"In the gleaming city by the sea, contentment comes bottled and wired. Streets clean, bellies full, minds empty. They call it progress — this strange paradise where no one wants for anything, yet hunger still gnaws beneath the surface. Synthetic pleasures pulse through the veins of the willing while outside, reality withers untouched. What use is freedom when we've forgotten how to dream? The machines didn't conquer us; we handed them the keys to our souls and called it convenience. Now we drift, eyes glazed, through the marketplace of artificial bliss, collecting experiences that leave no mark, chasing sensations that fade like morning mist. This is how utopia dies — not with revolution, but with a satisfied sigh." (Circles of Paradise)<h3 id="buzz-donation-appreciated-it-helps-to-train-faster.">BUZZ donation appreciated - it helps to train faster.</h3><h3 id="realism-ea">Realism EA</h3>This version trained at 1024px so please don't expect it to be stable at resolutions higher 1440 by longest side. It was reported that this version gives mostly broken images at 1440x1904. Also, please note that eyes and fingers need some improvements. Sometimes they are off. Constructive feedback on what to improve appreciated!<h3 id="anime-and-cute-nai-ill">Anime and Cute NAI-Ill</h3>First attempt to make a (semi)realistic NoobAI model. Based on several popular NoobAI resources, it was trained on realistic datasets from CinEro V6. Training performed in several stages at different resolutions (1024 and 1216).Version prior the Realism AI can render 1216x1440, 1280x1600 picture in Text-to-Image mode without noticeable duplication issues (during my test I've got only few outputs with duplication at 1440x1904 resolution).Most first posts by me made with minimal ComfyUI workflow. Only one run without ADetailer or HiRes Fix techniques, but with few nuances. This single run from the start is done in 2Mpx resolution and special node "Detail Daemon Sampler" is used. Detail Daemon let's you add mode micro-details while keeping CFG low (low contrast image with high small details). Probably Forge and A1111 with conventional samplers will give less details.Try rendering High Resolution images first. Let me know if hires render will give you duplication issues.<h3 id="instructions">Instructions</h3><ol><li>Use my ComfyUI workflow attached as training data (Detail Daemon custom nodes required + some optional nodes which might be replaced with others).</li><li>Start with high resolution depending on max VRAM on your computer (1280x1600 or 1216x1440 for example).</li><li>Drag my samples into any ComfyUI workflow to look at prompts.</li><li>Try usual samplers like Euler + Beta or Euler A + Karras. Experiment with more noisy DPM++ 3M SDE for more unstable results.</li><li>Steps 30 or more recommended. CLIP Skip 1 is recommended, but 2 also might work with less realistic results.</li></ol>

CoP | NAI | ILL | EPS | Circles of Paradise

TL;DR; NoobAI Anime style checkpoint converted to photorealism realism by multiple training sessions with special dataset (read details below). It was tested side-by-side with classic NoobAI model (see OPs for comparison). BUZZ support es really needed to continue develop this branch of NAI.Tech DetailsFor training I rendered anime-ish pictures with classic NAI checkpoint and converted these pictures into realism using a very complex ComfyUI workflow which respects the composition but totally rewrite color palette and textures using CinEro V6 realistic model. This way I created the dataset combining the original prompts and composition of NAI Anime, but have realistic look. Hope you will feel this approach promising.

Files with SHA256: f2599ea96402a11b9ffffed14e204c6c7478199b41230b4e814f8ab5933180f7