How to run SD 1.5 locally.

Still worth running.

Stable Diffusion 1.5 came out in October 2022, made by RunwayML and CompVis with Stability’s compute. It’s a model of about 0.9B parameters with one text encoder, CLIP ViT-L, trained at 512 × 512.

It’s small and quick, it runs on almost any graphics card from the last few years, and it has years of fine-tunes, LoRAs and ControlNets behind it, including anime models like Anything, Counterfeit and MeinaMix. Its model card is honest about the limits: no legible text, weak faces and complex scenes, and English prompts only.

For new work on 8 GB or more, SDXL gives better images at 1024 pixels. SD 1.5 is the pick for 4 to 6 GB cards, old laptops, huge batches, or a fine-tune and LoRA set you already rely on.

Files you need.

One checkpoint, which holds the model, the text encoder and a VAE. A better VAE is optional and small.

• Checkpoint v1-5-pruned-emaonly-fp16.safetensors ComfyUI/models/checkpoints/ or v1-5-pruned-emaonly.safetensors, fp32, 4.3 GB
  2.1 GB Download
• VAE vae-ft-mse-840000-ema-pruned.safetensors ComfyUI/models/vae/ optional: cleaner colours and faces than the built-in VAE
  0.3 GB Download

The fp16 file makes practically the same images as the fp32 one at half the download. The 7.7 GB v1-5-pruned.safetensors carries extra weights for training and isn’t needed to make images. Fine-tunes from Civitai go in the same folder and load the same way.

models/
├── checkpoints/
│   └── v1-5-pruned-emaonly-fp16.safetensors
└── vae/
    └── vae-ft-mse-840000-ema-pruned.safetensors   (optional)

What fits your computer.

A 512 × 512 image at 20 to 25 steps. Comfy estimates its default SD 1.5 workflow at 3.1 GB of video memory.

• 4 GBTight

  Runs. ComfyUI’s own guide puts SD 1.5 at about 4 GB.

• 6 GBFits

  Runs smoothly. A GTX 1660 Super makes a 512 image in under 20 seconds.

• 8 GBFits

  Room for LoRAs, ControlNets, hires passes and batches.

• 12 GB+Fits

  Large batches. At this size SDXL is the better model.

• Mac 8 GB+Fits

  The 2.1 GB file runs on any Apple Silicon Mac. The lightest way to try ComfyUI on one.

• Laptop GPUs
• 8 GB graphics cards
• RTX 3060 Ti
• AMD Radeon
• Mac with Apple Silicon

Set it up.

1. 1

   Open the default workflow

   A fresh ComfyUI opens with an SD 1.5 workflow. In the template browser it’s Image Generation. It expects v1-5-pruned-emaonly-fp16.safetensors.

2. 2

   Download the checkpoint

   Put the file in ComfyUI/models/checkpoints, then press R or restart ComfyUI so Load Checkpoint lists it.

3. 3

   Add the better VAE

   Optional. Add a Load VAE node with vae-ft-mse-840000-ema-pruned and connect it to VAE Decode instead of the checkpoint’s VAE output.

4. 4

   For anime fine-tunes: clip skip

   Models from the NovelAI line (Anything, Counterfeit, MeinaMix and their merges) expect clip skip 2. Put CLIP Set Last Layer at -2 between the checkpoint and the prompt nodes.

5. 5

   Run

   Keep the size at 512 × 512 or close to it, such as 512 × 768. For bigger images, generate at 512 and upscale, or send the result through a second KSampler at partial denoise.

Settings that work.

Steps

20 to 25

CFG

7 to 8

Sampler

euler or dpmpp_2m

Scheduler

normal or karras

Size

512 × 512

Negative

yes

Clip skip

-2 for anime

VAE

vae-ft-mse

Comfy’s default template uses 20 steps, CFG 8, euler and normal. dpmpp_2m with karras at CFG 7 is the other common default. SD 1.5 has no official recommendation beyond that, and fine-tunes often list their own on their download page.

Fewer steps

Speed LoRAs exist for SD 1.5 too, from LCM, Hyper-SD and TCD. Hyper-SD’s SD 1.5 LoRAs run at the step count in their name with ddim and sgm_uniform at CFG 1, like the SDXL ones. The Hyper-SD15-8steps-CFG-lora keeps a real CFG, so negative prompts still work.

When it goes wrong.

Value not in list: ckpt_name: 'v1-5-pruned-emaonly-fp16.safetensors' not in []

The most common first-run error. The default workflow needs the checkpoint in models/checkpoints. Download it, then press R.

Faded, washed-out colours

The VAE inside the checkpoint. Load vae-ft-mse-840000-ema-pruned with a Load VAE node.

Doubled people and repeated objects at large sizes

SD 1.5 was trained at 512 and, as Comfy’s template notes, doesn’t do well at large sizes. Generate at 512 and upscale.

lora key not loaded

The LoRA is for SDXL or another model. SD 1.5 only takes SD 1.5 LoRAs.

mat1 and mat2 shapes cannot be multiplied with a ControlNet

The ControlNet is for SDXL. Use an SD 1.5 ControlNet.

Unsupported head_dim: 160 with --use-sage-attention

SageAttention doesn’t handle one of SD 1.5’s layer sizes. Start ComfyUI without that flag for SD 1.5.

Questions.

Sources: SD 1.5 model card (mirror), Comfy-Org SD 1.5 archive, ComfyUI text to image tutorial, sd-vae-ft-mse, Hyper-SD, ComfyUI issue #8826, #884, #7352.