bigASP v2.5

>Key Features:
Massive training scale: Trained on 13 million images with 150 million training samples
Flow Matching technology: Replaces SDXL's broken noise schedule, resulting in better image structure and dynamic range
Reduced failure rate: Significantly fewer mangled generations, duplicate limbs, and "little buddies" compared to v2
Expanded concept coverage: Includes anime/furry data for wider creative possibilities while maintaining photorealism
Better dynamic range: Can generate proper dark, light, and high-contrast images unlike standard SDXL
JoyCaption Beta One: Updated captioning system for more flexible prompting styles

>Technical Requirements:
ComfyUI only: Requires specific workflow with ModelSamplingSD3 node
Specific resolutions: Works best with predefined resolutions (832x1216, 1024x1024, etc.)
Precise sampler settings: Needs careful tuning of shift, CFG (3-6), and PAG (~2.0) parameters
Euler sampler recommended: Other samplers mostly fail or underperform

>Pros:
For experimenters: If you enjoy pushing boundaries and don't mind technical challenges
Better image quality: Flow Matching fixes many SDXL structural issues
Unique capabilities: Combines photorealism with expanded creative concepts from anime dataset
Research purposes: Learn from cutting-edge modifications to SDXL architecture

>Cons
No LORA support: Can't use existing LORAs or train new ones
No merging capability: Incompatible with standard SDXL models
Steep learning curve: Requires extensive parameter tweaking
Limited tool support: Only works in ComfyUI, not in other UIs
Experimental nature: Not suitable for production or casual use

Download if you're an advanced user who wants to experiment with a technically superior but challenging model that produces higher quality images than standard SDXL when properly configured.

https://civitai.com/models/1789765/bigasp-v25