Stable Diffusion 3.5 Medium: What's Different From SDXL and Flux

Stable Diffusion 3.5 Medium is a 2 billion parameter text-to-image model from Stability AI, built on a Multimodal Diffusion Transformer (MMDiT-X) architecture. It installs through Diffusers or ComfyUI like this site's other Stable Diffusion coverage, but its license differs: free under $1 million in annual revenue, Enterprise License required above that.

How it differs from SDXL, SD1.5, and Flux

This site already covers SD1.5 and SDXL through ComfyUI. SD 3.5 Medium is not a bigger version of either; it is a newer model family built on the MMDiT-X architecture, the same transformer-based design line as SD3 and Flux, rather than the older U-Net design behind SD1.5 and SDXL. At 2B parameters it sits well under Flux's parameter count and under SD 3.5 Large, which is why Stability AI positions it as the mid-tier option: lighter to run than the Large variant, more capable than SD1.5 on prompt adherence and text rendering.

Hardware: no fixed VRAM number is published

Stability AI does not state a minimum VRAM figure for SD 3.5 Medium on the model card. What the card does document is BitsAndBytes quantization support for cutting memory use on smaller GPUs, covered below. If a full-precision run does not fit your card, quantize before assuming you need more hardware.

Install with Diffusers

The Diffusers library is the fastest way to run SD 3.5 Medium from Python. Install the package, then load the pipeline by its Hugging Face repo id.

import torch
from diffusers import StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-medium",
    torch_dtype=torch.bfloat16,
)
pipe = pipe.to("cuda")

Generate an image

Call the pipeline with a prompt, a step count, and a guidance scale, then save the resulting image. The model card's own example uses 40 steps and a guidance scale of 4.5 as a starting point.

image = pipe(
    "A capybara holding a sign that reads Hello World",
    num_inference_steps=40,
    guidance_scale=4.5,
).images[0]
image.save("capybara.png")

Lower VRAM with BitsAndBytes quantization

For GPUs that cannot hold the full BF16 model, Stability AI documents 4-bit quantization through BitsAndBytes. It loads the transformer component in 4-bit precision and offloads the rest of the pipeline to CPU when idle, which trades some speed for a smaller memory footprint.

from diffusers import BitsAndBytesConfig, SD3Transformer2DModel
from diffusers import StableDiffusion3Pipeline
import torch

model_id = "stabilityai/stable-diffusion-3.5-medium"

nf4_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)
model_nf4 = SD3Transformer2DModel.from_pretrained(
    model_id,
    subfolder="transformer",
    quantization_config=nf4_config,
    torch_dtype=torch.bfloat16,
)

pipeline = StableDiffusion3Pipeline.from_pretrained(
    model_id,
    transformer=model_nf4,
    torch_dtype=torch.bfloat16,
)
pipeline.enable_model_cpu_offload()

Running it in ComfyUI

Stability AI's own SD3.5 GitHub repo is a command-line reference implementation, not a ComfyUI package; the repo itself points users to ComfyUI for node-based inference rather than shipping workflow files. In practice that means downloading the SD 3.5 Medium checkpoint from Hugging Face and loading it in ComfyUI the same way you would load an SDXL checkpoint, using the checkpoint loading steps already covered in this site's ComfyUI setup guide.

Licensing: read this before commercial use

SD 3.5 Medium ships under the Stability AI Community License, not a permissive license like MIT or Apache. It is free for individuals and organizations with under $1 million in annual revenue. Above that threshold, you need an Enterprise License from Stability AI, not the free community terms.

FAQ

How much VRAM does Stable Diffusion 3.5 Medium need?

Stability AI has not published a specific VRAM figure on the model card. Use the BitsAndBytes 4-bit quantization path above if a full-precision run does not fit your GPU.

Is Stable Diffusion 3.5 Medium free for commercial use?

Yes, for individuals and organizations under $1 million in annual revenue, under the Stability AI Community License. Above that revenue threshold, an Enterprise License from Stability AI is required.

How is SD 3.5 Medium different from SDXL or Flux?

It uses a different architecture, MMDiT-X, versus SDXL's U-Net design, and sits at 2 billion parameters, smaller than Flux and than SD 3.5 Large. Its license also carries a revenue-based commercial cutoff that does not necessarily apply to the SDXL, SD1.5, or Flux checkpoints covered elsewhere on this site.

Can I run Stable Diffusion 3.5 Medium in ComfyUI?

Yes. Stability AI's own GitHub repo is a command-line reference implementation and points to ComfyUI for node-based inference rather than including its own workflow files, so load the checkpoint in ComfyUI the same way as any other Stable Diffusion model.