When creating illustrations for a video, have you run into this too? The character comes out right but the pose doesn't; you finally get the composition right and then there's no room left for subtitles. Generating one image over and over, the most exhausting part is having to spot the problem and rewrite the prompt yourself every time.

To solve this, four developers from the 声临其境 team built 「千图百炼 · AI生图监工」(Thousand Images, Hundred Refinements · AI Image-Generation Supervisor), letting AI take over that rework loop for creators.

The Creation Process

The user's initial request: describe the image you want, the reference style and the conditions that must be met. The system fills in the rest of the requirements, generates a prompt and produces an image.

How the system works:

• It inspects the image along five dimensions: match with the goal, hard constraints, generation defects, visual quality and consistency with the reference.
• Obvious problems are fixed automatically and the image is regenerated; aesthetic suggestions are left to the user to decide on.

What you get: the prompt, score, cost and reasoning behind the changes for every round are all available to review. Once the quality target, round limit or budget cap is reached, the best image so far is delivered.