AngeliumiX V.6: Merge Workflow & Improvements
🎯 Objective
AngeliumiX V.5 is a photorealistic-focused SDXL merge, designed to move away from the "AI-perfect" look and closer to natural, diverse, and lifelike portraits. The goal is to achieve:
More natural faces: Realistic skin tones, subtle imperfections, and less artificial perfection.
Diversity: Better representation of various ethnicities, facial structures, and body types.
Photorealism: Results that resemble photography rather than 3D renders.
🔧 Merge Recipe
Base Checkpoints (Safetensors)
the best of the best !
VAE
SDXL_vae.safetensors
🎨 Recommended Use Cases
Best for: Portraits, photorealistic people, fashion, lifestyle.
Works well with: Subtle lighting prompts, cinematic styles, natural photography.
Sampler: DPM++ 2M SDE Karras or DPM++ 3M SDE Karras.
CFG Scale: 2.5–4 (for realistic results).
STEPS : 25 minimum
📏 Resolution Guidelines
Resolution
1024 × 1024 1:1 Square Portraits, social media
1152 × 896 9:7 Horizontal Close to 4:3
896 × 1152 7:9 Vertical Close to 3:4
1216 × 832 19:13 Close to 3:2
832 × 1216 13:19 Close to 2:3
1344 × 768 7:4 Close to 16:9
768 × 1344 4:7 Close to 9:16
1536 × 640 12:5 Wide horizontal
640 × 1536 5:12 Tall vertical
1024 × 1280 4:5 Vertical
😊 Positive Prompts
Useless to describe more than 77 tokens.
I create a prompt helper here with Grok 4, you enter in your language some elements and Grol do the rest. Inscription is free and free to use it : https://poe.com/SDXL-prompthelper
⚠️ Negative Prompts
Default: "3D render, octane render, CGI, simplified background, wrong perspective, text, watermark, error."
For architecture: Add "wrong architecture" to reduce distortions.
More complete : 3D, CGI, doll, deformed, bad anatomy, blurry, watermark, text, artificial, waxy skin, bad hands, ugly, distorted, unnatural colors, fake, airbrushed, bad lighting, lowres, jpeg artifacts, oversaturated, uncanny valley, instagram filter, bad shadow, floating hair, cloned face
🔄 Workflow Notes
The ComfyUI workflow is included for easy replication.
Copy-paste the workflow to comfyui and load the merge and start experimenting.
💡 Improvements Over Previous Versions
More natural faces: Less "AI-perfect," with realistic skin tones and subtle details.
Enhanced diversity: Better handling of different ethnicities and body types.
Photorealism: Closer to photography than 3D renders.
📌 Final Notes
The merge is versatile and works well for a variety of photorealistic applications.
Experiment with samplers, CFG scales, and resolutions to fine-tune results.
Description
Angeliumix 4 — SDXL Photorealistic Master Blend
Description:
Angeliumix 4 is the latest evolution in the Angeliumix series — a hand-crafted fusion of the finest SDXL photorealistic models, blended to deliver flawless anatomy, superior prompt adherence, and cinematic realism.
This version refines skin texture, depth, and lighting for true-to-life results, whether you’re creating portraits, full-body shots, fashion editorials, or cinematic scenes.
Key Features:
Ultra-realistic detail with natural skin tones and lifelike textures.
Consistent, accurate anatomy for dynamic and natural poses.
High prompt fidelity — captures even nuanced descriptive cues.
Optimized for both close-up shots and full-scene compositions.
Best For:
📸 Portrait photography simulation
🎭 Character design & cosplay concepts
🎬 Cinematic scenes & storytelling renders
👗 Fashion, editorial, and lifestyle imagery
Angeliumix 4 is designed for artists who demand photorealism without sacrificing creative control.
FAQ
Comments (5)
I tested v2.5, 3.1, and 4.0 side by side using the same seed. Currelty, 2.5 and 3.1 are the most realistic, with better color balance and tone; 2.5 is closer to "realistic" photography on more seeds than 3.1. Version 4.0 has a different ("colder") color balance, and tends to overexpose; it requires a slightly lower CFG than the previous versions, and even then it is somewhat less photography-like, being a bit brighter and higher contrast than needed. I suggest addressing these issues in the potential v4.1.
It can be seen by comparing https://civitai.com/images/93386682 and https://civitai.com/images/91727210
ok I think it's about too much donuts details. thank you for your feedback
4.1
Angelium 4.1 is extremely strange: it mildly disrespects styles (e.g. vintage photo, noir) and it mildly disrespects ethnicities (e.g. dark-skinned, hispanic), leaning to the same body type and face most of the time. Generations look synthetic in a way I cannot explain; it is something like the original Flux perhaps. 4.1 demonstrates severe aesthetic alignment to a certain style. Generation results on the same seeds change drastically between 2.5/3.1 (similar) and 4.1 (drastically different). White balance still leans towards daylight (much more so than 2.5 or 3.1).