Tool
STABLE DIFFUSION GENERATOR
Positive + Negative two-field prompt builder: 25 styles / 18 lightings / 18 negative-prompt tags / 5 model variants / 5 samplers / Steps + CFG sliders / 12 one-click presets. Output works in Automatic1111, ComfyUI, Forge, fal.ai, RunPod.
PRESETS · Quick presets
Describe the scene in plain English — subject + composition + environment.
Style00Pick 1–2.
Mood00Pick 1–2.
Colors00Pick 1–2.
Quality / Detail00Pick 2–3. Phrases like 'Highly detailed' / '8k' / 'Masterpiece' are SD-specific quality keywords the models trained on.
Negative prompt00Things to avoid. Every SD UI has a separate Negative prompt field. Pick 3–5.
SDXL 1.0 = general default. SD 3.5 Large = newest. Turbo / Lightning = low-step variants.
DPM++ 2M Karras is the broadest-quality default. LCM is fastest but needs LoRA support.
20–40 is standard. Turbo / LCM models work at 4–8 steps.
304–50
1–20. Photorealism 5–7, illustration 7–9. Above 10 risks oversaturation.
71–20
HOW TO USE
- Type the focal subject (English, the more specific the better)
- Pick style / mood / lighting / color / quality, or '🎲 Randomize all' for a baseline
- Pick negative-prompt tags (3–5 recommended; common: bad anatomy, low quality, watermark)
- Pick model + sampler, adjust Steps / CFG sliders (these aren't embedded in the prompt — they're sidebar controls in your SD UI)
- Paste positive into Automatic1111, ComfyUI, or fal.ai; paste negative into the negative prompt box; mirror model / sampler / steps / cfg in the UI sidebar.
FAQ
Q01Is this generator free? Do I need an account?+
Completely free, runs in your browser (no API calls, nothing stored). No signup — open and use.
Q02How does Stable Diffusion differ from Flux / Midjourney?+
All three are image generators, but SD is the actual open-source ecosystem — you can run it locally (ComfyUI / Automatic1111), download LoRAs / checkpoints for style specialisation, and chain ControlNet for composition control. Flux is also open but SD's surrounding ecosystem is older and deeper. MJ is closed-source SaaS. SD fits advanced users who want self-hosting, fine control, or LoRA workflows.
Q03Why split into Positive and Negative prompt sections?+
SD models learn 'what to draw' AND 'what to avoid'. Negative prompt isn't 'not mentioning' — it's explicitly telling the model to actively avoid. Want clean hands? Put 'deformed hands, extra fingers' in negative; the model actively suppresses them. Flux / MJ don't have this field (different architecture).
Q04How should I set Steps / CFG / Sampler?+
Steps (denoising iterations) 20–40 is standard; Turbo / Lightning models 4–8. CFG Scale (prompt adherence) 5–7 for photorealism, 7–9 for illustration; >10 risks oversaturation. Sampler defaults to DPM++ 2M Karras (broadest quality); LCM is fastest but needs LoRA support.
Q05Which model should I pick — SDXL / SD 3.5 Large / SD 1.5?+
SDXL 1.0 is the current general-purpose default; SD 3.5 Large is Stability's newest (late 2024), best at text-in-image; SD 1.5 is the lightweight classic with the deepest LoRA library on Civitai. Turbo / Lightning are reduced-step variants of SDXL. Start with SDXL 1.0.
Q06Why aren't model / sampler / steps / cfg in the copied string?+
Every SD UI (Automatic1111 / ComfyUI / Forge / fal.ai / RunPod) handles those as sidebar controls, not prompt content. We surface them in the preview as a 'Recommended settings' block so you can mirror them in your platform's UI.
Q07Why ≤ 350 chars recommended for positive prompt?+
SD uses CLIP / OpenCLIP / T5 for text embedding; these weight the first 75 tokens highest, with later tokens fading. Within 350 chars almost everything gets through. To emphasise a token use the (word:1.5) weight syntax (supported in Automatic1111, partial in ComfyUI) — we don't expose this UI yet; you can hand-edit the copied string.
MORE TOOLS