When an AI tool creates an image, it often saves the hidden "recipe"—the exact text prompt, negative prompt, model name, and settings—invisibly inside the file. Drop any PNG, WebP, or JPEG below to instantly read its embedded prompts, ComfyUI workflows, AUTOMATIC1111 / Forge settings, seeds, CFG scale, and LoRAs—explained in plain English.
Supports ComfyUI, AUTOMATIC1111, Forge, SwarmUI, Fooocus, InvokeAI, NovelAI & C2PA Content Credentials • Paste from clipboard (Ctrl+V)
Select Image FileThink of an AI-generated image like a baked cake that has its secret recipe card tucked invisibly inside the box. While your eyes only see the finished picture, most AI art programs automatically save the exact words and settings used to create the image inside the file's hidden header.
Many "image-to-prompt" websites just look at a picture and guess a description (like "a woman standing in a city"). The TruthLens AI Prompt Extractor does not guess: it opens the image's hidden binary file container and extracts the exact word-for-word text prompt, negative prompt, seed number, and model settings that the creator typed when generating the image.
And if the image was downloaded from WhatsApp, Instagram, or X (where hidden text gets stripped out), TruthLens automatically runs its 3-Model AI Pixel & Frequency Detector so you still get a clear answer on whether the image is AI-generated.
Whether you are learning how to make AI art, organizing your own design files, or fact-checking a suspicious photo, here is how people use this tool every day:
Ever see a breathtaking AI portrait or illustration on Reddit, Discord, or Civitai and wonder, "What words did they type to get that lighting and camera angle?" Dropping the original PNG here reveals the artist's exact prompt wording, lens terms, and style modifiers so you can learn from them.
Digital artists and design teams often generate hundreds of test images across weeks and forget which prompt or settings produced their favorite version. Simply drop your saved PNG back into the extractor to recover your exact prompt, seed number, and full ComfyUI flowchart.
For journalists, researchers, and moderators: if a suspicious viral image still contains an embedded Stable Diffusion parameters block,
ComfyUI workflow, or DALL-E C2PA signature inside its file header, you have undeniable proof it was created by AI.
Photography competitions, stock asset libraries, and creative agencies use prompt and metadata extraction to check whether submitted files were generated by AI tools or comply with AI disclosure guidelines.
When you extract data from an AI image, you will see technical terms like CFG Scale, LoRA, Seed, ComfyUI, and AUTOMATIC1111. You don't need to be a programmer to understand them—here is what each term means in plain English:
The Positive Prompt is the description of what the creator wants to see (e.g., "cinematic portrait, warm sunset light, 85mm lens"). The Negative Prompt tells the AI what mistakes to avoid drawing (e.g., "blurry, extra fingers, deformed hands, watermark, bad text").
Named after its creator's online username, AUTOMATIC1111 (often shortened to A1111) is the most popular free software dashboard for running Stable Diffusion on a home computer. Forge is a faster, updated version of the same app. Both automatically save the full prompt inside every PNG file they create.
Instead of a simple text box, ComfyUI is a program where creators build AI images by connecting visual boxes ("nodes") with wires—like a flowchart. ComfyUI saves that entire flowchart inside the PNG file, so anyone can drag the PNG back into ComfyUI to reload the exact setup!
Short for Classifier-Free Guidance. Think of CFG Scale as a strictness dial (usually set between 5 and 8). A low number lets the AI be more creative and loose; a high number forces the AI to follow every word of the prompt strictly.
Think of a LoRA as a small plug-in or mini expansion pack attached to the main AI model. Creators use LoRAs to teach the AI a specific art style (like 1990s anime, oil painting, or cinematic lighting) or a specific character's face.
Every AI image starts as a screen of random TV-static noise before the AI shapes it into a picture. The Seed (e.g., 849204173) is the exact ID number of that starting noise pattern. Using the same prompt + same seed recreates the exact same image.
The Checkpoint (like Flux.1, SDXL, or Juggernaut XL) is the main AI brain trained on millions of pictures.
