Stable Diffusion: every generation from SD 1.5 to 3.5
Stable Diffusion by Stability AI: SD 1.5, 2.1, SDXL, SD3 and 3.5, with text encoders, sizes, licenses, and which generation to use for which job.
What Stable Diffusion is
Stable Diffusion is the text-to-image line from Stability AI. It changed architecture twice: the first three generations are latent diffusion models built on a U-Net, and Stable Diffusion 3 moved to a Multimodal Diffusion Transformer (MMDiT). The line also has a video side line. Because so many community fine-tunes exist, the generation you pick often matters more than the specific checkpoint.
Generations in order
Stable Diffusion 1.x
Stable Diffusion 1 covers v1.4, v1.5 and their re-uploads. The v1.5 model card describes a latent diffusion model with a CLIP ViT-L/14 text encoder feeding a U-Net by cross-attention, fine-tuned at 512x512 for 595,000 steps on the laion-aesthetics v2 5+ dataset. The license is CreativeML OpenRAIL M. The original runwayml repository is deprecated, and the card we read is a mirror of it. We did not source a release month for this generation, so we give none.
Stable Diffusion 2 (November 2022)
Stable Diffusion 2 was announced on November 24, 2022. It switched to an OpenCLIP text encoder, and the announcement lists default resolutions of 512x512 and 768x768. The release also included a 4x super-resolution upscaler, a depth-to-image model and an updated inpainting model. The SD 2.1 card lists a 0.9B-parameter model, OpenCLIP-ViT/H as the text encoder, resolutions up to 768x768, and the CreativeML Open RAIL++-M license.
Stable Diffusion XL (July 2023)
Stable Diffusion XL 1.0 was released on July 26, 2023. The announcement gives 3.5B parameters for the base model and 6.6B for the full base-plus-refiner ensemble. It uses two text encoders (OpenCLIP-ViT/G and CLIP-ViT/L), and the refiner is optional: the card says the base model works on its own. The license is CreativeML Open RAIL++-M. SDXL is the base for a very large share of community checkpoints, including turbo and lightning distillations and ControlNet models, which is why its hub is the biggest in this series.
Stable Diffusion 3 and 3.5
Stable Diffusion 3 replaces the U-Net with an MMDiT and uses three text encoders: OpenCLIP-ViT/G, CLIP-ViT/L and T5-xxl.
- SD3 Medium has 2B parameters. The diffusers card we read lists a non-commercial research license, with Creator and Enterprise licenses for commercial use.
- SD 3.5 was released on October 22, 2024, with the Medium variant added on October 29. Large has 8.1B parameters per the announcement (the model card says 8 billion) and targets 1 megapixel output. Large Turbo is a distilled version that generates in 4 steps. Medium is 2.5B parameters (the announcement figure) and was built to run on consumer hardware. The cards list the Stability Community License: free for research, non-commercial use and commercial use below $1M in annual revenue, with an Enterprise License above that.
The Medium card says its MMDiT-X design has self-attention modules in the first 13 transformer layers, and warns of edge artifacts when T5 prompts exceed 256 tokens.
Which one to use today
- Largest ecosystem of LoRAs, ControlNets and fine-tunes: SDXL. Most community adapters target it.
- Smallest and most legacy-compatible: SD 1.5, which runs at 512x512 and has the widest tool support, but is the oldest architecture here.
- Best prompt following and text rendering from Stability weights: SD 3.5 Large or Medium; SD3 and 3.5 use T5-xxl alongside the CLIP encoders. Check the Community License revenue cap before commercial use.
- Few-step generation: SD 3.5 Large Turbo (4 steps), or the SDXL turbo and lightning derivatives in the SDXL hub.
- Depth-conditioned or upscaling workflows from the SD2 era: Stable Diffusion 2.
For new work with no existing adapters, start with SD 3.5 Medium or Large. If you depend on community checkpoints, stay on SDXL.
Side lines
- Stable Video Diffusion: Stable Video Diffusion generates short clips from a still image. The XT model produces 25 frames at 576x1024, up to about 4 seconds, cannot be steered by text prompts, and uses the stable-video-diffusion-community license. The card says it was fine-tuned from a 14-frame model.
Running Stable Diffusion
All generations load in Diffusers, and SD 3.5 is also listed for ComfyUI on its card. You can run them on Aquanode GPUs.
Stable Diffusion generations
Every Stable Diffusion generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Stable Diffusion 1 | 3 | 860M | 3.8 GB | RTX 3070 | $0.088/hr |
| Stable Diffusion 2 | 5 | 866M | 3.9 GB | RTX 3070 | $0.088/hr |
| Stable Diffusion XL | 34 | 611M to 2.6B | 11.5 GB | V100 | $0.110/hr |
| Stable Diffusion 3 | 3 | 2.1B to 8.1B | 4.7 GB | RTX 3070 | $0.088/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Other Stable Diffusion lines
Side lines from Stability AI: specialised or companion models outside the main Stable Diffusion generations.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Stable Video Diffusion | 2 | 1.5B | 6.8 GB | RTX 3070 | $0.088/hr |
Stable Diffusion compared with other models
Side-by-side VRAM, context length and license for models of a similar size.
- FLUX.1-dev vs stable-diffusion-xl-base-1.0
- FLUX.1-schnell vs stable-diffusion-xl-base-1.0
- Qwen-Image vs stable-diffusion-xl-base-1.0
- Qwen-Image-2.1 vs stable-diffusion-xl-base-1.0
- FLUX.1-dev vs sdxl-turbo
- FLUX.1-schnell vs sdxl-turbo
- Qwen-Image vs sdxl-turbo
- Qwen-Image-2.1 vs sdxl-turbo
- FLUX.1-dev vs stable-diffusion-v1-4
- FLUX.1-dev vs sd-turbo
- FLUX.1-dev vs stable-diffusion-3.5-medium
- FLUX.1-schnell vs stable-diffusion-v1-4
- FLUX.1-schnell vs sd-turbo
- FLUX.1-schnell vs stable-diffusion-3.5-medium
- Qwen-Image vs stable-diffusion-v1-4
- Qwen-Image-2.1 vs stable-diffusion-v1-4
- Qwen-Image-2512 vs stable-diffusion-v1-4
- FLUX.1-Krea-dev vs stable-diffusion-v1-4
- Qwen-Image vs sd-turbo
- Qwen-Image-2.1 vs sd-turbo
- Qwen-Image-2512 vs sd-turbo
- FLUX.1-Krea-dev vs sd-turbo
- Qwen-Image vs stable-diffusion-3.5-medium
- Qwen-Image vs stable-diffusion-3.5-large
- Qwen-Image vs stable-diffusion-3-medium-diffusers
- Qwen-Image-2.1 vs stable-diffusion-3.5-medium
- Qwen-Image-2512 vs stable-diffusion-3.5-medium
Best models by task
Stable Diffusion models appear on these ranked task pages.
Sources
- https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5
- https://stability.ai/news-updates/stable-diffusion-v2-release
- https://huggingface.co/sd2-community/stable-diffusion-2-1
- https://stability.ai/news-updates/stable-diffusion-sdxl-1-announcement
- https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
- https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers
- https://stability.ai/news-updates/introducing-stable-diffusion-3-5
- https://huggingface.co/stabilityai/stable-diffusion-3.5-large
- https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
- https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.