About Text-to-Image Generation
Text-to-image generation lets you create custom images directly inside the Abyssale Design Editor by describing what you want to see. The quality of your prompt is the single biggest factor in the quality of your result, this article walks you through how to write prompts that consistently produce strong images.
Accessing Text-to-Image Generation
Select any image layer on your canvas
An AI toolbar appears at the bottom of the canvas, click Text to image
The toolbar transforms into the AI Text to Image prompt bar, where you'll write your prompt, pick a model, and generate
This is also where the tips in this article come into play most directly: the prompt field is where structure and detail matter, and Abyssale gives you two built-in helpers right above it:
Suggestion tags (e.g. Enhance lighting, Improve text, Refine branches, Boost color) : one-click additions that cover some of the same ground as the "Key Elements" below, without you having to type them out
Enhance : rewrites your prompt for clarity and detail while keeping its substance and direction intact. It's a good first pass, but it works best on a prompt that already has a clear subject and structure to build on, not on a single vague word
Writing Effective Prompts
Write in plain sentences, not labeled fields
Abyssale's AI models (Nano Banana, Gemini, GPT-image, and others) are multimodal models, the same kind of model you'd have a conversation with, adapted to also produce images.
They read a prompt the way a person would read a scene description, not the way an older diffusion model parses a bag of keywords. That means the best prompts read like a well-written paragraph describing the image, not a form with labeled fields.
Avoid writing your prompt as a template like this:
SUBJECT: [...] STYLE: [...] COMPOSITION: [...] LIGHTING: [...]
Two reasons to skip this format:
It can leak into the image. Some models render text with a high degree of accuracy, which also means literal labels like
SUBJECT:orTECHNICAL SPECIFICATIONS:occasionally get rendered as typography somewhere in the picture, since the model treats them as text to include rather than metadata to ignore.It encourages keyword stuffing. A bullet list under each label tends to turn into a pile of disconnected adjectives, which is harder for the model to weave into one coherent scene than a few well-formed sentences.
Think in categories, write in sentences
You don't need to abandon structure, you just need to fold it into natural language. Before writing your prompt, it still helps to think through the same categories used before:
Subject : what's the main focus, and what specifically distinguishes it?
Style : what artistic or photographic approach should this look like?
Composition : how is it framed, from what angle, with what arrangement?
Lighting & atmosphere : what light, time of day, and mood?
Environment : what's the setting and background?
Technical finish : what quality, focus, or rendering qualities matter?
The difference is in how you write it down: combine these into a couple of flowing sentences instead of a labeled outline.
Example
Instead of a labeled template, describe the scene the way you'd explain it to a photographer:
A luxury resort infinity pool set against a lush tropical landscape, shot in the style of contemporary architectural and resort-lifestyle photography. Wide horizontal composition with a linear perspective along the pool's edge and a calm, balanced reflection in the water, camera at a mid-level angle. Natural daylight with partial cloud cover casts soft shadows under the umbrellas, giving the scene a bright, serene, tropical ambiance. In the background, dense palm trees and a modern resort building with balconies complete the setting. Sharp architectural detail throughout, high dynamic range, professional real-estate photography quality.
This covers exactly the same ground as a labeled template would, subject, style, composition, lighting, environment, technical finish, but reads as something the model can actually picture as one coherent scene.
Key Elements to Include
Whatever the category, aim for specific, visual, concrete language over vague or abstract terms.
1. Main Subject
Describe the primary focus in detail, including size, color, and material where relevant.
