AI Image Prompt Guide: Master Z-Image, Qwen & LoRA Prompting
Model-specific prompting techniques, optimal settings, example prompts, and honest limitations for every Imagera generation mode
Every AI model interprets prompts differently. A prompt that works on Z-Image may fail on Qwen T2I — and Qwen Edit needs an entirely different approach. This guide covers the exact techniques, settings, and prompt formulas that work for each of Imagera's generation modes, plus what each model genuinely struggles with so you waste fewer credits.
Why Model-Specific Prompting Matters
Most AI image guides give generic advice. The problem: Z-Image, Qwen T2I, and Qwen Edit are fundamentally different models with different architectures, training data, and optimal parameters. A prompt engineered for one model can produce garbage on another. This guide teaches you to prompt each model the way it was designed to work — and tells you honestly when a model is the wrong tool for the job.
Step-by-Step Guide
Start with a Clear, Specific Subject
Begin every prompt with your main subject. Be specific about who or what you want to see. "A woman" gives the AI too many options. "A professional businesswoman in her 30s with auburn hair and glasses" gives it a clear target. The more specific your subject, the fewer regeneration attempts you'll need.
Pro Tips
- • Use specific descriptors: age, gender, features, expression, pose
- • Name the subject's action: "standing", "walking through", "looking at camera"
- • Include distinguishing details: "wearing a navy blazer", "holding a coffee cup"
- • For objects: specify material, size, condition — "a weathered leather journal" vs "a book"
Define the Setting and Environment
The environment provides context, mood, and atmosphere. "Modern Tokyo street at night with neon reflections on wet pavement" tells the AI exactly what you want. "City" leaves everything to chance. Include location, time of day, weather, and key environmental elements.
Pro Tips
- • Be specific about location: "a sunlit Italian cafe terrace" vs "outside"
- • Include atmospheric details: fog, rain, golden hour light, overcast sky
- • Describe the relationship between subject and environment
- • Add depth cues: foreground elements, background details, middle ground
Specify the Visual Style
Style is the single most impactful element after subject. Without a style directive, the model defaults to whatever it saw most in training data. Reference specific styles: "editorial fashion photography", "cinematic film still", "Studio Ghibli anime", "dark fantasy digital painting". The more specific, the more consistent.
Pro Tips
- • Photography styles: "editorial", "street photography", "product shot", "portrait"
- • Art styles: "oil painting", "watercolor", "concept art", "digital illustration"
- • Reference known aesthetics: "Wes Anderson color palette", "cyberpunk neon"
- • Combine styles for unique looks: "cinematic portrait with film grain and muted tones"
Add Lighting and Technical Details
Lighting separates amateur prompts from professional ones. "Golden hour backlighting with rim light on hair" produces dramatically different results than no lighting instruction at all. Include camera details for photorealistic work: lens type affects perspective, aperture affects depth of field.
Pro Tips
- • Lighting keywords: "golden hour", "rim lighting", "studio softbox", "dramatic shadows", "Rembrandt lighting"
- • Camera lens: "85mm portrait lens" (compressed), "35mm" (natural), "wide-angle" (dramatic)
- • Depth of field: "shallow depth of field", "bokeh background", "everything in focus"
- • Camera angle: "low angle hero shot", "overhead flat lay", "eye level"
Use Composition and Mood Keywords
Explicitly stating composition prevents the AI from defaulting to centered, symmetrical arrangements. Mood keywords influence color grading, atmosphere, and emotional tone. These subtle additions make the difference between a technically correct image and one that feels intentional.
Pro Tips
- • Composition: "rule of thirds", "negative space on the left", "leading lines", "symmetrical"
- • Mood: "melancholic", "energetic", "peaceful", "tense", "whimsical"
- • Color: "warm tones", "desaturated", "high contrast", "pastel palette"
- • Texture: "grainy film texture", "smooth gradient", "sharp details"
Use Negative Prompts to Remove Unwanted Elements
Negative prompts tell the AI what to avoid. They're essential for preventing common issues like distorted anatomy, unwanted styles, or quality problems. Think of negatives as guardrails — they won't make a bad prompt good, but they prevent a good prompt from going wrong.
Pro Tips
- • Anatomy fixes: "deformed hands", "extra fingers", "distorted face", "extra limbs"
- • Quality guards: "blurry", "low resolution", "jpeg artifacts", "noise"
- • Style exclusions: "cartoon", "anime", "3D render" (when you want photorealism)
- • Keep negatives focused — a huge list of negatives has diminishing returns
The Prompt Formula
Use this structure for consistent results across all models: [Subject] + [Setting] + [Style] + [Lighting] + [Composition/Mood] + [Technical Details]. Not every prompt needs all six elements — a product shot might not need composition keywords, and a portrait might not need environmental details. But covering at least Subject + Style + Lighting produces reliably good results.
