Extract ComfyUI workflow from PNG online: Free prompt & node viewer
View and recover ComfyUI workflows embedded in generated PNG images without installing ComfyUI or Python. Extract positive prompts, checkpoint models, sampler parameters, and download clean workflow JSON files locally.
Inspect ComfyUI workflows without launching Python
ComfyUI generates stunning AI artwork and embeds the entire node execution graph directly inside the generated PNG file. But when you find an image on Civitai, Discord, or Reddit, viewing that workflow usually requires firing up Python, starting ComfyUI, and downloading missing custom nodes.
Yellorn solves this with zero setup: drop any ComfyUI PNG directly into the editor. The decoder parses the PNG text chunks client-side in under 50 milliseconds, giving you instant access to positive prompts, model checkpoints, LoRA configs, sampler settings, and the full node graph.
What metadata gets extracted
- Clean summary (
__summary__): Checkpoint models, LoRA weights, sampler name, scheduler, step count, CFG scale, seed, dimensions, and extracted positive and negative prompts. - Execution prompt (
prompt): The raw API dependency graph. This is the exact payload ready to POST directly to ComfyUI's/promptREST endpoint for headless automation. - Canvas workflow (
workflow): The UI layout containing node coordinates, link wires, group boxes, and custom note annotations.
1-click transform tools
Once you load a ComfyUI PNG into Yellorn, open the Transform Menu in the editor toolbar to run automated extractions:
- Extract Canvas Workflow: Strips outer envelopes and isolates the pure UI workflow JSON. Save this file to drag and drop directly back onto an active ComfyUI canvas.
- Extract Execution Prompt (API): Extracts the headless execution graph formatted for headless Python scripts and batch rendering pipelines.
- Extract Positive Prompt: Isolates and cleans the raw conditioning prompt text, discarding negative tokens and node envelopes.
- Extract Negative Prompt: Copies only the negative conditioning prompt for reuse in other workflows or WebUIs.
100% private, client-side execution
Unlike online web tools that upload your files to remote servers, Yellorn executes 100% locally in your browser sandbox using efficient Uint8Array stream readers.
Your images, prompts, and custom model names never touch any server. It is completely safe for confidential commercial creative work, proprietary pipeline development, and sensitive artwork.
Frequently asked questions
How does ComfyUI embed workflow data into PNG images?
When ComfyUI renders an image, it serializes the entire node graph and prompt execution payload into JSON strings and writes them into the PNG file's metadata chunks (tEXt, zTXt, or iTXt). The visual pixels remain standard image data, while the hidden text chunks contain the full recipe.
Can I extract the workflow without installing ComfyUI or Python?
Yes. Drop any ComfyUI-generated PNG directly onto Yellorn. Yellorn parses the PNG chunks client-side in your browser using WebAssembly and Web Workers. It extracts the positive prompt, negative prompt, checkpoint models, LoRAs, sampler settings, and full UI workflow JSON in under 50 milliseconds.
Are my images or private prompts uploaded to your servers?
No. Yellorn operates 100% locally on your device. The image bytes never leave your browser window, making it completely safe for proprietary artwork, confidential commercial concepts, and sensitive NSFW workflows.
How do I import the extracted workflow back into ComfyUI?
In Yellorn, use the Transform Menu to click 'Extract Canvas Workflow'. Copy or download the clean JSON file, then simply drag and drop the JSON file directly onto the ComfyUI browser canvas to restore the exact node layout and wires.
What if my workflow JSON has syntax errors or won't load in ComfyUI?
Use Yellorn's 'Fix Broken JSON' tool. It automatically repairs missing commas, unescaped quotes, and corrupted node brackets, outputting valid schema-compliant JSON that ComfyUI loads without blank-canvas crashes.
Related
Where to go next
Try a fix in the editor or browse more articles.