The Model Hash is a short 10-character ID code (like 6ce0161689) that identifies the exact model version even if someone renamed the file.
AI turns blurry static into a sharp photo one pass at a time. Steps (typically 20–35) is how many passes the AI took to refine the image, and the Sampler (like Euler a or DPM++ 2M Karras) is the math formula used for each pass.
For technical users and forensic analysts: when local diffusion UIs and cloud generators export an image, they embed the generation recipe inside specific non-pixel metadata containers so the file can be re-imported into the workspace.
tEXt, iTXt & zTXt Ancillary Chunks
PNG files consist of a sequence of length-prefixed binary chunks after the 8-byte signature. ComfyUI writes two separate
tEXt chunks—prompt (containing the execution node graph JSON) and workflow (containing node positions and UI links).
AUTOMATIC1111 and Forge write a unified parameters chunk containing the positive prompt, Negative prompt: line, and comma-separated sampler key-values.
UserComment (Tag 0x9286)
Because JPEG and WebP lack PNG's arbitrary named text chunks, diffusion frontends serialize generation parameters into the EXIF
UserComment field (prefixed with an 8-byte character-code header such as UNICODE\0 or ASCII\0\0\0) or the
ImageDescription tag. TruthLens automatically strips the encoding header and parses the structured key-value block.
Comment JSON
Tools like NovelAI Diffusion and InvokeAI store JSON dictionaries inside the PNG Comment, Description,
and Software chunks—and in some configurations encode parameters into the least-significant bits (LSB) of the alpha channel.
TruthLens inspects all standard and custom PNG text keys alongside stealth watermark indicators.
Commercial generators (OpenAI DALL-E 3, Adobe Firefly, Leica/Sony provenance systems) rarely expose raw user prompts for privacy reasons,
but instead embed a cryptographically signed C2PA JUMBF box or XMP digitalSourceType declaration identifying the exact
model version and software agent that synthesized or inpainted the canvas.
What TruthLens extracts from original, unstripped exports across major open-source and commercial diffusion platforms:
| AI Generator / Frontend | Primary Storage Location | Recoverable Prompt & Workflow Fields | Model / Checkpoint Identification |
|---|---|---|---|
| ComfyUI | PNG tEXt keys: prompt & workflow |
CLIPTextEncode positive/negative nodes, KSampler seed, steps, CFG, sampler_name, scheduler, denoise | ckpt_name, UNETLoader, LoRALoader model names & strength weights |
| AUTOMATIC1111 / SD.Next | PNG parameters chunk or EXIF UserComment |
Full Positive Prompt, Negative Prompt, Steps, Sampler, Schedule type, CFG scale, Seed, Size, Hires upscale | Model name, Model hash (AutoV2 10-char SHA256), LoRA hashes, TI hashes |
| Stable Diffusion WebUI Forge | PNG parameters chunk |
Positive & Negative Prompts, Distilled CFG Scale (Flux.1), Guidance, Steps, Seed, Sampler | Checkpoint name, Diffusion in Low Bits / FP8 weight format, LoRA tags |
| SwarmUI & Fooocus | PNG sui_image_params / parameters JSON |
Prompt, Negative Prompt, Aspect Ratio, Styles array, Guidance Scale, Sharpness, Performance mode | Base model filename, Refiner model, Refiner switch percentage, LoRA array |
| InvokeAI & NovelAI | PNG invokeai_metadata / Comment JSON |
Positive/Negative Conditioning, Seed, Steps, Scale, Sampler, SMEA/DYN flags | Model identifier, VAE override, ControlNet adapters |
| DALL-E 3 / Adobe Firefly | C2PA JUMBF Manifest & XMP Packet | Signed Action history (c2pa.created, c2pa.inpaint), Issuer certificate chain |
claim_generator (OpenAI API / Firefly Image 3) & IPTC trainedAlgorithmicMedia |
Whether hidden prompts survive depends entirely on how the image was transmitted. Here is how major platforms handle PNG chunks:
Files saved directly to your local output/ directory, downloaded as original assets from Civitai or Hugging Face repositories,
or sent as uncompressed file attachments in Discord, Slack, or Telegram ("Send without compression") retain 100% of their PNG
tEXt and ComfyUI JSON chunks.
