ChatGPT Stable Diffusion & Flux Prompt Generator Prompt
You are an expert in open-source image generation models including Stable Diffusion XL, SD 3.5, and Flux.1.
Category
🎨 Image AI
Difficulty
Advanced
Models
3
Last Updated
2026-06-28
Works with
📄 Example output
⚠️ Common Mistakes
❓ FAQ
⚙️ Fill in your variables
📋 Prompt
You are an expert in open-source image generation models including Stable Diffusion XL, SD 3.5, and Flux.1.
Concept: [image concept]
Style: [style — photorealistic/illustration/anime/cinematic/painterly/3D render]
Ratio: [aspect ratio — 1:1/16:9/9:16/4:3]
Model: [model version — SDXL/SD 3.5/Flux.1 Dev/Flux.1 Schnell]
Task: Write optimised prompts for [model version]:
1. POSITIVE PROMPT (what to include):
[Subject description], [environment], [style], [lighting], [camera/lens], [quality tags]
Format: comma-separated keywords + short phrases, most important first
2. NEGATIVE PROMPT (what to exclude):
[Unwanted elements]
Standard negative prompt for quality improvement
3. RECOMMENDED SETTINGS:
CFG scale, steps, sampler, ControlNet (if useful)
4. MODEL-SPECIFIC NOTES: What makes this model different and how to leverage it
5. VARIATIONS: 3 prompt variations exploring different interpretations
Concept: [image concept]
Style: [style — photorealistic/illustration/anime/cinematic/painterly/3D render]
Ratio: [aspect ratio — 1:1/16:9/9:16/4:3]
Model: [model version — SDXL/SD 3.5/Flux.1 Dev/Flux.1 Schnell]
Task: Write optimised prompts for [model version]:
1. POSITIVE PROMPT (what to include):
[Subject description], [environment], [style], [lighting], [camera/lens], [quality tags]
Format: comma-separated keywords + short phrases, most important first
2. NEGATIVE PROMPT (what to exclude):
[Unwanted elements]
Standard negative prompt for quality improvement
3. RECOMMENDED SETTINGS:
CFG scale, steps, sampler, ControlNet (if useful)
4. MODEL-SPECIFIC NOTES: What makes this model different and how to leverage it
5. VARIATIONS: 3 prompt variations exploring different interpretations
CONCEPT: Futuristic city at night
STYLE: Cinematic/photorealistic | RATIO: 16:9 | MODEL: Flux.1 Dev
NOTE ON FLUX.1: Flux.1 Dev understands natural language much better than SDXL. Write in complete sentences rather than keyword lists. Flux handles complex spatial relationships and text in images significantly better than SD models.
POSITIVE PROMPT (Flux.1 Dev — natural language):
'A cinematic photorealistic aerial view of a sprawling futuristic megacity at night, rain-slicked streets reflecting neon signs in cyan and magenta, flying vehicles trailing light streaks between glass-and-steel towers, thick fog rolling through the lower levels, dramatic atmospheric perspective with sharp foreground skyscrapers fading to haze in the distance, shot on IMAX camera with anamorphic lens, golden hour mixed with artificial lighting, ultra-detailed architecture, blade runner aesthetic without being derivative, 8K resolution'
NEGATIVE PROMPT:
'cartoon, anime, painting, illustration, daytime, sunshine, cheerful, low quality, blurry, watermark, text, logo, oversaturated, flat lighting'
RECOMMENDED SETTINGS (Flux.1 Dev):
Sampler: Euler or DPM++ 2M
Steps: 20–30
CFG: 3.5–4.5 (Flux needs lower CFG than SDXL)
Aspect: 1344×768 (16:9 for Flux)
SDXL VERSION (keyword-based):
'futuristic megacity, nighttime, rain, neon reflections, cyan magenta lighting, aerial view, flying cars, volumetric fog, atmospheric perspective, cinematic, blade runner aesthetic, ultra detailed, (masterpiece:1.3), 8k, (photorealistic:1.2), anamorphic lens, shallow depth of field'
NEGATIVE SDXL:
'(worst quality:1.4), (low quality:1.4), cartoon, anime, daylight, flat lighting, ugly, deformed, watermark, text'
STYLE: Cinematic/photorealistic | RATIO: 16:9 | MODEL: Flux.1 Dev
NOTE ON FLUX.1: Flux.1 Dev understands natural language much better than SDXL. Write in complete sentences rather than keyword lists. Flux handles complex spatial relationships and text in images significantly better than SD models.
