Recently, I started an experiment with style reference (sref) and personalization codes in Midjourney’s latest version, 6.1. To my disappointment, the images generated became particularly dull and uninspiring. They lost detail and diversity, producing extremely boring results. Here is a window in my attempts to recalibrate how Midjourney understands my prompts.
This project is an exploration into concept art and product design using Midjourney V6. I focused on blending fashion and consumer electronics references. I also use almost opposed movie references that create these interesting ambiguous in-betweens spaces.
In an attempt to regain the creative edge of my previous work, I reverted to version 5.2, removed the personalization codes and began the painful process of reteaching the system the style I wanted. It was (and still is in many ways) a long and difficult journey, but as my images started to improve, I cautiously resumed experimenting with version 6.
Style reference codes, or srefs, require a process that can only be described as “mining.” You need to send batches of recursive prompts, identify which codes show potential and then test them with specific prompts to see what they produce. This means investing time and resources into discovering hundreds of codes, most of which yield uninspiring or even horrible results.
Adding to the complexity, recursive commands can only be used in fast hour mode. This forces users who wish to explore this dimension of Midjourney to invest in fast hours, essentially paying for the privilege of experimentation. Discovering and selling sref codes has become a business where users bear all the costs. It’s like a casino where, by chance, some users stumble upon codes that might be interesting—but 99% are absolutely useless.
This trial-and-error process doesn’t just consume resources; it also affects how the AI understands user preferences. In my case, exploring srefs that didn’t match my usual style completely sent my Midjourney model in the wrong direction.
While rerolling itself doesn’t directly teach the AI new preferences, it plays a role in the overall learning process of Midjourney. When we reroll and then choose to upscale or create variations of specific results, we indirectly provide feedback to the system about preferred outputs. All the sub-generations and forks of sub-generations created during this process provide a path of influence on the desired outcome, shaping how the AI learns and adapts. Frequent rerolls therefore contribute to Midjourney’s dataset, helping the AI understand the range of interpretations for given prompts. Midjourney has introduced a personalization feature that learns from user preferences. While rerolling itself doesn’t directly affect this, the images users choose to keep or discard after rerolling contribute to the AI’s understanding of individual tastes.
This process can be conceptualized as “preference scaffolding,” where user interactions act as incremental supports that help structure and refine the AI’s understanding of what is desired. Much like in cognitive development, where scaffolding guides learners to higher levels of competence, the rerolling and branching of generations layer implicit signals that shape the AI’s evolving output. Each fork or reroll acts as a subtle cue, creating a cumulative framework that steers the AI toward aligning with user preferences over time.
This mechanism can also be understood through “latent feedback dynamics.” This term captures the hidden, underlying interactions through which indirect signals—such as rerolls, selections and forks—dynamically influence the AI’s adaptation process. These dynamics highlight the interplay between user behavior and the system’s evolving responses, emphasizing how preferences are communicated non-explicitly but meaningfully. Frequent rerolls, coupled with the decisions users make after these rerolls, create an ongoing loop of implicit feedback, allowing the AI to interpret and adjust to nuanced and evolving user expectations.
This content has been co-generated using an AI, using a series of notes and references for inspiration. The images are generated using Midjourney.