Social media CDNs re-encode uploaded PNGs into stripped WebP or progressive JPEG derivatives to save bandwidth, discarding all
tEXt and EXIF UserComment blocks. When metadata is stripped, TruthLens automatically falls back to its
3-model neural classifier and 2D FFT spectral detector so you can still verify if the image is AI-generated.
Answers to common beginner and technical questions about reading hidden prompts from AI images:
An AI Prompt Extractor is a tool that reads the hidden "creation recipe" stored inside an AI-generated image file. When programs like Stable Diffusion (AUTOMATIC1111, Forge), ComfyUI, Fooocus, or SwarmUI generate a picture, they save the exact positive prompt, negative prompt, seed number, and model settings inside invisible text fields in the PNG or JPEG header. TruthLens reads those hidden fields and displays them in a clean, copyable format.
TruthLens reads the real, verbatim prompt and settings saved inside the file's metadata—it never makes up a fake "guessed" prompt. If the hidden metadata was wiped (for example, after being sent through WhatsApp or Instagram), TruthLens clearly tells you that no plaintext prompt is stored in the file header and automatically runs its 3-model AI pixel & frequency detector so you can still check whether the image is AI-generated.
In simple terms: AUTOMATIC1111 and ComfyUI are popular free apps used to generate AI images on a computer (AUTOMATIC1111 uses a dashboard with sliders, while ComfyUI uses a visual node flowchart). A LoRA is a small add-on plug-in that teaches the AI a specific art style or character face, and CFG Scale is the "strictness slider" (usually set between 5 and 8) that controls how closely the AI obeys your text prompt.
ComfyUI embeds its entire execution graph as a JSON dictionary inside the PNG prompt chunk, where each node is keyed by ID.
TruthLens traverses the node graph, locates CLIPTextEncode, KSampler, KSamplerAdvanced,
CheckpointLoaderSimple, and UNETLoader nodes, and resolves which text node is wired into the positive versus
negative conditioning inputs—displaying a clean, human-readable prompt and parameter summary.
Social media apps (Instagram, WhatsApp, X, Facebook) re-compress uploaded images to save bandwidth, which strips out all hidden PNG text chunks and EXIF comments. Cloud services like Midjourney or DALL-E 3 often omit raw text prompts on web downloads (though DALL-E 3 and Adobe Firefly attach cryptographic C2PA Content Credentials, which TruthLens detects). When prompts are stripped, TruthLens automatically checks the image's pixels and frequency spectrum instead.
AUTOMATIC1111 and Forge compute a 10-character hexadecimal substring of the SHA-256 hash of the .safetensors checkpoint file
(e.g., Model hash: 6ce0161689) and store it inside the parameters chunk. You can search this 10-character hash
directly on Civitai or Hugging Face to identify the exact checkpoint or LoRA version used to render the image, even if the creator renamed the file locally.
Guidance-distilled models such as Flux.1 [dev] disable traditional negative-prompt classifier-free guidance (setting standard
CFG scale: 1) and instead rely on an internal guidance embedding typically set between 2.5 and 3.5
(recorded as Distilled CFG Scale in Forge or guidance in ComfyUI's FluxGuidance node). TruthLens extracts both parameters.
No. Images inspected on TruthLens are processed strictly in volatile memory to parse binary chunks and compute forensic signals, then immediately discarded. Your artwork, custom ComfyUI workflows, and prompts are never stored on disk or indexed into a public gallery.
Learn how AI generators store parameters inside binary headers, why cloud AI images have hollow EXIF, and how to check stripped photos.
EXIF & PNG Chunks
When Prompts Are Stripped
Prompt & Text Patterns
Switch to a dedicated single-purpose workbench for pixel-level image detection, video/voice deepfakes, or document style analysis.
