Stable Diffusion Model Downloads: Versions, Licenses, and Install Folders

Before you download

Choose a checkpoint by architecture, exact file, license, and workflow. SDXL checkpoints are not interchangeable with SD 1.5 checkpoints, and v-prediction models can require different scheduler settings. Read the license before commercial use.

“Verified” here means the model identity, named version, file, repository-reported display size, and license display were checked against the linked official, creator, or explicitly labeled mirror source on July 23, 2026. It does not mean OfflineCreator downloaded, installed, benchmarked, or quality-ranked the model.

File size is not runtime VRAM. None of the retained source cards publishes one universal minimum-VRAM figure, so this catalog does not invent one.

Quick catalog

rank on the rendered cards is catalog order, not a quality score.

Official base checkpoint

Stable Diffusion XL Base 1.0

Use Stable Diffusion XL Base 1.0 when you want the original official SDXL checkpoint rather than a community fine-tune.

• File: sd_xl_base_1.0.safetensors
• Repository display size: 6.9 GB
• License: OpenRAIL++
• Source: official Stability AI Hugging Face repository

The official card does not state a simple minimum VRAM. Start from the runtime's SDXL requirements and your selected resolution, precision, and offload strategy.

SDXL community checkpoints

Juggernaut X v10

Juggernaut X v10 is pinned to Juggernaut-X-RunDiffusion-NSFW.safetensors, not a generic “latest Juggernaut” label.

• Repository display size: 7.1 GB
• License metadata: CreativeML Open RAIL-M
• Additional creator term: paid API deployment requires explicit licensing
• Source: RunDiffusion creator repository

The card does not publish a minimum VRAM. This is a file-catalog entry, not a claim that Juggernaut is the quality winner.

RealVisXL V5.0

RealVisXL V5.0 is pinned to RealVisXL_V5.0_fp16.safetensors.

• Repository display size: 6.9 GB
• License: OpenRAIL++
• Source: SG161222 creator repository

Do not substitute an older V4 repository or infer a VRAM floor from the 6.9 GB checkpoint size.

DreamShaper XL v2 Turbo

DreamShaper XL v2 Turbo is pinned to DreamShaperXL_Turbo_v2_1.safetensors.

• Repository display size: 6.9 GB
• License: OpenRAIL++
• Card example: six steps and CFG 2
• Source: Lykon creator repository

The example settings are version-specific guidance, not a universal promise about speed or output quality.

Anime and tag-based checkpoints

Pony Diffusion V6 XL

Pony Diffusion V6 XL is represented by a verified re-export of the official CivitAI weights.

• File: ponyDiffusionV6XL_v6StartWithThisOne.safetensors
• Repository display size: 6.9 GB
• License: modified Fair AI Public License 1.0-SD
• Material term: monetized hosted inference is restricted without permission
• Setup: Pony-specific prompt conventions and CLIP skip 2

The re-export is labeled as such; it is not described as the original creator repository.

NoobAI XL V-Pred 1.0

NoobAI XL V-Pred 1.0 is pinned to NoobAI-XL-Vpred-v1.0.safetensors.

• Repository display size: 7.1 GB
• License: Fair AI Public License 1.0-SD plus creator additions
• Creator term: commercialization of the model, derivatives, and generated products is prohibited
• Setup: v-prediction-compatible configuration; creator recommends Euler

Do not load this as an ordinary epsilon-prediction SDXL checkpoint and assume default scheduler behavior will match.

SD 1.5 community checkpoints

Realistic Vision V6.0 B1 no-VAE

Realistic Vision V6.0 B1 no-VAE is pinned to Realistic_Vision_V6.0_NV_B1_fp16.safetensors.

• Repository display size: 2.1 GB
• License: CreativeML Open RAIL-M
• Important packaging note: no VAE; the creator recommends a separate VAE
• Source: SG161222 creator repository

The smaller file does not by itself establish a VRAM minimum.

ChilloutMix fp16

ChilloutMix fp16 is explicitly labeled as a Hugging Face mirror duplicated from Chillout2023/Chilloutmix.

• File: Chilloutmix-ema-fp16.safetensors
• Repository display size: 2.5 GB
• License metadata: modified CreativeML Open RAIL-M
• Additional card terms: Dreamlike Diffusion restrictions, including hosted-inference and attribution conditions

Use this entry only if the mirror provenance and inherited terms are acceptable for your use case.

Held and removed entries

• CyberRealistic XL: held because creator-card and repository-tree version, size, and license wording conflict.
• epiCRealism: held until one exact creator-authorized checkpoint file, version, license, and size are pinned.
• Beautiful Realistic Asians V7: removed because only third-party mirrors established the claimed file and no authoritative creator source was confirmed.
• Detail Tweaker and Background Detail Enhancer: removed because LoRAs belong to the separate LoRA catalog.
• Stable Diffusion 3.5 entries: held until gated files and required dependencies can be authenticated and verified.

Checkpoints are not LoRAs

A checkpoint supplies base model weights. A LoRA modifies a compatible checkpoint and belongs in a different folder. Use the ComfyUI LoRA model download list for architecture-matched adapters.

File format and precision

Prefer a source-provided .safetensors inference file. Pick fp16 only when the creator provides and identifies it. Do not assume every file called pruned, EMA, or fp16 is interchangeable. A checkpoint with “no VAE” in its name may require a separate VAE.

Safetensors narrows the arbitrary-code risk of pickle-based weight files, but download sources, installers, extensions, and custom nodes still require review.

Install folders

For ComfyUI, place ordinary checkpoint files in ComfyUI/models/checkpoints/. For Forge or AUTOMATIC1111, place them in models/Stable-diffusion/. Some newer architectures use separate diffusion-model, text-encoder, and VAE files; follow the exact official workflow rather than forcing every file into the checkpoint folder.

Refresh or restart the UI after copying files. LoRAs do not go in either checkpoint folder.

License check

A free download is not unrestricted commercial permission. Review the linked card before hosted inference, paid APIs, client work, merging, redistribution, or commercial use of outputs. Preserve a copy of the terms for the exact version you download.

For evaluated cross-architecture recommendations, continue to Best Uncensored CivitAI Models.

The catalog intentionally favors stable identity and provenance over the old twelve-item count. Unsupported LoRAs and unclear mirrors were replaced with an official base checkpoint, exact filenames, version labels, source classifications, install destinations, and model-specific license cautions.

Re-check every linked source immediately before publication because repository files, gates, and terms can change.