRA you want to use. At the top search\nbar simply enter the text Zemanova.<\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Once you key in Zemanova you will see all the available\nLoRAs to use. I would highly recommend Starring <\/strong>the ones you plan to use\nfrequently as this will make for faster retrieval in the future. Then you\nsimply need to hit the My Starred<\/strong> button the next time you want to load\na new LoRA.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

There are 3 LoRAs which have been created which are very\ngood to use to generate realistic photos of Veronika. They are all similar with\nsome slight differences between them: Zemanova V2<\/strong>, Zemanova V3<\/strong>\nand ZEMANOVA Fluxed<\/strong>. I would recommend using either the Zemanova V2<\/strong>\nor Zemanova V3<\/strong> LoRAs for the absolute best results. Here’s a comparison\nof all 3 LoRAs:<\/p>\n\n\n\n

ZEMANOVA Fluxed<\/strong> LoRA Model (created by ManOfCulture)<\/p>\n\n\n\n

Link: https:\/\/tensor.art\/models\/775472460280703407<\/a><\/p>\n\n\n\n

From the generated images I can tell that the LoRA is based\nprimarily on images published by Penthouse. As Penthouse posted all their pics\nin the reverse orientation, all the of the MoC generated images result in a\nmirror imaged Veronika.<\/p>\n\n\n\n

Still results in very realistic facial images however.<\/p>\n\n\n\n

Several of the images used to train the LoRA must have been\nmagazine scans as every once in a while you can see stripes in the final image.<\/p>\n\n\n\n

Views from the rear are not as accurate … occasional dark\npatches on the butt cheeks and little or no sidewinders visible.<\/p>\n\n\n\n

Zemanova V2<\/strong> LoRA (created by t-man)<\/p>\n\n\n\n

Link: https:\/\/tensor.art\/models\/796834248926639537<\/a><\/p>\n\n\n\n

Trained using only images after her first set of implants.<\/p>\n\n\n\n

The most accurate LoRA in representing images of Veronika\nfor all orientations. The training images included high resolution pics of\nVeronika from all angles.<\/p>\n\n\n\n

Best ass and sidewinder pics.<\/p>\n\n\n\n

2 front teeth with slight gap accurately represented.<\/p>\n\n\n\n

Zemanova V3<\/strong> LoRA (created by t-man)<\/p>\n\n\n\n

Link: https:\/\/tensor.art\/models\/795395486422106135<\/a><\/p>\n\n\n\n

Trained using only images after her second implants.<\/p>\n\n\n\n

There were fewer naked images from all angles to train the\nLoRA with for this vintage resulting in slightly reduced accuracy for rear\nimages, however frontal and side images are very accurate.<\/p>\n\n\n\n

Second implant breast shape perfectly rendered.<\/p>\n\n\n\n

Find below some sample images from all 3 LoRAs using the\nEXACT same prompt (ie input):<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

When the LoRA was trained it used some specific parameters\nto create the model. These parameters should automatically load when you add\nthe Veronika LoRAs. Deviating or overriding these values will adversely affect\nthe final generated image outcome as more “weight” will be applied to\ncertain variables affecting the final outcome away from the Veronika?s true\nlikeness as taught.<\/p>\n\n\n\n

Please use these defaults as those are the values which\nproduce the most accurate likeness of Veronika, deviating from the default will\nnegatively alter Veronika?s appearance.<\/p>\n\n\n\n

The first critical parameter is the ?strength?.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

For Zemanova V2<\/strong> the default strength is 0.82<\/strong>.\nFor Zemanova V3<\/strong> it is 0.8<\/strong>. <\/p>\n\n\n\n

The final setup of the Model\nand LoRa should look like this (in this case for Veronika V2):<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Next a few other critical parameters which must be set are\nthe Aspect Ratio, Sampler<\/strong>, Sampling Steps<\/strong>, and Guidance Scale<\/strong>.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The Aspect Ratio is merely the size and orientation of the\nresulting image. These sizes are primarily driven by the current capabilities\nof the underlying image generation models and so deviating from the defaults\nwill likely fail to render a useful result. Note that there might be a\ndifference between the current max resolution which can be selected if you have\na FREE vs PRO membership. The image shown above is for a PRO membership.<\/p>\n\n\n\n

For the Sampling Method, please use the following: Sampler =\neuler<\/strong>, Sampling steps = 20<\/strong>; Guidance Scale = 3.5. <\/strong>Again,\ndeviating from these parameters will adversely impact the final rendered result\nfrom looking like Veronika as the LoRA was trained using these specific\nparameters.<\/p>\n\n\n\n

If you scroll further down you will see an option to provide\nUpscaling as part of the initial image generation. This feature will require\nadditional credits per image, but it is still less expensive to perform the\nUpscaling at the initial image generation phase as opposed to first creating\nthe image and then upscaling as a secondary step. I find the built in Upscale\nfunction to be extremely good and beneficial in creating a really accurate\nlooking image. Sometimes if the image is too small, some of Veronika?s features\nget distorted because there simply aren?t enough pixels to accurately render\nit. See the image below for suggested upscaler settings, I found that these\nhave worked extremely well. A max 2X upscale is recommended to not introduce\ntoo many artifacts. The specific R-ESRGAN 4x+ upscaler works exceptionally\nwell, so recommend using it.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

Now that the Model and LoRA are loaded and everything has\nbeen setup, you are now ready to start and create some images.<\/p>\n\n\n\n

Now comes the most important step ?. Entering your Prompt<\/strong>,\nwhich will describe exactly what you want your image to portray. Enter your\nprompt on the top making sure to include the specific keyword to activate the\nLoRA you selected. The specific keywords for the 3 LoRAs are as follows:<\/p>\n\n\n\n