Make a scene

What actually renders this

A Radeon AI PRO R9700 (32 GB, ROCm) in this homelab, running Qwen-Image-2.1 at 20 steps. The picture is generated on that card, the progress comes from the engine itself, and the file is served from this site. No third-party service is involved at any point: your browser talks to this domain and nothing else, and the prompt never leaves the network it was composed on.

There is no prompt box on purpose

Every button above is a fragment written and reviewed on this machine. When you press PLAY, your browser sends the identifiers of the buttons you picked — not text — and the sentence is assembled server-side. Anything that is not one of these identifiers is refused before it reaches the GPU. That is the whole safety model: there is nothing to type, so there is nothing to smuggle in.

What is inside the file

ComfyUI writes its whole recipe into every PNG it saves: the model and VAE filenames, the sampler settings, the negative prompt, the seed, and the full node graph. Useful on our side, none of your business on yours — so the relay removes every text chunk from the file before serving it and writes back only what we are willing to state:

• nfv:source — needforvram.com/scenes
• nfv:engine — the model, the step count and the sampler it ran on
• nfv:gpu — the card that rendered it
• nfv:asked — the sentence your buttons produced, the same one shown above
• nfv:size, nfv:recipe, nfv:rendered — dimensions, the identifiers you pressed, UTC time

No EXIF, no GPS, no camera or author field, no IP, no user agent, no engine graph. The untouched original stays on the lab machine so a render can be reproduced or audited. The pixels are never re-encoded: only the metadata chunks differ from the engine’s file.

Do not take our word for it — list the chunks of the file you just downloaded:

python3 -c "import sys,struct
d=open(sys.argv[1],'rb').read();i=8
while i<len(d):
    n=struct.unpack('>I',d[i:i+4])[0];t=d[i+4:i+8]
    if t in (b'tEXt',b'iTXt',b'zTXt'): print(d[i+8:i+8+n].split(b'\0')[0].decode())
    i+=12+n" picture.png

You should get the seven nfv: lines and nothing else (exiftool -a -G1 picture.png says the same thing). If you ever find more than that in a file served from here, that is a bug worth reporting.

The queue is fair, and small

One picture renders at a time. You get one job at a time, a 10-second pause between two jobs, 8 pictures per 10 minutes and 40 per day; the whole site shares a daily ceiling. When a limit bites, the page tells you the exact number of seconds to wait. If the card is asleep, the page says so instead of pretending. Pictures are public: they stay on the site and may be reused with a link back.

If a picture is wrong

There is no moderation queue behind a card in a garage, so: use the Report button under each result. Reports keep the picture and the recipe so it can be reviewed and pulled.