Most bad results are not a model failure. They are an underspecified request. Learning how to write ai image generation prompts is mostly learning to say the four or five things the model cannot guess, then reaching for the controls — negative prompts, reference uploads, LoRA, inpainting — that some tools expose and others do not. This guide covers both halves: the sentence you type, and the switches worth hunting for when the sentence is not enough.
The five slots a usable prompt fills
A prompt is not a wish. It is a specification with predictable gaps. Fill these five and you eliminate most re-rolls:
- Subject — who or what, with the two or three attributes that matter. "A woman" leaves everything open. "A woman in her sixties, short silver hair, wearing a canvas apron" closes it.
- Composition — framing and distance. Close-up, three-quarter portrait, wide establishing shot, overhead flat lay. This single word changes more pixels than any adjective.
- Lighting — soft window light, hard noon sun, single rim light against a dark background. Lighting is what separates "AI picture" from "photograph."
- Style or medium — 35mm film photo, gouache illustration, isometric 3D render, pixel art sprite. Pick one; stacking four styles produces mush.
- Setting and mood — where it happens and how it should feel, in a few words rather than a paragraph.
One weak prompt, rewritten
Here is the same idea at three levels of specification. These are illustrations written for this guide, not output from any particular tool.
| Version | Prompt | What you get |
|---|---|---|
| Vague | a coffee shop, cozy | A generic interior. Every element is the model's default. |
| Better | interior of a small coffee shop, warm afternoon light, cozy | Right mood, but framing and subject are still uncontrolled. |
| Specified | wide interior shot of a narrow coffee shop, barista at an espresso machine in the mid-ground, late afternoon sun through a front window, warm shadows, 35mm film photograph, muted colors | Subject, composition, lighting, and medium are all pinned. Re-runs vary in detail, not in concept. |
Two habits do the heavy lifting. Change one variable at a time — if you rewrite the subject and the lighting together, you cannot tell which fix worked. And keep a text file of prompts that landed; the strongest prompt library is your own. Some tools bake this in: ImgPilot ships a prompt library and generation history alongside multi-turn refinement, and Craftura AI includes a prompt enhancer that expands a short line before generating.
What a negative prompt actually does
A negative prompt is a second box where you list what should not appear — extra fingers, text, watermarks, a cluttered background. It is a separate input, not a "no" in the main prompt, and the distinction matters: writing "no text" inside a positive prompt can just as easily summon text.
Negative-prompt fields are far less common than prompt-engineering advice implies. Across the image generation category, only a handful of tools document one:
- Craftura AI — freemium, multiple models including Flux, with custom and negative prompts plus a credit system.
- Imgi.in — paid, generates with DALL-E 3 and FLUX Schnell and lists negative prompt support alongside art-style and mood presets.
- Avatarify AI — freemium photo-to-avatar generator offering custom prompt and negative prompt, running on daily free credits.
- AI App Icon Generator — paid, one-time credit packs, with prompt and negative-prompt controls and image-to-image redesign.
- ImagineQr — freemium prompt-to-QR-art generator with optional negative prompts and a watermark-free free tier.
If your tool has no negative box, the workaround is positive phrasing of the thing you want instead: "plain concrete wall" beats hoping the model honors "not cluttered."
Upload instead of describing
Describing a specific face, room, or product in words is the hardest possible way to get it. Around 140 tools in this category accept an uploaded photo or reference image, which converts a paragraph of description into a file.
Ideogram.ai builds a consistent character from a single reference photo, holding the same face across poses, outfits, and scenes, and is documented for rendering legible text inside images. Whisk AI takes the idea further with three separate uploads — subject, scene, and style — blended by Google's Gemini and Imagen 3 models, so you steer with pictures rather than adjectives. SciFig accepts text, sketches, reference images, PDFs, or lab photos as input for publication-style scientific figures, and PixelLab keeps game sprites on-style through reference-based editing.
