Electrum Cinnamon
A new year and a new model.
I did a number of test with this model and wrote down what I found. If you want a quick look the article is here: Sexiam - Electrum Cinnamon Merge Test
Pros:
Electrum Cinnamon is a little different from my other models in that it pushes more toward realism in textures, anatomy, and lighting. Skin has a nice matte finish, free of the usual glassy shine you often see in generations. Clothing, surfaces, and materials all have realistic microtextures. Overall, it's a solid stylized photorealistic merge with great detail.
Cons:
Electrum Cinnamon doesn't have quite the same range of faces as some of my other models. That's a sign of slight overfitting from using certain lighting and realistic LoRAs during the merge process. Girls often end up with similar faces unless you prompt for more specific details.
The model also really likes close-ups and cowboy shots. An easy way to work around these limitations is to adjust your prompt weights—something like (full body:1.2) will reliably force it to zoom out.
There's also a subtle earth-tone tint to most images, which comes from the merging process. You can counter that by adding simple [color] background. For example, simple blue background will give the whole image a cooler blue color scheme, especially if you've already prompted details about the background and environment. If you haven't described the environment, it'll just give you a solid color backdrop.
Recommended Settings:
Steps: 24–32
Sampler: Euler a or DPM++ 2M Karras
CFG: 5–8
Resolution: Standard 1 MP SDXL sizes (e.g., 1024x1024, 768x1344, 896x1152, etc.)
VAE: Default or the model's embedded VAE
Negative Prompt:
worst quality, low quality, displeasing, text, , watermark, bad anatomy, text, artist name, signature, hearts, deformed hands, missing finger, shinny skin, child, children, loli, monochrome, bad framing, poor framing,
Quality Tags:
examples are all the same settings and seed
Prompt: 1girl, ghost in the shell, kusanagi motoko, holding gun, highleg leotard, cyberpunk, bodysuit,,, looking at viewer, on the edge of a building looking over a cyberpunk city at night, nighttime, dark sky,
masterpiece, best quality, very aesthetic, ultra-HD, cinematic lighting, depth of field, (blurred background), dramatic lighting,
masterwork, masterpiece, best quality, depth of field, (sketch, lineart, crosshatching:1.2),
newest,absurdres,incredibly absurdres, best quality, amazing quality, very aesthetic, limited palette,
masterwork, masterpiece, best quality, (masterpiece:1.2), (best quality:1.2),(very aesthetic:1.2), (absurdres:1.2), newest,
masterpiece, newest,absurdres,incredibly absurdres, best quality, amazing quality, very aesthetic, (anime_screencap, retro:1.3),
vibrant hues, dynamic pose, blue palette, realistic, dynamic pose, abstract background, (4k, detailed, intricate details, best quality, amazing quality, masterpiece), vdsur, surreal, colorful background (simple black background), (volumetric lighting),
Description
Electrum Cinnamon Anime is my attempt at adding an anime-style aesthetic to the merge. The model itself still retains its flexibility and can produce a range of styles, but the baseline bias leans more toward anime-inspired features in the characters.
When generating at standard SDXL resolutions (around 1MP), small faces in wide or long shots may appear softer—especially in highly detailed, photorealistic compositions. If a face occupies only 5–10% of the image, it represents a relatively small number of pixels. At that scale, there simply isn’t enough spatial resolution to cleanly resolve fine features such as eyelashes, subtle lighting gradients, and micro-texture.
Because the model must distribute detail across the entire image, complex backgrounds, textured clothing, and lighting effects can further reduce the effective detail allocated to small subjects. As a result, facial features may lack sharpness at base resolution.
Upscaling to 2MP or higher—via High-Res Fix, latent upscaling, or detail refinement passes—provides greater pixel density. This gives the diffusion process more spatial bandwidth to define edges and refine facial structure, resulting in noticeably sharper features in longer shots.
In contrast, anime-style rendering holds up better at lower resolutions. Stylized models rely more on bold shapes, flat colors, simplified gradients, and strong line work rather than high-frequency micro-detail. These structural elements remain crisp even at lower pixel counts, which is why anime compositions often appear sharper at 1MP compared to photorealistic generations.
FAQ
Comments (4)
Looks terrific and you can see the care and thought that went into this in your post. Can’t wait to put this model through its paces.
I waited quite a bit to review this Checkpoint. First, It works well with complex ideas and characters, they are mostly all represented without Lora and you can accomplish very unique character builds with the checkpoint alone. Its diverse ability with resolution is amazing. You may get some stretching, but overall, it does the job all the way up to the 1536 area. It holds style well and allows for realistic overlap with its anime background. The only downside is that the Adetailer is going to struggle if you need it for character consistency. You will have to bump the denoising and that tends to effect hair textures with pretty much anything trained on other loras or checkpoints. I dont hate it. I like to use it to make hair gradient colors. Downsides, If you want to use your images in workflows like, anime to realism or editing software, the equipment tends to struggle with the high detail, high artifact images, You wont be able to make them better than what they already are with this checkpoint. The other issue would be pose can be incredibly effected by resolution. You will have to consider poses based on that to make your ideas work. Overall, I am over 10000 images into this checkpoint and its honestly top 3 of any image generation equipment available as of mid 2026. I leave my prompts and metadata on the images if you want to try it out. Definitely try using realism leaning loras in the face detailer. Its pretty cool.
Thanks for the review.
A lot of Illustrious models struggle with a latent image larger than 1 megapixel, and from tests you get a weird stretching of the figure's proportions. If you're going for something larger than standard SDXL sized images, I wouldn't go over 1.5 MP. The size of an image and it's origination also has a large impact on SDXL models.
Becuase the model is trained on micro-details, it's sensitive to detail patterns if the denoise value is too high. You might need to set it as low as .3 if you're using the model in an adetailer workflow. More 2D style models don't have that same issue, and you can go as high as .55 denoise before running into errors. Generally the more a model is capable of photo-real details, the more issues adetailer system seems to have with adding too much detail into the image.
You can easily upscale image with a two-stage upscaling system. If you set the second stage to .35 denoise with an img2img system, you can process the second stage at 5MP if you have the VRAM. After that you can use SD Ultimate Upscaler to go past 20 MP with a denoise of .25
I would also recommend using SEGS inpaiting systems with models like this one. SEGS is based on a differential diffusion system that is better at blending details into an image rather than overwrite them with standard crop and stitch systems.
@Sexiam I appreciate the follow up and i love that you document your makes. I truly believe you make this hobby more professional. My comments are basics only. i found a large amount of success by using my character lora in the adetailer at higher denoising to bridge the gap between realism and non with the model. I do a none masked area 1st round and on the second the lora hits it at full weight and I have the most life like images that a non upscaled, unpolished run can get me. I even run your checkpoint on faces in other checkpoints. You can look through my style and see how well its worked for me. I will try out your recommendations and I hope it brings you Joy that it brings me great success in my attempts and your prompts have taught me a lot as well as your use guides. I prompt completely different because of the one for this. Sincerest of appreciation for the time it takes to tune a checkpoint of this quality. Looking forward to the future. Keep it coming. The group definitely needs it.