Image models

Uncensored means the model can render adult, NSFW or otherwise restricted content without refusing or returning a blank/blurred output. The catalog distinguishes two levels: NSFW models retain some safety conditioning but handle adult content cleanly, while unrestricted models have had safety filters removed entirely and will render anything they can prompt.

Each model card lists recommended VRAM. As a rough guide: SD1.5 runs on 6–8 GB, SDXL on 8–12 GB, FLUX FP8 on 12–16 GB, and FLUX FP16 on ~24 GB. The /can-i-run tool will check a specific GPU against a specific checkpoint. CPU-only inference works but is roughly 30–60× slower.

SDXL has the largest LoRA library and runs on modest hardware. FLUX produces the strongest prompt adherence and detail at higher VRAM cost. Pony Diffusion XL uses different prompting (score and source tags) and excels at anime, illustration and stylized art. Most practitioners keep two or three checkpoints and switch based on task.

A checkpoint is a full model weight file (typically 2–22 GB) you load as the base. A LoRA is a small (50–500 MB) delta that’s added on top of a compatible checkpoint to inject a style, character or capability. LoRAs only work with the base family they were trained on — an SDXL LoRA won’t work on FLUX, and a Pony LoRA usually won’t work on plain SDXL.

Most listed checkpoints use permissive licenses (CreativeML Open RAIL, FLUX Non-Commercial, Apache 2.0). License details are on each model’s detail page. Hosting and using uncensored content is your responsibility; verify your local regulations and platform terms before generating, sharing or selling output.

Each model card links to its primary host — usually Hugging Face or CivitAI. Click through to the model’s detail page for the direct download link, sample prompts, and the recommended sampler settings.