An image to prompt generator writes a text description of an image that you can feed back into an image model. It does not recover the original prompt. For one image, ask a chat assistant for a prompt in your model's format, or use Midjourney's Describe if you work in Midjourney. For many images, ComfyUI's Qwen3.5 Text Generation template captions a folder on your own machine with a built-in node and an Apache 2.0 model. Edit every result before you use it.
An image to prompt generator looks at a picture and writes a description you can paste into an image model. It does not recover the prompt the creator used; it writes a new one that should lead somewhere similar. For one image, a chat assistant does the job. For a folder of them, ComfyUI does it free on your own machine.
Checked on 14 October 2026 against each tool's own docs or model card: Midjourney's Describe page, the CLIP Interrogator repository, the Florence-2 and Qwen3.5 model cards and Comfy's Qwen3.5 workflow. We did not score output quality, so there are no star ratings here. An earlier version ranked tools with stars and gave speed figures we could not source; they are gone.
Quick answer: which tool for which job
The image to prompt tools compared
Six tools cover nearly every use. They differ less in quality than in where they run and how much you can steer them.
| Tool | What it gives you | Cost | Best for |
|---|---|---|---|
| Midjourney Describe | Four suggested prompts in Midjourney's style | Needs a Midjourney plan | Midjourney users |
| A chat assistant (Claude, ChatGPT, Gemini) | Any format you ask for, with follow-up questions | Depends on your plan | One image at a time, for any model |
| CLIP Interrogator | A BLIP caption refined with CLIP terms | Free, MIT licence | Older Stable Diffusion prompt styles |
| Florence-2 | Short, detailed or more detailed captions, plus OCR | Free, MIT licence | Fast batch captions in your own scripts |
| ComfyUI Qwen3.5 template | Free-form text from an instruction you write | Free, Apache 2.0 model | Batches inside ComfyUI, no custom nodes |
| JoyCaption | Captions written for training diffusion models | Free open weights | LoRA datasets |
A few details from the sources. CLIP Interrogator combines OpenAI's CLIP and Salesforce's BLIP, its README recommends ViT-L-14/openai for Stable Diffusion 1.x, and its default settings use about 6.3 GB of VRAM, or 2.7 GB in low-VRAM mode. Florence-2 dates from a November 2023 paper, its large version has about 0.8 billion parameters, and it takes task prompts such as <DETAILED_CAPTION>. JoyCaption's model card describes it as built "for the community to use in training Diffusion models" and names Llama 3.1 underneath, so read the terms of both before commercial use.
Midjourney Describe
Describe is Midjourney's own reverse prompt. On midjourney.com you right-click an image, or drop one into the describe zone, and Midjourney suggests four prompts (announcement). You can send any of them to the Imagine bar or run all four.
- Four prompt suggestions in Midjourney's vocabulary
- Different on every run for the same image
- Quick to test, since you can run all four
- Built for Midjourney's current models
- The original prompt
- Saved: suggestions vanish when you refresh
- Free: it needs a Midjourney plan
- Tuned for Flux, SDXL or other models
For prompt structure once you have a suggestion, see our Midjourney prompts guide.
Using a chat assistant
For a single image going to any model, a general assistant that accepts image uploads is the most flexible tool, because you choose the format. Our free image to prompt generator does this in one step: upload a photo and it writes a prompt formatted for Midjourney, FLUX, Nano Banana or Stable Diffusion, with a negative prompt. In your own assistant, the instruction matters more than the assistant. This one keeps output short, ordered and free of names you should not use:
Describe this image as a single prompt for an image model.
Order: subject, action, setting, framing and camera angle,
lighting, colour and mood, medium or style.
Use plain nouns and adjectives. No names of real people,
brands or artists. Under 70 words.Then adjust for the target. Stable Diffusion and SDXL respond to comma-separated phrases; Flux and newer models read full sentences well; Midjourney has its own parameters. If a word in the result is doing nothing, cut it. The prompt engineering guide covers the general method, and the free AI influencer prompt generator builds a character prompt from scratch.
