# Stable Diffusion 3.5

Stable Diffusion 3.5 is Stability AI's diffusion-transformer text-to-image family. It encodes prompts
with three text encoders at once: **CLIP-L**, **CLIP-G** and **T5**. It uses a 16-channel VAE. InvokeAI
supports two checkpoints:

- **SD3.5 Medium**: the smaller model and the one InvokeAI's default settings are tuned for.
- **SD3.5 Large**: the larger model.

Both appear in InvokeAI under the **SD 3.x** base.

## License

SD 3.5 Medium and Large are released under the **Stability AI Community License**: free for individuals and organizations with less than $1M in annual revenue, above which an Enterprise license is needed. The downloads are gated: accept the license on the model's HuggingFace page and add your HuggingFace token under **API Keys** in the Model Manager.

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

## Hardware

The Diffusers downloads are about **16 GB** for Medium and **28 GB** for Large, including the text
encoders and VAE. Medium is the better choice on smaller cards. See
[System Requirements](/start-here/system-requirements/) for general guidance, and the
[Low VRAM guide](/configuration/optimization/low-vram-mode/) if you run out of memory.

## Installing

Both models are **starter models** in the Model Manager. Search for **SD3.5 Medium** or **SD3.5 Large**.
This family has no starter bundle.

| Starter model | Source                                   | Size   |
| ------------- | ---------------------------------------- | ------ |
| SD3.5 Medium  | `stabilityai/stable-diffusion-3.5-medium` | ~16 GB |
| SD3.5 Large   | `stabilityai/stable-diffusion-3.5-large`  | ~28 GB |

SD3 models install in **Diffusers** format only. Each Diffusers folder contains everything the model
needs: the transformer, both CLIP encoders, the T5 encoder and the VAE. InvokeAI does not recognize
single-file or GGUF SD3 main models.

:::note[Gated downloads]
If the download fails with a permission error, open the model's HuggingFace page, accept its license
terms, and add your HuggingFace token in the Model Manager's API Keys section. See
[Model Manager](/users-guide/models/introduction/#external-models) for details.
:::

### Optional component overrides

By default, every component comes from the main model. In the **Components** section next to the model,
you can replace any of them with a model you have installed separately:

- **T5 Encoder**
- **CLIP L**
- **CLIP G**
- **VAE**: only SD3 VAEs are offered

The T5 encoder can also be one of the single files ComfyUI uses: `t5xxl_fp16`, or `t5xxl_fp8_e4m3fn_scaled` for
about half the VRAM. A single-file T5 run on the CPU is held in float32 (about 19 GB of RAM for
T5-XXL).

The CLIP encoders can also be the single files ComfyUI uses, `clip_l` and `clip_g`; SD 3 then still tokenizes
with the main model's tokenizers.

## Generation settings

When you select an SD3 model, InvokeAI applies these defaults, which come from the SD3.5 Medium
reference settings:

| Setting    | Default     |
| ---------- | ----------- |
| Steps      | 40          |
| CFG Scale  | 4.5         |
| Resolution | 1024 × 1024 |

- **CFG Scale** applies true classifier-free guidance. Stability's reference setting for **Large** is
  about **28 steps at CFG 3.5**. Use that as a starting point when you switch to Large.
- **Width and height** must be multiples of **16**.
- **Negative prompts** are supported and always used.
- **No scheduler choice.** SD3 always uses its own flow-matching sampler, so the Scheduler control is
  hidden.
- **Prompt weighting** (Compel syntax) does not apply to SD3. See [Prompt Syntax](/users-guide/image-generation/advanced-prompt-syntax/#attention-weighting).

## Canvas

On the Canvas, SD3 supports text-to-image, image-to-image, inpainting and outpainting. ControlNet,
reference images and regional guidance are **not** available for SD3.

## LoRA

SD 3 and SD 3.5 LoRAs are supported in the layouts trainers publish them in: the diffusers DreamBooth script (a folder
with `pytorch_lora_weights.safetensors`, or the file itself), ai-toolkit, PEFT adapters, SimpleTuner's LyCORIS files
(LoKR) and kohya sd-scripts (`lora_unet_joint_blocks_…`, or `lora_unet_transformer_blocks_…` for trainers that keep
diffusers' module names). A LoRA that also trains the CLIP-L and CLIP-G text encoders applies to them too; T5 layers
in a LoRA are skipped with a warning, because the T5 encoder takes no LoRA. DoRA files, and files with layers that
match no SD 3 module, are refused when you install them, as are files that mix SD 3 Medium and Large widths, name
blocks beyond the model's depth, or name only SD 3 modules without one complete layer to tell the size by.

- **Medium and Large.** A LoRA is trained for one size: SD 3 Medium and SD 3.5 Medium share one, SD 3.5 Large (and
  Large Turbo) the other. Invoke records the size of LoRAs and of SD 3 models when you install them, and the concept
  picker only offers LoRAs that fit the selected model. SD 3 models installed with an earlier version carry no size;
  the picker then offers every SD 3 LoRA. To add it, select the model in the Model Manager and use **Re-identify**
  (which also resets the model's default settings). A LoRA that does not fit is refused when the generation starts,
  before the transformer is moved to the GPU. Text-encoder work, including CLIP LoRAs, can already have run by then.
- **SD 3.5 Medium LoRAs on SD 3 Medium.** SD 3.5 Medium adds a second attention to its first 13 blocks. A LoRA that
  trains it does not fit SD 3 Medium and is refused there.
- **FP8 Storage.** With FP8 Storage the LoRA is applied alongside the weights instead of being merged into them; it then
  stays in VRAM for the whole generation, and InvokeAI reserves room for it.

Earlier versions identified some SD 3 LoRAs as Qwen-Image LoRAs and others not at all. Select such a LoRA in the Model
Manager and use **Re-identify** to fix it. If the file is one Invoke now refuses (for example a LoRA with layers that
match no SD 3 module, or an incomplete layer), Re-identify reports the reason and the model stays as it was; remove it
or install a corrected file.

## PiD super-resolution decode

SD3 works with **PiD** decoding. PiD replaces the VAE decode with a 4× super-resolving pixel diffusion
pass. Two decoders are available as starter models: **PiD Decoder SD3 (2K)** and **PiD Decoder SD3
(2K to 4K)**. Each is about 5 GB and installs the shared Gemma-2 caption encoder as a dependency. See
[PiD Super-Resolution Decode](/users-guide/image-generation/advanced-topics/pid-decode/).

## Workflow editor

The SD3 nodes are **Main Model - SD3**, **Prompt - SD3**, **Denoise - SD3**, **Image to Latents - SD3**
and **Latents to Image - SD3**. **Latents to Image - SD3 + PiD (4x SR)** decodes with PiD instead of the
VAE.

The **T5Encoder** input of **Prompt - SD3** is optional. SD3 was trained with text-encoder dropout, so
leaving T5 unconnected still produces images and saves the time and memory of loading T5. When T5 is connected, it reads at most **256 tokens** of the
prompt.