Instead of: "A watch"
Use: "A luxury chronograph watch with a rose gold case, black ceramic bezel, and textured leather strap"
Instead of: "A modern building"
Use: "A 30-story glass skyscraper with a curved facade, steel accents, and a distinctive spire"
2. Style and Approach
Name the artistic style, the overall aesthetic, and any influences, as part of the same sentence describing the subject:
"...shot as high-end product photography with a clean white background and simple composition"
"...rendered as a photorealistic 3D visualization inspired by Scandinavian design principles"
"...illustrated as minimalist line art with a warm, casual, natural-light feel"
3. Composition
Describe framing, camera angle, and arrangement in the same natural flow:
"Vertical composition, subject centered with negative space at the top"
"Three-quarter perspective, items arranged in a grid with even spacing"
"Aerial view with a dynamic diagonal composition"
4. Lighting and Atmosphere
Describe the light source, time of day, and mood as part of the scene, not as a separate list:
"Two-point studio lighting with the key light at 45 degrees and a soft fill for gentle shadows"
"Early morning golden-hour light with a light mist, calm and warm"
"A single spotlight from above creating strong contrast and deep, dramatic shadows"
5. Technical Finish
Fold in quality and rendering expectations at the end of the sentence rather than as a bullet dump:
"...sharp focus throughout, high dynamic range, suitable for large-format printing"
"...shallow depth of field with the subject crisply in focus, rich and accurate color reproduction"
Note: Technical language like this is also where your model and quality/ratio choice do real work, a prompt asking for "high-resolution, sharp detail" will go further on a model and quality setting built for that than on a fast, low-quality pass.
Using Image References Alongside Your Prompt
Beyond text, you can attach up to 4 image references to guide a generation, useful for locking in a style, a color palette, a product shape, or a composition that's hard to put into words. When you use references:
Keep your prompt focused on what should change or combine, rather than re-describing everything already visible in the reference
If you want the image currently in your layer included as a reference, add it explicitly, it isn't included automatically
Note that not all models accept image references, so check the model's capabilities before relying on this
Common Use Cases
Product Photography
A white ceramic coffee mug with a glossy finish, shot as professional product photography from a 45-degree angle, centered, with a slight elevation. Soft, diffused studio lighting creates subtle shadows and clean reflections on the surface, against a light gray seamless studio background. Sharp focus throughout, high-key lighting, commercial quality.
Brand Illustrations
An abstract tech company logo concept, minimalist geometric design, balanced and centered layout built from simple shapes and clean lines. A deep blue gradient as the primary color with light gray accents, kept to a minimal palette. Vector-style quality, crisp edges, professional finish.
Tips for Better Results
Be specific and detailed : use precise descriptions, include measurements or proportions when relevant, and specify exact colors rather than general terms.
Write in sentences, not fragments : a prompt that reads naturally is easier for the model to turn into one coherent scene than a list of disconnected keywords.
Focus on visual elements : describe what can be seen, avoid abstract concepts ("innovation," "trust"), and use concrete visual terminology instead.
Include technical requirements : mention quality expectations, focus, and finish as part of the description, not as an afterthought.
Iterate instead of restarting : after a generation, refine your prompt from the selected result rather than writing a new one from scratch. Small, targeted changes ("make the lighting warmer," "remove the reflection") tend to work better than rewriting the whole description.
Common Mistakes to Avoid
Ineffective Prompt
"modern business tech startup professional corporate brand marketing design with blue colors and abstract shapes dynamic movement innovation leadership trust quality service solutions global reach"
Problems:
Keyword stuffing with no sentence structure for the model to interpret
No clear focus, too many competing concepts
Abstract, non-visual terminology ("innovation," "trust," "leadership") that has no concrete visual meaning
Better Version
A modern tech company's visual identity, shown as a clean, minimal composition with a professional business aesthetic. Dynamic geometric shapes in a balanced, centered layout, using professional blue tones as the primary color and cool gray as an accent, on a modern gradient background. The overall mood is professional, trustworthy, contemporary, and clean.
The second version keeps the same intent but expresses it as a coherent, visual scene rather than a string of buzzwords.
Limitations and Considerations
Generated images may vary from your exact description, treat the first result as a draft to iterate on, not a final answer
Complex prompts might need several attempts, or a switch to a different model, to achieve the desired result
Each generation consumes AI credits, calculated from your model, quality, dimension, number of image references, and number of variations, check the cost shown next to the generate button before launching
Not all models support the same ratio, quality, or image-reference options, check the model list if a setting you expect isn't available
If a generated image contains unwanted text or labels, check whether your prompt included formatting artifacts (like field labels or brackets) that the model may have rendered literally
Remember: Text-to-image generation works best with clear, detailed, natural-language descriptions. Take time to craft your prompts as you would explain the image to a person, and use the suggestion tags and Enhance button to help fill in the details you might miss.