Captioning a folder in ComfyUI
ComfyUI now does this with no custom nodes. Comfy's Qwen3.5 Text Generation template runs on the built-in TextGenerate node, and the docs say it "can describe images and generate detailed captions" and "extract prompts and generation parameters from reference images". The Qwen3.5 model is listed under Apache 2.0 on its model card.
- 1Update ComfyUI
Missing nodes in the template usually mean an older version.
- 2Download one model size
2B for low VRAM, 4B for most consumer GPUs, 9B for higher quality. Files go in models/text_encoders.
- 3Open the template
Templates, then search for Qwen3.5: Text Generation.
- 4Load an image and write the instruction
Use the instruction above, or your caption style for training.
- 5Set the length and run
The default maximum length is 256 tokens. The output is a plain text string.
For a whole folder, put the graph in a loop through the API or a batch loader, and save each caption next to its image with the same file name. That pairing is the format most LoRA trainers expect, which is why captioning is the main reason people batch this at all. Our character LoRA dataset guide covers what a good training caption says and leaves out. New to ComfyUI? Start with what ComfyUI is.
Can you recover the real prompt?
Sometimes, and no tool is needed. ComfyUI's built-in save nodes store the whole workflow, prompts included, inside the PNG or video they write. Drag such a file onto the ComfyUI canvas and the original graph opens. Files that were re-encoded, screenshotted or posted to a social app have usually lost it, and then every tool on this page is guessing.
- 01Is it an original file?
A PNG or video straight from the creator, not a repost.
- 02Drag it onto ComfyUI
If a workflow opens, you have the real prompt and settings.
- 03Nothing opens?
The data is gone or was never there.
- 04Then generate a description
And treat it as a draft to edit.
The same applies in reverse: if you share originals and want to keep your prompts private, start ComfyUI with --disable-metadata. The workflow library guide explains what a shared file carries.
What it is good for, and where to stop
- Learning the vocabulary for a lighting or camera style you like
- Turning a mood board into starting prompts for your own images
- Captioning your own training images before a LoRA run
- Describing your existing brand photos to keep new images on style
- Building a searchable library of prompts from your own work
For consistent people specifically, prompts only go so far: see the AI influencer prompt guide and negative prompts for realistic people, then the AI Influencers program, which covers prompt engineering, LoRA training and ControlNet for one fictional persona. The US Copyright Office's AI pages explain how it treats AI-generated work.
Image to prompt: FAQ
What is an image to prompt generator?
A tool that looks at an image and writes text describing it, worded so you can feed it back into an image model. It is the reverse of normal generation. It does not recover the prompt someone actually used, which is usually not stored anywhere; it writes a new description that should lead to a similar picture. Treat the output as a first draft.
What is the best free image to prompt generator?
For one image, a chat assistant that accepts uploads, asked for a prompt in the format of the model you use. For many images on your own machine, the Qwen3.5 Text Generation template in ComfyUI, which runs on a built-in node and an Apache 2.0 model. For prompts tuned to Midjourney, its own Describe tool, which needs a Midjourney plan.
Can I get the exact prompt used to make an AI image?
Only if the creator left it in the file. ComfyUI saves its workflow, prompts included, inside the PNG or video it produces unless metadata is switched off, so dragging such a file onto the ComfyUI canvas shows the real prompt. Images that were re-encoded or posted to social apps have usually lost that data, and then any tool can only guess.
How does Midjourney Describe work?
On midjourney.com you right-click an image or drop it into the describe zone, and Midjourney suggests four prompts. Each run gives a different set, and the suggestions disappear when you refresh the page. You can send one prompt to the Imagine bar or run all four. Midjourney does not claim they are the original prompt.
Is it legal to use image to prompt on someone else's image?
Describing an image is not the risk; copying it is. Use a description to learn the lighting, framing and style vocabulary, then make something of your own. Do not recreate a specific copyrighted image, a brand's character or a real person's likeness, whatever tool wrote the prompt. This is general information, not legal advice.
Which image to prompt tool is best for LoRA captions?
One you can run on a whole folder with the same instruction, so every caption follows one style. The ComfyUI Qwen3.5 template, Florence-2 and JoyCaption are all used for this. Whatever you use, read and correct the captions before training, and keep the trigger word and the features you want the LoRA to learn consistent.