Most "prompt help" tools focus on text. Paste a question, get a better question. Useful for ChatGPT and Claude. Useless for the part of your AI workflow that is actually expensive: image generation.
Generating an image with Midjourney, DALL-E 3, Flux, Imagen, Nano Banana, or SDXL costs 10 to 100 times what a text completion costs in tokens or credits. Bad image prompts fail in ways that are harder to debug than bad text prompts:
A 4/10 text prompt produces vague but recognizable text.
A 4/10 image prompt produces a recognizable shape that is missing the lighting, composition, lens choice, or subject specificity you wanted in the first place.
You burn the credit. You look at the output. You sigh. You start over. Often without understanding why the prompt failed.
This is what image-prompt QA solves.
Text vs image prompts: different failure modes
Text prompt failures are usually about constraints:
- No audience defined, the model guesses
- No length set, output is too long or too short
- No tone, output is generic
- No success criteria, you cannot tell a good output from a bad one
Image prompt failures are about physical specification:
- Subject ambiguity. "A woman with a dog" vs "a 30-year-old Black woman in a yellow rain jacket walking a Shiba Inu."
- No composition cue. "A coffee cup" vs "a coffee cup, centered, shallow depth of field, golden hour, 50mm lens, eye level."
- No style anchor. "Fantasy castle" vs "fantasy castle, watercolor, Studio Ghibli style, soft pastel palette."
- No medium. Photorealistic? Illustration? Oil painting? 3D render? Matte painting?
- No lighting. Golden hour? Overcast? Studio softbox? Rim light? Cinematic? Neon?
- Conflicting modifiers. "Minimalist baroque" or "vintage futuristic" confuse the model.
- Negative prompt missing. What should not appear (no text, no extra fingers, no background clutter).
- Aspect ratio not stated. Defaults rarely match what you actually need.
A good image-prompt rubric scores all of the above before you hit generate.
The image prompt QA rubric
This is the rubric we use inside FixMyPrompt when you paste an image prompt.
Subject specificity (0 to 25 points)
Who or what is in frame? Demographics, clothing, expression, pose, action. The vaguer the subject, the more the model improvises. Improvisation rarely matches your mental image.
Composition (0 to 15 points)
Camera angle, framing (close-up, medium, wide), subject placement (rule of thirds, centered), depth, perspective. Without these, you get whatever the model defaults to. Usually centered, eye level, medium distance.
Lighting (0 to 15 points)
Time of day, source direction, hardness, color temperature. "Golden hour, side-lit, warm tones" produces something predictable. "A photo at a place" does not.
Style / medium (0 to 15 points)
Photo? Illustration? 3D render? Oil painting? Reference an artist or movement when possible. "Wes Anderson color palette" gives Flux something to anchor to. "Pretty colors" does not.
Technical specs (0 to 10 points)
Lens (24mm, 50mm, 85mm), camera (DSLR, film, smartphone), film stock (Portra 400, Cinestill 800T), resolution, render engine (Octane, Unreal, V-Ray) for 3D-style outputs.
Negative space / exclusions (0 to 10 points)
What should not be in frame. "No text, no extra hands, no background people." Particularly important for Midjourney and SDXL.
Aspect ratio + format hints (0 to 10 points)
--ar 16:9 for landscape, --ar 2:3 for portrait, --ar 1:1 for square. Stating this explicitly avoids cropping disasters.
A prompt that scores 80+ on this rubric will produce something close to your mental image on the first try. A prompt under 50 will need three to five regenerations to dial in. At $0.04 to $0.20 per image, or one to ten credits per generation in subscription tools, that adds up fast.
Worked example
Here is a real prompt a user pasted into FixMyPrompt last week:
hero image for a coffee brand landing page
Score: 18 / 100.
Issues flagged:
- Critical. No subject specificity. "Coffee" could mean beans, a cup, a barista, a plantation, a bag, or a packaging shot.
- Critical. No medium. Photo, illustration, 3D, or watercolor?
- Major. No composition. Hero images need specific aspect ratios and focal points.
- Major. No lighting. Coffee photography lives or dies on lighting.
- Major. No style anchor. "Coffee brand" could mean Starbucks, Blue Bottle, Stumptown, or Death Wish. Wildly different aesthetics.
- Minor. No negative prompt. Hero images often need "no text overlay" because the brand name gets added later by a designer.
Improved prompt generated by the QA:
Hero shot of a single ceramic pour-over coffee cup on a reclaimed wood counter, centered, shallow depth of field. Steam rising. Warm morning light streaming from camera-left through a window with soft shadows. Hand-poured aesthetic, indie specialty coffee shop vibe (Blue Bottle / Stumptown reference). Shot on 50mm, f/1.8, natural light, no flash. Muted earth-tone palette: cream, walnut, brass. Negative: no text, no logos, no people in frame, no clutter on counter. Aspect 16:9, landscape orientation, photorealistic.
Score: 91 / 100.
That is the difference between a prompt that needs six regenerations and one that lands on attempt two.
Why no other tool does this
Most "prompt optimizer" tools work for text only. They have one rubric (clarity, specificity, structure) and one rewrite engine. Image prompts need a different rubric (composition, lighting, medium, lens) that text-prompt rubrics do not capture.
Browser-extension optimizers cannot do image-prompt QA because the input UX (Midjourney's /imagine slash command, DALL-E's chat box, Flux's text field) does not fit a "click sparkle button to rewrite" pattern.
A web-based QA tool is the only practical place to grade image prompts before you spend the credit.
How to use FixMyPrompt for image prompts
- Go to fixmyprompt.net/try. Free, no signup, three free reports per day.
- For deeper analysis, sign up free and switch the prompt mode to "Image prompt."
- Paste your draft image prompt.
- Get a scored report with composition, lighting, style, and technical breakdowns.
- Copy the improved prompt into Midjourney, DALL-E, Flux, Imagen, Nano Banana, or SDXL.
The rubric is model-agnostic. It works with any image-gen tool.