Creating Characters with Fake Words: Unveiling a Midjourney Semantic Imagination Experiment

A Fascinating AI Creative Experiment

In Reddit's AI creative community, a creator shared an ingenious approach: first use a random word generator to produce a batch of completely nonexistent "fake words," then feed these meaningless terms into Midjourney and let the AI generate character designs based on these fictional vocabulary.

This seemingly simple experiment actually touches on a fundamental question about today's text-to-image models: When the input prompt doesn't exist in any real language system, how does the AI interpret and "fill in the blanks" to produce an image?

Why Fake Words Can Still Make Midjourney Generate Coherent Characters

The Associative Mechanism of Phonetics and Morphology

Even when a word is randomly assembled, it's still composed of real letters, syllables, and root fragments. For models like Midjourney that are trained on massive image-text pairs, these fragments aren't meaningless — the model has learned extensive implicit associations between "word forms and visuals" during training.

The core source of this associative capability is the CLIP (Contrastive Language-Image Pre-training) model. CLIP is a multimodal model released by OpenAI in 2021, trained through contrastive learning on approximately 400 million internet image-text pairs, learning to map text and images into the same high-dimensional vector space. In this shared space, semantically similar text descriptions and images are pulled closer together, while semantically unrelated ones are pushed apart. Because of this, even when fake words are input, CLIP's text encoder will still find a reasonable position in this vector space based on its subword composition, and the visual concepts associated with that nearby region will be activated for image generation.

For example, a fake word containing syllables like "-oth," "gr-," or "-ax" might be adjacent to certain fantasy, mechanical, or biological concepts in the training data, thereby guiding the model to output characters with corresponding aesthetics. This is why these fake-word characters often look remarkably coherent, as if they were always meant to look that way.

The Token-Level Decomposition Principle

From a technical perspective, the model doesn't reject a fake word as an unrecognizable whole. Instead, it uses a tokenizer to break it down into several subword tokens. The tokenization method universally adopted by modern large language models and multimodal models is the BPE (Byte Pair Encoding) algorithm. BPE was originally a data compression technique that was later introduced into natural language processing to solve the open vocabulary problem. Its core approach is: starting from individual characters, repeatedly merging the most frequently occurring adjacent character pairs in the training corpus, gradually building a fixed-size subword vocabulary. For example, "unhappiness" might be split into "un"+"happiness" or "un"+"happ"+"iness." This means that even when encountering completely unfamiliar fake words, the BPE algorithm won't throw an error or refuse to process — it simply breaks them down into known subword fragments.

Each subword carries a certain semantic vector, and ultimately these vectors are combined and weighted to form a semantic representation that the model "considers reasonable." This synthesized semantic representation is then projected into the model's Latent Space. Latent space is one of the core concepts in generative AI — in the diffusion model architecture used by Midjourney, images aren't generated directly in pixel space. Instead, the step-by-step denoising process is completed in a compressed, low-dimensional latent space, then restored to a full image through a decoder. This latent space can be understood as an abstract map of all possible images, where each coordinate point corresponds to a possible visual output, and adjacent points typically represent visually similar images.

In other words, AI never comes back "empty-handed" — even when facing purely fictional input, it mobilizes the closest concepts in the entire latent space to fill in the gaps. Once a fake word is encoded as a vector, the model locates the corresponding region on this conceptual "map" and samples from that region to generate images. This is an intuitive demonstration of the generative model's characteristic of "never saying I don't know."

Practical Value of the Fake Word Experiment for AI Creation

Breaking Free from Prompt Path Dependency

Many creators fall into fixed prompt patterns when using text-to-image tools like Midjourney: "a beautiful girl," "cyberpunk city," "highly detailed"... While these high-frequency word combinations produce stable results, they also tend to generate highly homogeneous outputs.

Understanding this phenomenon from a technical perspective, there's a key parameter in text-to-image models called CFG Scale (Classifier-Free Guidance Scale), which controls how closely the generated result adheres to the text prompt. When using common high-frequency prompts, a high CFG value makes the model generate images tightly centered around concepts that have been strongly reinforced by extensive training data, leading to highly convergent outputs. Since fake words aren't strongly bound to any specific visual pattern, they essentially give the model greater freedom to explore less-visited regions of the latent space, producing results with more diversity and surprise.

Using random fake words as seed inputs is essentially a strategy of actively introducing randomness and escaping the creative comfort zone. It forces the model into less frequently activated regions of the training data, thereby producing visual outputs with greater originality and unfamiliarity. For creators who need to design unique IP characters, game monsters, or concept art, this is a low-cost method for sparking inspiration.

Reverse Binding of Names and Appearances

Even more interesting is that this workflow provides a "name first, form second" character design logic. The traditional process typically involves conceiving a character before naming it, but the fake word experiment reverses this — an unfamiliar combination of syllables appears first, and then the AI gives it flesh and form. Characters generated through this reverse binding develop a natural consistency between their names and appearances, making it ideal for building naming systems in fictional worlds.

This method is particularly valuable in game design and fantasy literature creation. When J.R.R. Tolkien built Middle-earth, he first created language systems like Elvish, then built characters and worldviews around these languages. While the fake word experiment doesn't approach Tolkien's linguistic depth, its creative path "from sound to image" shares a similar aesthetic intuition — certain syllable combinations naturally evoke specific emotional associations, and AI happens to be able to visualize those associations.

The Boundaries of AI Understanding: Reflections from the Fake Word Experiment

This small experiment also leaves some thought-provoking questions. When we say AI "understood" a fake word, what exactly did it understand? The answer is clearly not the word's meaning itself, but rather the statistical associations implied by its form.

From a machine learning perspective, the fake word experiment essentially touches on the OOD (Out-of-Distribution) generalization problem. When a model receives input outside the distribution of its training data, ideally it should express uncertainty — for example, telling the user "I don't recognize this word." But generative models are designed with the opposite goal: they're trained to "always produce an output." This characteristic is viewed in academia as a mild form of "hallucination." The model uses nearest-neighbor interpolation to map unknown inputs onto some weighted combination of known concepts, generating a seemingly reasonable result based on that.

This reminds us that text-to-image models don't truly "understand" prompts — they perform interpolation and fitting within a vast network of image-text associations. They excel at creating smooth transitions between known concepts but can never create something completely outside the training distribution from scratch. So-called fake-word characters are essentially some rearrangement and recombination of training data — they always look "plausible" but never "entirely new," because the model can only creatively recombine visual concepts it has already seen.

For creators, understanding this is important: AI is a powerful association engine, not an independent imagination agent. True creativity still comes from how humans design inputs, curate outputs, and assign meaning. This game of creating characters with fake words perfectly demonstrates this kind of "human-AI collaborative imagination."

Conclusion

From a random word generator to Midjourney's character gallery, this experiment — though small — vividly reveals the operational logic and creative potential of generative AI. It's both an entertaining demonstration of technical characteristics and a practical technique for breaking creative inertia. Next time you're running low on inspiration, try feeding the AI a few nonexistent words — it might bring unexpected surprises.

Key Takeaways