If you enjoy my contribution to this community, feel free to <a target="_blank" rel="ugc" href="https://ko-fi.com/aramintastudio">buy me a coffee</a>: the more caffeine I drink the more models I can create 😅Comparison gallery here: <a target="_blank" rel="ugc" href="https://civarchive.com/posts/20652937">Fv6-Fv5</a> and NSFW illustration comparison between <a target="_blank" rel="ugc" href="https://civarchive.com/posts/9197527">Gv1-Cv6-Fv2</a>.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/98d68d4c-483a-42aa-8a17-93a3b088665a/width=525/98d68d4c-483a-42aa-8a17-93a3b088665a.jpeg" />Current SOTA model in my experiment:<ul><li>SDXL Base model: Gv4 is the most balanced model allowing both realistic and styled NSFW and SFW images. Better aesthetic than Fv6 but less photorealistic.</li><li>SDXL photorealist (SFW and NSFW) model: Fv6 is the way to go for hyper-realism including realistic NSFW images but it mostly lacks the styling capabilities of Gv4.</li><li>SDXL Illustration : Gv4 (SFW and NSFW). Cv6 is however still worth a try if you are not into NSFW images.</li></ul><ul><li>Flux model: Flux1-A1</li></ul>Image Generation Settings for SDXL modelsDPM++ 2/3M SDE / Karras or Exponential are always a good bet with 25+ steps and CFG around 5-7. But DPM++ SDE / Karras with less steps (e.g. 12) and higher CFG (8-11) is worth a try as is Euler Ancestral / Normal for a less detailed result which may also be interesting with a model like Fv6 to get something less "noisy".The default CLIP Skip of 2 is also a good bet, but using 1 or 3-4 is also worth trying: 1 push more towards prompt adherence and 3-4 give sometimes a better result than the default focusing more on the "concepts".Image Generation Settings for Flux modelsMy preferred settings are DPM++ 2M / beta or sgm_uniform or DDEIS / normal for the sampler / scheduler, beta giving a bolder stronger image. For a more subtle image, Euler / simple or beta seems a good bet.CFG seems to have a huge impact on the final image and be very sensitive even to small variations.<ul><li>For photos, CFG should remain low (1.5-2.5) to avoid plastic skin.</li><li>For fine art and illustration it is more complicated because it depends on the medium. For "rough" styles (painting, watercolours etc.), CFG should stay quite low in the 1.5-2.5 range but for anime or comic style, CFG needs often to be pushed further to achieve the desired style (3-6 or more).</li></ul>If the image is messy/malformed or blurred, it is often because the CFG/steps are inappropriate for this image, but it is not always easy to know whether CFG/steps must be increased or decreased (at least to me 😊).There is for sure a lot to learn concerning Flux behaviour which is quite different than SDXL and we will need to adapt.<h3 id="workflow">Workflow</h3>All my images are uploaded with embedded ComfyUI workflow which is alas incompatible with CivitAI processing and most often prompt cannot be retrieved. You can however download the original PNG image with the workflow included by clicking on the "DOWNLOAD" icon in the image viewer.Why I try to mostly publish images straight from my model with maybe a bit of a Lora (mine or some detail enhancer), I also sometimes now use Controlnet to get better more detailed compositions more easily: in this case the source image is obviously not in the workflow, but I guess you can still use the image I publish as a source if you want to make a variation :)<h3 id="past">Past</h3><hr />Starting from E serie, models are evolving sometimes by merging with other models (thanks to other contributors!), but mostly via training on my own dataset: a modest dataset (~2000 images currently), but I try to somehow compensate with quality and originality.Starting with Fv1, I have included many synthetic images I created using previous versions: playing hard with prompt and retouching when necessary the result in Photoshop in order to have a dataset which contains many original images.<hr />The core idea behind this model was to create a versatile tool by merging some of the best existing models which fits my personal taste (photography and fantasy art to make it simple). My primary goals were:<ol><li>Photorealism: The ability to produce stunningly realistic images of both people and objects/nature.</li><li>Flexibility: The ability to create highly stylized images, allowing for artistic expression through various styles and combinations of artists. I am from a older generation and comes from Europe so "style" does not mean for me "Japanese kawaii aime with boobs" or "DC Comics cartoon with lots of superheroes and voluptuous blonde babes" but more from a univers of Frank Frazetta, Milo Manara, Boris Vallejo, H.R.Giger, Wojtek Siudmak and such fantasy art masters: there are boobs involved for sure, but the style is somewhat different :P</li><li>As I doesn't like being limited in my exploration of the human body, the idea is also to have a fairly capable NSFW model. However, dur to the nature of available training images in the data sets, NSFW often comes with a strong bias toward either porn photos or porn Japanese anime and impact the flexibility (typically as soon as you use the word "sexy" in your prompt you need to weight in the style). This point is thus NOT the priority for the base model but is pushed forward in the NSFW model.</li></ol>

The Araminta Experiment (SDXL+Flux)

Based on Ev2 and merged with a Lora I trained on about 800 hand picked, high quality images, carefully retouched (think watermark) and manually captioned. And the result is beyond my expectation and validate this approach for further improvement.I will now concentrate on improving my dataset :)This version is basically again a step forward compared with Ev2 with almost everything quite noticeably better:<ul><li>Better NSFW (still room for improvement though)</li><li>Better illustration styles (will improve it further in the future because I like that)</li><li>Better faces and skin texture</li><li>Follow better the prompt</li></ul>Ev3 behave a bit differently thanEv2 though and parameters or maybe prompt may have to be tweaked a bit, but it is because it has more details natively and is better at understanding the prompt.

Files with SHA256: c8324bf27c51649c603d568a7293e8f56dc2064a618b9afa8ef9b0b6ebba33b1