- Subject: Who or what — be specific with descriptors
- Setting: Where and when — location, time, atmosphere
- Style: Artistic approach — photography type, art style, aesthetic
- Lighting: Light source, direction, quality — the biggest impact multiplier
- Composition/Mood: Arrangement, emotion, color palette
- Technical: Camera settings, resolution keywords, quality modifiers
Z-Image Text-to-Image — Prompting Guide
Z-Image (z_t2i) is Imagera's highest-quality generation mode. It uses a custom architecture optimized for detail and coherence. Z-Image responds well to structured, descriptive prompts with clear style direction. It excels at photorealistic portraits, detailed scenes, and LoRA-driven generation.
- Optimal steps: 8 (the "magic number" — going beyond 12 can "burn" the image)
- Optimal CFG: 1.0 (range 0.0-2.0 — much lower than SDXL's typical 7-9)
- Resolution: Full HD (1920x1080, 1080x1920, 1080x1080)
- Keep prompts descriptive and specific — Z-Image handles natural language well
- Include quality keywords: "highly detailed", "sharp focus", "professional"
- Works well with LoRAs — use up to 4 chained LoRAs for combined styles
- Cost: 5 credits per generation
Z-Image Image-to-Image — Prompting Guide
Z-Image I2I (z_i2i) transforms an existing image based on your prompt. The key parameter is denoise: at 0.5 (optimal), you get a balanced transformation that keeps the structure of your input while applying your prompt's style and modifications. Higher denoise = more creative freedom but less fidelity to the original.
- Optimal denoise: 0.5 for balanced transformation (range 0.1-1.0)
- Low denoise (0.1-0.3): Subtle style changes, keeps most of the original
- High denoise (0.7-1.0): Major transformation, less original structure preserved
- Optimal steps: 8, CFG: 1.0 — same as T2I
- Prompt should describe the DESIRED OUTPUT, not the input image
- Works great for style transfer: upload a photo, prompt with an art style
- Cost: 5 credits per generation
Qwen Text-to-Image — Prompting Guide
Qwen T2I (qwen_t2i) is a 20-billion parameter model with native text rendering capability and up to 4K resolution output. It supports Lightning mode for 4-step turbo generation and has 95%+ compatibility with CivitAI LoRAs. Qwen T2I handles natural language prompts well and can render readable text in images — something most AI models fail at.
- Standard mode: 25 steps, CFG 1.0 — high quality, slower
- Lightning mode: 4 steps — fast iterations, good for experimenting
- CFG range: 1.0-5.5 (higher CFG = more prompt adherence, but artifacts above 5.0)
- Resolution: Up to 4K (3840x2160) — much higher than Z-Image
- Text rendering: Use double quotes around text you want in the image: "STOP", "SALE"
- LoRA support: Up to 4 LoRAs chained, 100,000+ CivitAI models compatible
- Lightning LoRA is included by default (strength 0.8)
- Cost: 5 credits (standard), varies for lightning
Qwen Edit — Prompting Guide (Different Technique)
Qwen Edit (qwen_edit) is fundamentally different from T2I modes. Instead of describing an image you want to create, you describe what you want to CHANGE about an existing image. Think of it as giving instructions to a photo editor: "Change the background to a sunset beach" or "Make the dress red instead of blue". Qwen Edit can combine 1-3 images with a text instruction.
- Standard: 20 steps, CFG 2.5. Lightning: 4 steps, CFG 1.0
- Prompt style: INSTRUCTIONS, not descriptions — "Change X to Y", "Add Z to the scene"
- Optimal prompt length: 50-200 characters — be concise and specific
- Specify what to KEEP and what to CHANGE — "maintain the original pose, change the outfit"
- For complex edits: break into multiple steps — structure first, then details
- Direction is from VIEWER perspective — "move head left" means viewer's left
- Supports bilingual editing (English + Chinese) with accurate text preservation
- Cost: 15 credits (standard), 10 credits (lightning)
Qwen Edit — 7 Specialized Editing Tools
Qwen Edit powers 7 specialized sub-features, each optimized for a specific editing task. All use the same underlying model but with prompts tailored to their use case:
- AI Background Studio: "Replace the background with [scene description]" — works for product shots, portraits
- AI Angle Changer: "Show this person from a 45-degree side profile" — rotates subjects to new angles
- AI Look Book: "Change the outfit color to emerald green" — recolors and swaps clothing designs
- Photo Style Transfer: "Transform this into a Studio Ghibli anime style" — applies artistic styles
- AI Product Photography: "Place this product on a marble countertop with soft studio lighting"
- Smart Detail Enhancer: "Remove flyaway hair and clean up skin details" — fine detail refinement
- AI Text Editor: "Change the sign text to OPEN" — modifies text in images while preserving font style
LoRA Prompting — How Custom Models Change Everything
LoRAs (Low-Rank Adaptations) are small trained model additions that modify how the base model interprets prompts. When you add a LoRA, it shifts the model's "vocabulary" — a photorealistic LoRA makes everything more photo-like, a style LoRA applies a consistent aesthetic. Understanding this changes how you prompt.
