How to run SD 3.5 locally.

Large, Turbo or Medium.

Stability released SD 3.5 in October 2024. All three versions use three text encoders: CLIP-L, OpenCLIP bigG (CLIP-G) and T5-XXL.

• Large is the 8B model, made for images up to about one megapixel. The best quality of the three.
• Large Turbo is Large distilled to four steps. Same size, much faster, a little less detail.
• Medium is a 2.5B model with a newer layout (MMDiT-X). It handles 0.25 to 2 megapixels and fits smaller cards.

Stability notes that SD 3.5 gives more variation from seed to seed than earlier models, which suits exploring and makes one exact look harder to repeat. For anatomy, the Medium card recommends Skip Layer Guidance, covered under settings.

Files you need.

The easy way: one file

Comfy-Org packs the model, all three text encoders and the VAE into one checkpoint. It goes in models/checkpoints and loads with Load Checkpoint.

• Large sd3.5_large_fp8_scaled.safetensors ComfyUI/models/checkpoints/ model and encoders in fp8
  14.9 GB Download
• Medium sd3.5_medium_incl_clips_t5xxlfp8scaled.safetensors ComfyUI/models/checkpoints/ fp16 model, T5 in fp8
  11.6 GB Download

Both carry an fp8 T5, which doesn’t run on a Mac. See the Mac notes below.

Separate files

Stability’s own files hold the model without the text encoders, so the encoders come separately. This is also the route for Large Turbo and for GGUF.

• Model sd3.5_large.safetensors ComfyUI/models/checkpoints/ gated. Turbo and Medium (5.1 GB) in their own repos
  16.5 GB Download
• Text encoder clip_l.safetensors ComfyUI/models/text_encoders/
  0.2 GB Download
• Text encoder clip_g.safetensors ComfyUI/models/text_encoders/
  1.4 GB Download
• Text encoder t5xxl_fp8_e4m3fn_scaled.safetensors ComfyUI/models/text_encoders/ or t5xxl_fp16.safetensors, 9.8 GB, which a Mac needs
  5.2 GB Download

models/
├── checkpoints/
│   └── sd3.5_large.safetensors
└── text_encoders/
    ├── clip_l.safetensors
    ├── clip_g.safetensors
    └── t5xxl_fp8_e4m3fn_scaled.safetensors

Smaller files: GGUF

city96’s GGUF builds load through the ComfyUI-GGUF nodes and go in models/unet. There is a Large Turbo set with the same sizes. Large has no K-quants; Medium does.

A GGUF holds the model only. You also need the three text encoders and a VAE.

What fits your computer.

At 1024 × 1024. T5-XXL is bigger than Medium itself, and ComfyUI moves it out of the way before sampling, so system RAM matters too: Comfy’s examples ask for 32 GB of RAM or more for the fp16 T5.

• 6 GBTight

  Medium as GGUF, with the fp8 T5 offloaded to system RAM.

• 8 GBFits

  Medium, or Large as GGUF Q4_0 (4.8 GB). One Krita user on an RTX 4060 8 GB found SD 3.5 GGUF about twice as fast as Flux GGUF.

• 12 GBFits

  Medium at full precision, or Large as GGUF Q8_0 (8.8 GB).

• 16 GBFits

  Large fp8 all-in-one. Comfy estimates 14.9 GB.

• 24 GBFits

  Large at full precision with the fp16 T5.

• Mac 16 GBTight

  Medium as GGUF with the fp16 T5. No fp8 files.

• Mac 32 GB+Tight

  Large as GGUF Q8_0 with the fp16 T5, about 20 GB of files together.

• 8 GB graphics cards
• 12 GB cards
• RTX 3060 12 GB
• Mac with Apple Silicon
• AMD Radeon

Set it up.

1. 1

   Update ComfyUI

   Medium needs a ComfyUI from after its release; older versions stop with a size mismatch error. ComfyUI Desktop updates itself. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download

   An all-in-one checkpoint into models/checkpoints, or a model file plus the three encoders. For Stability’s files, accept the licence on Hugging Face first.

3. 3

   Open the template

   In the template browser, SD3.5 Simple loads sd3.5_large_fp8_scaled with one Load Checkpoint node. Swap in the Medium file if you use that.

4. 4

   For separate files, add the encoders

   Keep Load Checkpoint for the model and VAE, and add a TripleCLIPLoader with clip_l, clip_g and the T5. For GGUF, use Unet Loader (GGUF), the TripleCLIPLoader and Load VAE.

5. 5

   For Medium, add Skip Layer Guidance

   Put SkipLayerGuidanceSD3 between the model and the sampler. Stability recommends it for Medium’s structure and anatomy.

6. 6

   Run

   Use EmptySD3LatentImage for the size. The first run loads the T5 and takes a while.

Settings that work.

Large

Steps

28

CFG

3.5

Sampler

euler

Scheduler

sgm_uniform

Size

1024 × 1024

Shift

3 (default)

Negative

yes

Latent

EmptySD3LatentImage

28 steps at 3.5 is from Stability’s Large card. Comfy’s SD3.5 Simple template uses 20 steps at CFG 4, which is faster and close. ComfyUI sets SD 3.5’s shift to 3 on its own.

Large Turbo

Steps

4

CFG

1.2

Sampler

euler

Scheduler

sgm_uniform

ComfyUI’s examples say “set steps to 4 and cfg to 1.2”. Stability’s card uses no guidance at all, which is CFG 1 in ComfyUI.

Medium

Steps

40

CFG

4.5

Guidance

SkipLayerGuidanceSD3

Size

0.25 to 2 MP

From the Medium card. Keep prompts under 256 T5 tokens: the card warns of artefacts at the image edges beyond that.

Without T5

T5 is optional. Load only CLIP-L and CLIP-G with a DualCLIPLoader and SD 3.5 still works, with less memory and worse prompt following and text in images. At low CFG the difference is smaller.

When it goes wrong.

Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.

An fp8 file on a Mac. Use GGUF or full-precision weights and t5xxl_fp16.

size mismatch for joint_blocks.0.x_block.adaLN_modulation.1.weight

ComfyUI is too old for Medium. Update it.

clip missing: ['text_projection.weight']

Harmless. It shows after some updates and the image comes out fine.

Value not in list: clip_name2: 'clip_l.safetensors'

An encoder file is missing from models/text_encoders. Download it and press R.

Black images with a GGUF

Usually a wrong or half-downloaded VAE. On a Mac, Turbo GGUF black images at sizes other than 1024 were a PyTorch bug fixed in 2.6.

Mosaic-like images from Medium on a Mac

In the one report, the answer was a CFG set too high. Go back to 4.5 or lower.

unknown model architecture: 'sd3'

The GGUF was opened in LM Studio or llama.cpp. These files are for ComfyUI-GGUF only.

Out of memory before sampling starts

The T5 encoder. Use the fp8 T5 (not on a Mac), or run without T5.

Questions.

Sources: SD 3.5 Large card, Large Turbo card, Medium card, Stability’s announcement, Community License, ComfyUI SD3 examples, Comfy blog on Medium, Comfy-Org fp8 files, city96 GGUF, ComfyUI issue #12202, Krita AI Diffusion #1328.