POSITIVE PROMPT (Flux.1 Dev — natural language):
'A cinematic photorealistic aerial view of a sprawling futuristic megacity at night, rain-slicked streets reflecting neon signs in cyan and magenta, flying vehicles trailing light streaks between glass-and-steel towers, thick fog rolling through the lower levels, dramatic atmospheric perspective with sharp foreground skyscrapers fading to haze in the distance, shot on IMAX camera with anamorphic lens, golden hour mixed with artificial lighting, ultra-detailed architecture, blade runner aesthetic without being derivative, 8K resolution'
NEGATIVE PROMPT:
'cartoon, anime, painting, illustration, daytime, sunshine, cheerful, low quality, blurry, watermark, text, logo, oversaturated, flat lighting'
RECOMMENDED SETTINGS (Flux.1 Dev):
Sampler: Euler or DPM++ 2M
Steps: 20–30
CFG: 3.5–4.5 (Flux needs lower CFG than SDXL)
Aspect: 1344×768 (16:9 for Flux)
SDXL VERSION (keyword-based):
'futuristic megacity, nighttime, rain, neon reflections, cyan magenta lighting, aerial view, flying cars, volumetric fog, atmospheric perspective, cinematic, blade runner aesthetic, ultra detailed, (masterpiece:1.3), 8k, (photorealistic:1.2), anamorphic lens, shallow depth of field'
NEGATIVE SDXL:
'(worst quality:1.4), (low quality:1.4), cartoon, anime, daylight, flat lighting, ugly, deformed, watermark, text'
🏆
💡 Pro Tips
Best model for this prompt
ChatGPT
ChatGPT (GPT-4o / GPT-5)
Flux.1 uses natural language sentences — don't use keyword lists with Flux, write it as you'd describe a photograph to a photographer
For SDXL, weight important elements with (element:1.3) syntax — but use sparingly, above 1.4 causes distortion
Steps: 20–25 is usually sufficient for Flux; SDXL typically needs 30–40 for best quality
The negative prompt for quality ('worst quality, low quality, jpeg artifacts') is more important than the subject-specific negative prompts
Using SDXL prompt structure on Flux — they need completely different approaches
Too many brackets and weights in SDXL — it leads to element 'bleeding' and distorted outputs
Putting the most important elements at the END of the prompt — both models weight early tokens higher
Ignoring model-specific resolution — each model has native resolutions it was trained at; using wrong resolutions degrades quality
- What's the difference between Flux.1 Dev and Flux.1 Schnell?Flux.1 Dev: higher quality, slower, requires 20+ steps, best for final output. Flux.1 Schnell: 4-step generation, much faster, good for rapid iteration and prototyping, slightly lower quality. Schnell is non-commercial; Dev requires a license for commercial use.
- Is Stable Diffusion better than Midjourney?Midjourney produces more aesthetically polished results out of the box. Stable Diffusion is open-source, infinitely customisable, runs locally, and has no content restrictions. Many professionals use both: Midjourney for quick creative exploration, SD for fine-tuned production output.
- How do I run Stable Diffusion locally?ComfyUI (most powerful, node-based workflow) and Automatic1111 (most popular, simpler UI) are the two main options. Requires an NVIDIA GPU with 8GB+ VRAM for SDXL, or can run on CPU slowly. Flux.1 needs 16GB+ VRAM for full quality.
- What are ControlNet and LoRA?ControlNet: adds precise control over composition, pose, depth, and edges — lets you match a reference image's structure. LoRA (Low-Rank Adaptation): fine-tuned model add-ons that inject a specific style, character, or concept into generation. Both dramatically expand what's possible.