- LoRAs handle STYLE — your prompt should focus on CONTENT (subject, scene, composition)
- Don't fight the LoRA: if using a "Boring Reality" photorealistic LoRA, don't prompt for "anime style"
- LoRA strength (0.1-1.0) controls influence — start at 0.7 and adjust
- Up to 4 LoRAs can be chained — use different strengths to balance their influence
- CivitAI compatibility: 95%+ of CivitAI LoRAs work — paste URL directly
- Check LoRA trigger words — many LoRAs require specific activation keywords in your prompt
- If a LoRA isn't working, check the model page for recommended settings and trigger words
What AI Image Models Are NOT Good At — Honest Limitations
No AI model is perfect. Knowing what models struggle with saves you credits and frustration. These limitations apply across Z-Image and Qwen models, though to varying degrees:
- Hands and fingers: Still the biggest weakness — extra fingers, merged digits, impossible poses. Use "perfect hands, five fingers" in prompt and "deformed hands, extra fingers" in negatives. Even then, expect occasional issues.
- Complex text: While Qwen has better text rendering than most models, long sentences or unusual fonts still fail. Short text in quotes ("SALE", "STOP") works best. Paragraphs of text will produce gibberish.
- Counting and quantities: Asking for "exactly 3 birds" may produce 2, 4, or 5. AI models struggle with precise counting. Specify and hope for the best, or generate and regenerate.
- Complex multi-person scenes: More than 2-3 people in a scene dramatically increases error rates — merged bodies, extra limbs, inconsistent lighting. Keep scenes simple for best results.
- Teeth and ears: Often too small, overcrowded, pointed, or asymmetrical — especially in close-up portraits. Include "natural teeth, detailed ears" in prompts.
- Physics and shadows: Shadows may point in impossible directions. Objects may float. Gravity is optional. Don't rely on AI for physically accurate scenes.
- Very specific poses: "Person doing a handstand while holding a cup" will likely fail. Complex action poses with object interactions are beyond current models.
- Iterative quality loss (Qwen Edit): Each sequential edit degrades quality slightly. For major changes, do it in as few steps as possible rather than many small edits.
Qwen Edit — Known Limitations
Qwen Edit is powerful but has specific technical limitations you should be aware of before spending credits:
- Aspect ratio shifts: Output images may have slight zoom or crop differences from input. Non-square aspect ratios (e.g., 832x1216) produce more consistent results than square (1024x1024).
- Square resolution issues: 1024x1024 and similar square outputs can produce hallucinations, washed-out colors, or incoherent backgrounds. Use non-square ratios when possible.
- Background changes: The model may modify parts of the image you didn't ask it to change. Be explicit about what should stay the same: "keep the background exactly as is".
- CFG sensitivity: Keep CFG at 2.5 for standard mode. Values above 5.0 introduce artifacts. This is lower than what most diffusion model users expect.
- Complex text characters: Obscure or complex characters may not render correctly in a single step. Use multi-step editing for difficult text.
- Progressive degradation: Repeated edits on the same image reduce quality. Plan your edits to minimize iteration count.
