# Stable Diffusion 1.5 / 2.x

Stable Diffusion 1.x and 2.x are the original latent-diffusion UNet models. They are small and fast
by current standards, run on modest hardware, and have the largest library of community fine-tunes,
LoRAs, embeddings and ControlNets.

- **SD 1.x** (in practice, **SD 1.5** and its fine-tunes) is trained at **512×512** and uses a single
  CLIP text encoder. It is fully supported: text-to-image, image-to-image, inpainting and outpainting,
  control layers, reference images, regional guidance and upscaling.
- **SD 2.x** is treated as **legacy**. Models you install yourself load and generate, but InvokeAI
  ships no SD 2.x starter models and some Canvas features are not available for it (see
  [SD 2.x](#sd-2x)).

InvokeAI identifies the family, the **variant** (normal or inpainting; SD 2.x also has a depth
variant) and the **prediction type** (epsilon or v-prediction) automatically on install.

## License

| Model | License | Commercial use |
|---|---|---|
| SD 1.5 base, Dreamshaper 8, CyberRealistic v4.1, ReV Animated | CreativeML OpenRAIL-M | Allowed, subject to the license's use restrictions |
| Deliberate v5 | CC BY-NC-ND 4.0 | Companies and services must contact the author; the author places no restrictions on private individuals |

None of the starter downloads is gated.

This is a summary, not legal advice. Read the full license on the model's HuggingFace page before using a model commercially.

## Hardware

SD 1.5 is the lightest local image model InvokeAI supports: the
[System Requirements](/start-here/system-requirements/) table lists **4 GB VRAM** and 8 GB RAM as the
minimum at 512×512. On smaller cards, see [Low-VRAM mode](/configuration/optimization/low-vram-mode/).

## Installing

The easiest path is the **SD 1.5** bundle in the Model Manager's starter models. It
installs:

- **Dreamshaper 8** — a popular, versatile main model
- **EasyNegative** — a textual-inversion embedding for the negative prompt
- Three **IP-Adapters** for reference images — Standard, Precise and Face — plus their image encoder
- A set of **ControlNets**: Hard Edge Detection (canny), Inpainting, Line Drawing (mlsd), Depth Map,
  Lighting Detection (Normals), Segmentation Map, Lineart, Lineart Anime, Pose Detection (openpose),
  Contour Detection (scribble), Soft Edge Detection (softedge), Remix (shuffle) and Tile
- **SwinIR**, a 4× upscaler used by the Upscale tool

Other SD 1.5 starter models can be installed individually:

| Starter model | Notes |
| --- | --- |
| CyberRealistic v4.1 | Photorealistic; installs its companion **CyberRealistic Negative v3** embedding |
| ReV Animated | Fantasy and anime styles |
| Dreamshaper 8 (inpainting) | Inpainting version of Dreamshaper 8 |
| Deliberate v5 | Versatile general-purpose model |
| Deliberate v5 (inpainting) | Inpainting version of Deliberate v5 |
| QRCode Monster v2 (SD1.5) | ControlNet for scannable, stylized QR codes |
| T2I-Adapters: Hard Edge Detection (canny), Sketch, Depth Map | Lighter-weight alternatives to ControlNet |

SD 1.x and 2.x main models are self-contained: the UNet, text encoder and VAE all come in one install.
Both formats are supported:

- **Diffusers** folders (for example a HuggingFace repo ID such as `Lykon/dreamshaper-8`).
- **Single-file checkpoints** (`.safetensors` / `.ckpt`), the format most community fine-tunes are
  published in.

No HuggingFace token is needed for the starter models.

If a model was identified incorrectly, open it in the Model Manager and correct its **Variant** or
**Prediction Type**.

## Generation settings

Selecting an SD 1.5 model applies these defaults, which you can override globally or per model in the
Model Manager's **Default Settings**:

| Setting | SD 1.5 default |
| --- | --- |
| Width × Height | 512 × 512 |
| Steps | 30 |
| CFG Scale | 7 |
| Scheduler | Euler Ancestral |

- **Resolution** — dimensions must be multiples of 8. An image size equivalent to 512×512 in pixel
  count is recommended. For larger output, use the Upscale tool, or try
  [HiDiffusion](/users-guide/image-generation/advanced-topics/hidiffusion/).
- **Negative prompt** — fully supported.
- **Schedulers** — the full standard list (DDIM, DPM++ 2M Karras, Euler, Euler Ancestral, UniPC, and
  so on) is available.
- **Prompt syntax** — SD prompts support [Compel weighting and prompt functions](/users-guide/image-generation/advanced-prompt-syntax/#attention-weighting).

