Model

Trends

.ai

the best open source ai models

Stable Diffusion 1.5: the classic model, prompted properly

Q: Two heads or repeated bodies

Generating far above 512 on one axis. Generate at 512x768 and upscale with hires fix instead of asking for 1024 directly.

Q: Washed-out, grey images

Missing or weak VAE in some checkpoints. Use a checkpoint with a baked VAE, or load the standard SD 1.5 VAE; raise CFG by one.

Topic: Models

By Captain

Published 2026-10-02

Stable Diffusion 1.5 is still worth knowing: the right resolution, CFG and sampler, how to use its huge library of LoRAs and embeddings, and creative prompt ideas suited to a 512-pixel model.

Overview

Stable Diffusion 1.5

was released by RunwayML and Stability AI in late 2022 and trained at 512x512. It is small, fast and sits at the center of the largest collection of

LoRAs

embeddings

, ControlNets and fine-tunes ever made, thousands of which are in this catalog.

Its raw output is rough compared with newer models, but with a good

negative prompt

, a fine-tuned checkpoint or a couple of embeddings it produces distinctive images quickly, and it runs on almost any GPU.

Reference

Name	Type	What it is
Prompt	tags	Comma-separated descriptors. Weights (word:1.3) and emphasis are well supported.
Negative prompt	text	Essential. lowres, bad anatomy, bad hands, text, watermark, blurry; or an embedding like easynegative.
CFG	6-9	7 is the standard; go to 5 for soft painterly results.
Steps	20-30	Euler a at 25 or DPM++ 2M Karras at 20-25.
Resolution	512-768	512x512, 512x768 or 768x512. Larger sizes duplicate subjects unless you use hires fix.

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Generate at 512x768 (portrait) or 768x512 (landscape),
CFG
7, 25 steps.
Write a negative prompt first and reuse it; it matters more on 1.5 than on any later model.
Add one style LoRA at 0.7 or a style embedding; 1.5 takes LoRAs very well.
Enable hires fix (1.5x-2x, denoise 0.4-0.5) for the final image to reach 1024 on the long edge without doubled heads.
Keep a list of the seeds you like; 1.5 is seed-sensitive and the same seed with small prompt edits is a good way to iterate.

Examples

Classic prompt

portrait of an old lighthouse keeper, weathered face, wool sweater, stormy sky behind, oil painting, dramatic rim light, (highly detailed:1.2)

Negative: lowres, bad anatomy, bad hands, text, watermark, blurry, cropped

Vintage illustration

1950s travel poster of a mountain railway, flat colors, screen print texture, bold typography space at the bottom, teal and orange

Creative idea: tiny worlds

macro photo of a snail shell that is a spiral staircase for tiny glowing people, moss, dew drops, tilt-shift, soft morning light

Tips

Camera words (85mm, f/1.8, bokeh) and medium words (watercolor, linocut) shift 1.5 strongly.
If anatomy breaks, lower CFG by one and add a hand-fixing embedding rather than piling on more negative words.
ControlNet
(openpose, canny, depth) works best on 1.5 of all families; use it when you need a specific pose or layout.
For
inpainting
, pick one of the dedicated 1.5 inpainting checkpoints in this catalog instead of the base model.

Troubleshooting

Two heads or repeated bodies

Why it happens

Generating far above 512 on one axis.

How to fix it

Generate at 512x768 and upscale with hires fix instead of asking for 1024 directly.

Washed-out, grey images

Why it happens

Missing or weak VAE in some checkpoints.

How to fix it

Use a checkpoint with a baked VAE, or load the standard SD 1.5 VAE; raise CFG by one.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.