15 Example Prompts That Actually Work
Here are proven prompts for each Imagera generation mode, organized by use case. Copy, modify, and use these as starting points for your own work:
- Z-Image Portrait: "Professional headshot of a confident woman in her 40s, silver hair, minimal makeup, navy blazer, neutral gray studio background, soft diffused lighting, 85mm lens, shallow depth of field, sharp focus"
- Z-Image Landscape: "Misty mountain valley at sunrise, pine forest in foreground, golden light breaking through clouds, dramatic volumetric fog, wide-angle lens, National Geographic style photography, 8k detail"
- Z-Image Product: "Premium wireless headphones on dark slate surface, dramatic side lighting, soft gradient background, product photography, macro detail on texture and materials, clean commercial aesthetic"
- Z-Image Fantasy: "Ancient library with floating books and glowing runes, warm candlelight, dust particles in light beams, concept art style, rich warm color palette, detailed environment design"
- Z-Image with LoRA: "1girl, elegant evening gown, rooftop terrace at sunset, city skyline background, wind in hair, golden hour backlighting, fashion editorial photography" (with Boring Reality LoRA at 1.0)
- Qwen T2I Photorealistic: "Documentary-style photograph of a street food vendor in Bangkok, steam rising from wok, warm tungsten light mixing with blue neon signs, candid moment, 35mm film grain, authentic atmosphere"
- Qwen T2I with Text: "Modern coffee shop storefront with a neon sign reading \"OPEN LATE\", brick exterior, warm interior glow through windows, rainy evening, reflections on wet sidewalk, urban photography"
- Qwen T2I Artistic: "Ethereal underwater portrait of a dancer, flowing fabric creating abstract shapes, dappled light from above, deep blue-green color palette, fine art photography, surreal and dreamlike atmosphere"
- Qwen T2I 4K Detailed: "Macro photograph of morning dew drops on a spider web, each droplet reflecting the sunrise, extreme close-up detail, bokeh background of wildflower meadow, natural light, 4K resolution"
- Qwen T2I Lightning (fast): "Cute robot character sitting in a rainy cafe, anime style, warm interior lighting, window reflections, cozy atmosphere" (4 steps — great for quick iterations)
- Qwen Edit Background: "Replace the background with a tropical beach at sunset with palm trees and calm ocean" (input: portrait photo)
- Qwen Edit Style Transfer: "Transform this photograph into a Studio Ghibli anime style illustration, maintain the original composition and poses" (input: group photo)
- Qwen Edit Outfit Change: "Change the person's outfit to a formal black tuxedo with bow tie, keep the same pose and expression" (input: casual photo)
- Qwen Edit Text: "Change the text on the sign to read \"IMAGERA\" in the same font style and color" (input: image with sign)
- Qwen Edit Product Staging: "Place this product on a clean white marble surface with soft studio lighting and subtle shadow, professional product photography look" (input: product cutout)
Quick Settings Reference
Copy these optimal settings for each mode. These are the defaults that work best in most situations — adjust only when you have a specific reason:
- Z-Image T2I: Steps 8, CFG 1.0, Denoise 1.0 — Resolution: Full HD (1920x1080)
- Z-Image I2I: Steps 8, CFG 1.0, Denoise 0.5 — Lower denoise = more original preserved
- Qwen T2I Standard: Steps 25, CFG 1.0 — Resolution: up to 4K
- Qwen T2I Lightning: Steps 4, CFG 1.0 — Fast iterations, good for experiments
- Qwen Edit Standard: Steps 20, CFG 2.5, Shift 1.0 — For precise edits
- Qwen Edit Lightning: Steps 4, CFG 1.0, Shift 3.1 — Fast preview of edits
- LoRA Strength: Start at 0.7, adjust up (stronger effect) or down (subtler)
- Negative prompts: "blurry, distorted, deformed hands, extra fingers, low quality, watermark"
Common Mistakes That Waste Credits
These are the most frequent prompting mistakes we see. Each one wastes credits and produces disappointing results:
- Being too vague: "pretty landscape" gives unpredictable results — add specifics about location, time, style
- Overloading prompts: Cramming 20 concepts into one prompt confuses the model — focus on 3-5 key elements
- Conflicting instructions: "bright sunny night scene" or "minimalist maximalist" — the AI can't resolve contradictions
- Wrong model for the job: Using Z-Image T2I when you need Qwen Edit (or vice versa)
- Ignoring negative prompts: Skipping negatives when doing portraits (hands/face issues)
- Too many steps on Z-Image: Going beyond 12 steps "burns" the image — stick to 8
- High CFG on Qwen Edit: CFG above 5.0 introduces artifacts — use 2.5 for standard mode
- Expecting perfect text: Only short, simple text works reliably — don't try paragraphs
- Square resolution on Qwen Edit: Use non-square aspect ratios for better coherence
- Fighting the LoRA: Prompting for a style that contradicts the loaded LoRA's training
Expert Tips for Better Results
Frequently Asked Questions
How long should my AI image prompt be?
For Z-Image and Qwen T2I: 30-100 words works best. Include subject, style, lighting, and composition — but don't pad with unnecessary words. For Qwen Edit: 50-200 characters (much shorter) — it's an instruction, not a description. Too-long prompts dilute the model's attention and often produce worse results than concise, specific ones.
What is the difference between Z-Image and Qwen T2I?