The **Advanced** section of the Generate panel adds these SD-specific controls:

- **CLIP Skip** — skips the last layers of the text encoder (up to 12 for SD 1.5). Some fine-tunes,
  especially anime models, are designed to be used with it.
- **CFG Rescale** — intended for models trained with zero-terminal SNR; a value around `0.7` is
  suggested for those. Leave it at 0 for ordinary models.
- **Seamless Tiling** — generates images that tile without visible seams, on the X axis, Y axis or both.
- **VAE** and **VAE Precision** — override the model's bundled VAE, and choose fp32 or fp16 for
  encoding and decoding. VAE Precision is fp32 unless the model's default settings in the Model
  Manager say fp16.
- **HiDiffusion** — improves structure at higher resolutions. See
  [HiDiffusion](/users-guide/image-generation/advanced-topics/hidiffusion/).

## Inpainting models

Models with an inpainting variant (such as **Dreamshaper 8 (inpainting)** and
**Deliberate v5 (inpainting)**) are detected automatically. They are made for filling masked areas and
extending images on the Canvas, and usually blend new content into the surrounding image better than a
normal model does. The **Inpainting** ControlNet is an alternative that works with ordinary SD 1.5
models.

## LoRAs and embeddings

- **LoRAs** for SD 1.x (and SD 2.x), including LyCORIS variants, are supported. LoKr adapters may store
  either Kronecker operand whole or as a factorized `_a` / `_b` pair. Add LoRAs in the LoRA section of the
  Generate panel. LoRAs made for a different model family are not applied.
- **Textual-inversion embeddings** are supported in the positive and negative prompts. Type `<` in a
  prompt box to pick from the compatible embeddings you have installed — for example, `EasyNegative`
  in the negative prompt.

## Control layers

SD 1.5 supports **ControlNet** and **T2I-Adapter** models as control layers on the Canvas. A control
layer guides composition from an image — edges, depth, pose, line art and so on. See
[Adding a gallery image to a layer](/users-guide/canvas/introduction/#adding-a-gallery-image-to-a-layer) for creating a control layer from an image.

## Reference images (IP-Adapter)

SD 1.5 supports up to **five reference images**, using IP-Adapter models:

- **Standard Reference** — references an image with looser precision.
- **Precise Reference** — references an image more precisely.
- **Face Reference** — a precise reference adapted for faces.

Each reference image has a **Mode** — **Style**, **Composition** or **Both** — and a **Weight**. Under
**Advanced**, you can set the range of steps the reference is active for and, in Style mode, choose a
**Style variant** (Standard, Strong or Precise).

The starter IP-Adapters carry their own image encoder. The **CLIP Vision** selector next to the model
matters only for IP-Adapters installed from a single checkpoint file; if the chosen encoder is not
installed, it is downloaded on first use.

## Regional guidance

SD 1.5 supports **regional guidance** on the Canvas: each region can have its own positive prompt, its
own negative prompt, and its own IP-Adapter reference images.

## Upscaling

The **Upscale** tool supports SD 1.5 main models. It requires a **Tile** ControlNet for SD 1.5 (the
starter **Tile** ControlNet is part of the bundle), and the bundle's **SwinIR** model can serve as the
upscaler.

## Workflows

The workflow library includes **Text to Image - SD1.5** and **MultiDiffusion SD1.5**. In the workflow
editor, SD 1.x and 2.x models are loaded with the **Main Model - SD1.5, SD2** node and prompted with
**Prompt - SD1.5**. Other nodes that work with them include **Denoise - SD1.5, SDXL**,
**Apply LoRA - SD1.5**, **ControlNet - SD1.5, SD2, SDXL**, **T2I-Adapter - SD1.5, SDXL**,
**IP-Adapter - SD1.5, SDXL**, **Apply CLIP Skip - SD1.5, SDXL** and **Apply Seamless - SD1.5, SDXL**.

## SD 2.x

SD 2.x models (SD 2.0 and 2.1 and their fine-tunes) are supported for models you install yourself,
from a Diffusers folder or a single-file checkpoint. There are no SD 2.x starter models.

- **Defaults** — selecting an SD 2.x model applies 768×768, 30 steps, CFG 7 and Euler Ancestral. That
  suits the 768px v-prediction checkpoints. For the 512px `-base` checkpoints, change the size to
  512×512, or set it as the model's default in the Model Manager.
- **Prediction type** — SD 2.x checkpoints may be epsilon or v-prediction. InvokeAI detects this on
  install from the checkpoint or its scheduler config. If a model's output looks wrong, check its
  **Prediction Type** in the Model Manager.
- **CLIP Skip** goes up to 24, and **CFG Rescale**, **Seamless Tiling**, the negative prompt and
  regional prompts work as for SD 1.5.
- **Not available for SD 2.x on the Canvas**: control layers, reference images (IP-Adapter),
  HiDiffusion and the Upscale tool. In the workflow editor, the **ControlNet - SD1.5, SD2, SDXL** node
  does accept SD 2.x ControlNets.
- **LoRAs and embeddings** made for SD 2.x are supported.