ICBINP | Civitai vs Drag Your GAN

ICBINP is a Stable Diffusion 1.5 checkpoint merge built to push photorealistic image generation as far as a local SD 1.5 setup allows. The Civitai model page hosts multiple versions, with Mid 2024 as the latest major release updated in February 2025. Each variant ships as a pruned fp16 SafeTensor file with the SD-v2 840k VAE baked in, so you can load it in Automatic1111, ComfyUI, or any standard SD 1.5 interface without swapping VAEs.

Where most photorealistic checkpoints chase a single aesthetic, ICBINP grew through successive merges. It started from the RCNZ Gorilla With A Brick base, then folded in ten more photorealistic models at varying weights plus noise offset tuning. The Mid 2024 version adds a LoRA trained on Pexels stock photos merged with ICBINP Final, which tightens skin and lighting realism but still shows typical SD 1.5 quirks like cleft chin bias.

The author recommends DPM++ 3M SDE Karras or DPM++ 2M Karras at 20 to 30 steps with CFG between 2.5 and 5. Resolution sweet spot is 640x960 with a hires fix pass, though 768x1152 works if you hunt seeds to avoid duplicate heads. Beyond portraits, the checkpoint handles CGI characters and landscapes when prompted accordingly.

Downloads are free from Civitai under CreativeML Open RAIL-M. The Mid 2024 file (ICBINP_midYear.safetensors) weighs 2.21 GB and needs about 3.4 GB recommended VRAM. With 163,500+ downloads and 285 Very Positive reviews, it remains one of the most downloaded photorealistic SD 1.5 merges on the platform.

Drag Your GAN lets you reshape a generated image by dragging handle points to target spots on the canvas. Instead of training on labeled pose data or 3D priors, DragGAN moves pixels inside a pretrained GAN's image manifold so edits stay photorealistic even when occluded regions get hallucinated. The project page hosts the SIGGRAPH 2023 paper, demo clips, and links to open source code.

Most GAN editors rely on fixed training annotations or 3D meshes, which limits what you can deform. DragGAN pairs feature-based motion supervision with discriminative point tracking so handle points follow your cursor while staying on the object surface. That makes it useful for nudging animal poses, facial expressions, car angles, and landscape layouts without retraining the generator.

Researchers, graphics students, and computer vision engineers use it to prototype interactive image editing workflows. The Max Planck Institute page distributes the paper PDF and GitHub repo under a non-commercial Creative Commons license, with contact details for the lead author.