Z-Image produces the highest quality at Full HD resolution with optimal results at just 8 steps and very low CFG (1.0). It's best for detailed portraits, landscapes, and LoRA-driven generation. Qwen T2I is a 20B parameter model that supports up to 4K resolution, has Lightning mode for fast 4-step generation, can render text in images, and has broader CivitAI LoRA compatibility (95%+). Choose Z-Image for maximum quality, Qwen T2I for higher resolution or text rendering needs.
How do I use Qwen Edit vs T2I modes?
T2I (text-to-image) creates images from scratch based on a descriptive prompt. Qwen Edit modifies existing images based on instructions. Use T2I when you want to generate something new. Use Qwen Edit when you have an image and want to change specific elements — background, outfit, style, text, or details. Qwen Edit costs 15 credits (vs 5 for T2I) but saves time when you're refining rather than starting over.
Why do my AI-generated images have weird hands?
Hands are the hardest thing for any AI image model to render correctly. The anatomy is complex and highly variable across angles and gestures. To minimize issues: include "perfect hands, five fingers" in your prompt, add "deformed hands, extra fingers, missing fingers" to negative prompts, avoid poses where hands are the focal point, and use Imagera's Face Enhancer tool for post-generation fixes. Even with these techniques, expect occasional failures — this is a fundamental AI limitation across all models.
Can I put text in my AI-generated images?
Qwen T2I has better text rendering than most AI models. For best results: use double quotes around the exact text you want (e.g., "OPEN"), keep text short (1-3 words work best), avoid complex fonts or long sentences, and include the text early in your prompt. Z-Image has more limited text rendering ability. Neither model can reliably produce paragraphs of readable text — this is a current limitation of all AI image generators.
What are LoRAs and how do I use them with prompts?
LoRAs are small trained model additions that modify the base model's style, character, or aesthetic. They handle STYLE while your prompt handles CONTENT. When using a LoRA: focus your prompt on subject, scene, and composition (the LoRA handles the look). Check the LoRA's model page for trigger words — many require specific keywords. Start with strength 0.7 and adjust. You can chain up to 4 LoRAs simultaneously for combined effects. Imagera supports 100,000+ LoRAs from CivitAI — just paste the URL.
How can I make my AI images more photorealistic?
Use technical photography terms: camera type ("DSLR", "mirrorless"), specific lenses ("85mm f/1.4"), lighting setups ("natural window light", "golden hour"), and quality modifiers ("RAW photo", "8k detail"). Add "realistic skin texture, natural pores" for portraits. Use the "Boring Reality" LoRA at strength 1.0 on Z-Image. Add negative prompts: "cartoon, illustration, painting, CGI, unrealistic, plastic skin". Qwen T2I at high resolution (4K) with standard mode (25 steps) produces the most detailed photorealistic output.
Why does Qwen Edit change parts of my image I didn't ask it to?
This is a known Qwen Edit behavior. The model sometimes modifies background, facial details, or other elements you intended to keep. To minimize this: be explicit about what should stay the same ("keep the background, lighting, and expression exactly as is, only change the outfit"), use lower CFG values (2.5 is optimal — higher values make this worse), and prefer non-square aspect ratios which produce more consistent results. For critical edits, do a single targeted change rather than multiple modifications in one prompt.
What settings should I use for the best results?
For Z-Image: 8 steps, CFG 1.0, denoise 1.0 (T2I) or 0.5 (I2I). Never exceed 12 steps — images "burn". For Qwen T2I Standard: 25 steps, CFG 1.0. For Qwen T2I Lightning: 4 steps, CFG 1.0 — great for fast iterations. For Qwen Edit Standard: 20 steps, CFG 2.5. For Qwen Edit Lightning: 4 steps, CFG 1.0. LoRA strength: start at 0.7. These are Imagera's tested optimal defaults — deviate only when you have a specific reason.
How do I fix problems with my generated images?
For hand/face issues: use Face Enhancer or regenerate with stronger negative prompts. For wrong style: adjust your style keywords or add/remove LoRAs. For blurry output: check your resolution settings and ensure steps aren't too low. For Qwen Edit not following instructions: make your instruction more specific and explicit, reduce prompt length, and try non-square aspect ratios. For burned/overcooked Z-Image: reduce steps to 8 (never more than 12). For inconsistent Qwen Edit backgrounds: add "maintain original background unchanged" to your instruction.
Related Guides
LoRA Training Guide
Learn how to train custom LoRAs for consistent character and style generation — no GPU required
Image Upscaling Guide
Master AI upscaling to enhance image resolution to 4K without quality loss
Style Transfer Guide
Apply artistic styles to your images — Ghibli, oil painting, watercolor, and more
Video Enhancement Guide
Upscale and enhance video quality with AI — no download required