A better open image model in the no-credit tier

Venice's text side is solved. Images are the weak leg, and that's exactly where paying users are drifting away.

The no-credit image models (Chroma, Lustify, Z-Image Turbo, Anime WAI) are small and distilled. That's precisely why they're free — they're cheap to serve. It's also precisely why they fail: a distilled model places elements, it doesn't build scenes. They hold up with a single isolated subject and fall apart the moment there's a complex pose, real dirt, or camera intent.

This isn't a matter of taste. It's where the gap shows:

Test 1: "a framed painting on a wall of a woman, and a sign reading 'pasen a coger'". Chroma almost nails it, but invents a second sign full of gibberish. Z-Image Turbo renders the text perfectly and the anatomy falls apart.

Test 2: "a nude body face down in a forest, covered in dirt, completely inert, seen through the eyes of whoever left her there". Here the model doesn't have to place objects — it has to convey intent: that someone is standing over a body. None of the free models manage it. You get a nude in a forest, not a scene. And the nudity matters precisely because that's where anatomy and skin show, with no clothing to hide the mistakes.

The credit-billed models (Grok, GPT, Nano Banana) do deliver — but they're censored and they burn credits: ten images and your balance is gone. They can't be the daily workhorse.

And that's the point: far better open models already exist, the kind you see on Civitai and TensorHub — more realistic, with real skin. The filter is that they have to be open, commercially licensable, and permissive. Qwen Image or Krea are the right direction.

The ask: swap one of the small models in the no-credit tier for a current open model, and extend the "Free Models For Paid Tiers" experiment to a good image model as well.

And to close: it doesn't need to be a thousand images a day. A thousand images are worthless if they're mediocre. A hundred good ones beat a thousand average ones.