Flux AI Image Generator: How Picnow Renders Images in Under 0.5 Seconds

Which model Picnow uses

Text-to-image requests go to Juggernaut Lightning Flux, a model from RunDiffusion that we run through the Together AI inference platform. It belongs to the Flux family of models that Black Forest Labs released in 2024. Juggernaut is RunDiffusion's line of fine-tuned models, best known for photorealistic output, and the Lightning version is tuned to work with very few sampling steps.

We used FLUX.1 [schnell] before. When it stopped being offered as a serverless model on our provider, we switched to Lightning Flux, which fills the same role: cheap, fast and good enough to iterate on.

When you upload a photo, Picnow uses a different model: FLUX.1 Kontext [pro]. Kontext takes an existing image plus an instruction ('make it a watercolor', 'add snow') and edits the image instead of starting from scratch. It is slower than the Lightning model, so generate-while-typing is switched off while a photo is loaded.

Why four steps is enough: distillation

Diffusion models create an image by starting from random noise and removing a bit of that noise in each step. A standard model typically needs somewhere between 20 and 50 of these steps for a clean result, and every step is a full pass through a large neural network. More steps means more waiting.

Distillation shortens that process. A smaller number of steps is trained to land where the original model would end up after many steps, using the original model as the teacher. Names like Lightning, Turbo and schnell (German for fast) all point to models trained this way.

Picnow calls the model with 4 steps. That is the main reason an image takes a fraction of a second instead of several seconds. The cost is a little less fine detail in busy scenes and less reliable text, which we come back to below.

Keeping image sizes in check

The work a diffusion model does grows with the number of pixels, so image size has a direct effect on speed. Picnow offers presets for common formats: desktop, phone, Instagram post, Facebook cover and Twitter header. Each preset keeps the right aspect ratio but is rendered at half of the target dimensions.

The provider also requires width and height to be multiples of 64, between 128 and 2048 pixels, so every size is rounded to the nearest multiple of 64. The desktop preset, for example, has a target of 1792 by 1248 pixels. Halved that is 896 by 624, which is rounded to 896 by 640.

These sizes are big enough for screens, chats and social posts, and small enough to keep generation fast and cheap. If you need large prints, run the image through an upscaler afterwards.

How generate-while-typing works

Fast generation only feels instant if you do not have to press a button. Picnow watches the prompt field and waits until you pause typing for 300 milliseconds. Then it sends the prompt and shows the result. If you keep typing, the timer resets, so you do not fire a request on every keystroke.

A few rules keep this from wasting images. The prompt needs to be longer than two characters. If the prompt is the same as the last one that was generated, nothing is sent. And auto-generate pauses when you ask for more than one image at a time, or when you are editing an uploaded photo. In those cases you press generate yourself.

Each automatic generation counts as one image. In practice a short typing session produces a handful of images, and because each one only costs a fraction of a cent, exploring this way stays cheap.

What one image costs

Without an account you can generate 10 images per day for free. After that you buy a credit pack. There is no subscription. The Starter Pack is €5 for 500 images (€0.01 each), the Creator Pack is €12 for 1,500 images (€0.008 each) and the Power Pack is €30 for 5,000 images (€0.006 each). Credits never expire.

Credits are taken off your balance only after an image has been generated successfully. If the provider returns an error, you are not charged for it.

That pricing is only possible because each generation is so short. A 4-step model at a modest resolution uses far less compute than a 30-step model at full HD, and that difference ends up in the price per image.

Where the speed trade-off shows

A few-step model is not the best choice for every job. Crowded scenes with many people can lose detail. Longer text inside images is less reliable than one to four words. Hands and small objects sometimes come out wrong.

The practical answer is to iterate: keep the prompt short, add one detail at a time and regenerate when something is off. Our guide on how to write AI image prompts covers this in more detail. For edits to an existing image, upload it and describe the change, so the Kontext model can work from your photo.

If you want a broader comparison with tools like Midjourney, ChatGPT and local Stable Diffusion setups, read our overview of the fastest AI image generators.

FAQ

Is Picnow based on Flux? Yes. Text-to-image uses Juggernaut Lightning Flux, and photo edits use FLUX.1 Kontext [pro]. Both come from the Flux model family.

Can I change the number of steps? No. Picnow always uses 4 steps for text-to-image. That setting is what keeps generation under half a second.

Does generate-while-typing use my credits? Yes, each automatic generation counts as one image. You can turn auto-generate off and press generate yourself.

Is it free to try? Yes. You get 10 free images per day without signing up, and no credit card is needed.