Teaching Our AI to Learn Faster from Every Approve and Reject
The Fixed-Step Problem
Our content system learns from user feedback. When a business owner approves a post, the visual style behind it gets a small score boost. When they reject one, the score drops. Over time, the system converges on what each brand actually wants to publish.
The original implementation used a fixed adjustment — the same size boost for every approval, the same size penalty for every rejection. It worked, but it was slow to respond when a user's preferences were clear. A business owner who rejected the same bold-typography style five times in a row was sending an unambiguous signal, yet the system adjusted at the same cautious pace as the first rejection.
On the other side, a single accidental rejection could suppress a style that the user had consistently approved for weeks. Equal adjustments treat every signal as equally meaningful, but they are not.
Patterns Are Stronger Signals Than Individual Actions
We replaced the fixed adjustment with an adaptive one that looks at recent history. Before adjusting a score, the system now checks the last ten interactions with that specific style. If a style has been consistently approved — more than seventy percent of the time — approvals carry extra weight. If a style has been consistently rejected — less than thirty percent approval — rejections hit harder.
The range of adjustment widened from a single fixed value to a spectrum of three to fifteen points. A first-time rejection of a popular style nudges it gently. A fifth consecutive rejection of a struggling style pushes it down firmly. The system now responds proportionally to the strength of the signal.
This matters because taste revelation is not uniform. Some preferences become clear after two or three interactions. Others take longer. An adaptive system reaches the right answer faster in the common case without overreacting in edge cases.
Styles That Fail Should Stop Appearing
Beyond adaptive scoring, we added a dropout threshold. When a style's score falls below twenty, it is filtered out of the guidance that shapes content generation. The style is not deleted — it can recover if the user later approves content in that aesthetic — but it stops influencing new posts.
Previously, a style with a score of five would still appear in the guidance alongside styles scored at eighty-five. The system would occasionally generate content in the rejected style simply because it was still in the list. Now, consistently rejected styles go quiet until they earn their way back.
The guidance strings sent to the image generation pipeline also became more informative. Each style now includes its current score as a percentage, and an explicit avoidance section lists any styles that have fallen below the threshold. The AI generating images does not need to guess which styles are preferred — the signal is direct and quantified.
From Guidance to Conditioning
The mood board used to influence content through text alone — a description of preferred visual styles injected into image generation prompts. Text guidance is useful, but it is an indirect control. The AI interprets the description and translates it into visual choices, with all the approximation that involves.
We added a more direct channel. Reference images from the mood board are now passed as conditioning inputs to the image generation model. This is the same technology we use to maintain consistent influencer faces and accurate product appearances, but applied to overall visual style at a deliberately low influence level.
The effect is subtle by design. The reference images nudge the aesthetic — color temperature, lighting quality, composition tendencies — without overpowering the specific scene being generated. A mood board full of warm, natural-light photography will tilt generated images toward that warmth without forcing every image into the same template.
Threading Through the Entire Pipeline
The mood board previously influenced a single stage: the image generation prompt. We discovered that decisions made earlier in the pipeline — content planning and scene design — also benefit from knowing the brand's visual preferences.
The content calendar now considers preferred content types from the mood board when deciding what kind of posts to generate. A brand whose mood board favors lifestyle photography will see more lifestyle-oriented themes in their calendar. A brand that resonates with educational content will get more informational angles.
Scene planning, where the AI designs specific compositions before generating images, now receives mood board direction as well. If the mood board favors minimalist settings, the scene planner will design cleaner, less cluttered compositions. If it favors vibrant, energetic environments, scenes will include more color and activity.
By threading mood board intelligence through planning, scene design, and image generation, the system produces content that feels cohesive rather than randomly styled. The same visual sensibility that shaped the user's approvals now shapes the content from its earliest conception.
The Compounding Effect
These changes compound. Adaptive scoring means the system reaches accurate preferences faster. Style dropout means it stops wasting generations on rejected aesthetics. Visual conditioning means preferences are expressed through direct visual reference, not just text interpretation. Pipeline threading means preferences shape content from planning through execution.
A business owner who actively reviews their first batch of content will receive a noticeably more aligned second batch. By the third or fourth content cycle, the system has converged on their visual identity with a precision that would be difficult to achieve through explicit configuration alone.
The goal is not just faster learning but more complete learning — preferences that permeate every decision in the content pipeline, not just the final image generation step. When a business owner says the content "just gets it," that understanding was built from every approve and reject they ever tapped.