Practical rule: if a detail is non-negotiable — a face, a logo shape, a room layout — supply it as an image. Save prose for the things that are genuinely open.
LoRA, when a reference image is still not enough
A LoRA is a small add-on trained on top of a base model to teach it one specific subject or style — your face, a mascot, a house illustration style — so you can invoke it in every later prompt instead of re-describing it. It costs you a training step and a set of example images, and buys consistency across dozens of generations.
Roughly 22 tools here reference LoRA, in three distinct roles:
- Train your own: LoRA AI trains custom LoRAs from 10–20 images to keep characters and styles consistent, then generates with models such as Flux, Seedream, and Nano Banana. Krea offers LoRA fine-tuning on your own data inside a broader image, video, and 3D suite. Dreamlook is aimed at Stable Diffusion finetuning with LoRA file extraction and a Dreambooth API.
- Apply someone else's: Artroom AI pulls unlimited LoRAs via CivitAI integration and adds ControlNet pose, outline, and depth control. PixAI.Art and Holara both center on anime generation with large community model-and-LoRA libraries, Holara adding a tag-based prompt helper. CGDream exposes LoRA styles and 300+ filters on Flux models.
- Find and run them: Flux Lora is a free browsable catalog of FLUX.1 LoRA models sorted by base model and license, and Stability Matrix is a free one-click installer for running Stable Diffusion locally with models, LoRA, VAE, and CLIP.
Fix one region with inpainting instead of re-rolling
Inpainting means masking a region and regenerating only that area. When nine-tenths of an image is right and one hand, sign, or shadow is wrong, re-running the whole prompt gambles away the parts you liked. About 26 tools in this category document inpainting.
| Tool | Pricing | Documented editing scope |
|---|---|---|
| Dezgo | Freemium | Inpainting plus upscaling and background removal across an aggregated model list; a number of models are free |
| ImageFX.dev | Freemium | Inpainting, outpainting, and background removal, with premium models on paid tiers |
| FLORA AI | Freemium | Inpaint, outpaint, and crop on a collaborative canvas with credit rollover |
| APImage | Freemium | Inpainting and non-destructive editing, reusable characters and product assets, REST API |
| Letz.AI | Freemium | Inpainting, animation, and upscaling in one Studio, plus custom model training |
| Anifusion | Freemium | Infinite canvas with inpaint and pose control for manga and comic pages |
Two more worth knowing if your work is narrow: SciFig does region-based inpainting on scientific figures that stay editable afterward, and PixelLab offers style-consistent inpainting for pixel-art sprites.
Why the same words behave differently everywhere
Paste an identical prompt into three products and you get three different pictures, because you are usually not talking to the same model. Many products are interfaces over shared model families — Flux AI is built around FLUX models, CGDream runs on Flux Dev and Pro, Imgi.in calls DALL-E 3 and FLUX Schnell, and Dezgo aggregates Flux, Stable Diffusion, and others in one place. Our explainer on what Flux is and how it differs from other image models unpacks that layer, and the Model Frontends (SD/Flux) subcategory collects the tools whose main job is being a good interface to those models.
The practical consequence: a prompt is portable, but prompt tuning is not. Retune length, style keywords, and negative prompts when you switch products, and expect the first few generations on a new tool to be calibration rather than output.
How to write AI image generation prompts on a tool you have never used
Putting it together, this is the sequence worth following on unfamiliar software, and it is the order that wastes the fewest credits. Start with a short, fully specified prompt covering the five slots. Generate a small batch at low settings to see the model's defaults. Change one variable, re-run, compare. Add a reference image the moment a detail must be exact. Reach for a negative prompt only for recurring artifacts. Once a subject repeats across many images, train a LoRA. And repair with inpainting rather than re-rolling a near-miss.
Which product you run that loop in matters as much as the loop itself — see how image generators compare on output and price and the full working list of image generators by pricing model, or browse every option in the Image Generation category.