Morphological Addressing of Identity Basins in Text-to-Image Diffusion Models

Abstract: We demonstrate that morphological pressure creates navigable gradients at multiple levels of the text-to-image generative pipeline. In Study~1, identity basins in Stable Diffusion 1.5 can be navigated using morphological descriptors -- constituent features like platinum blonde,'' beauty mark,'' and 1950s glamour'' -- without the target's name or photographs. A self-distillation loop (generating synthetic images from descriptor prompts, then training a LoRA on those outputs) achieves consistent convergence toward a specific identity as measured by ArcFace similarity. The trained LoRA creates a local coordinate system shaping not only the target identity but also its inverse: maximal away-conditioning produces eldritch'' structural breakdown in base SD1.5, while the LoRA-equipped model produces ``uncanny valley'' outputs -- coherent but precisely wrong. In Study~2, we extend this to prompt-level morphology. Drawing on phonestheme theory, we generate 200 novel nonsense words from English sound-symbolic clusters (e.g., \emph{cr-}, \emph{sn-}, \emph{-oid}, \emph{-ax}) and find that phonestheme-bearing candidates produce significantly more visually coherent outputs than random controls (mean Purity@1 = 0.371 vs.\ 0.209, p<0.00001p < 0.00001 p<0.00001, Cohen's d=0.55d = 0.55 d=0.55). Three candidates -- \emph{snudgeoid}, \emph{crashax}, and \emph{broomix} -- achieve perfect visual consistency (Purity@1 = 1.0) with zero training data contamination, each generating a distinct, coherent visual identity from phonesthetic structure alone. Together, these studies establish that morphological structure -- whether in feature descriptors or prompt-level phonological form -- creates systematic navigational gradients through diffusion model latent spaces. We document phase transitions in identity basins, CFG-invariant identity stability, and novel visual concepts emerging from sub-lexical sound patterns.

• The paper demonstrates that morphological descriptors and phonesthemes enable precise targeting of identity basins in diffusion models, with hit rates rising to 70% after self-distillation.
• It employs LoRA adapters and push-pull conditioning to iteratively optimize latent space navigation and enhance model control over generated visuals.
• Key findings show that sound symbolism in nonsensical prompts reliably produces coherent visual clusters without explicit exemplars, offering new avenues for model interpretability.

Morphological Addressing of Identity Basins in Text-to-Image Diffusion Models

Introduction and Overview

This work presents an in-depth analysis of morphological structure as a navigational signal within text-to-image diffusion models, with a focus on Stable Diffusion 1.5. The central thesis is that both visual feature descriptors (e.g., "platinum blonde," "beauty mark") and phonological patterns (phonesthemes such as "cr-," "-oid") systematically define regions in the model’s latent space. Through a sequence of controlled studies, the authors demonstrate that these descriptors enable precise “addressing” of identity basins---dense attractor regions---even in the absence of explicit names or training images. Moreover, the authors establish that sound-symbolic structure in nonsense prompts produces emergent visual coherence surpassing random controls, pointing to intrinsic morphological gradients within the generative pipeline.

Morphological Addressing and Identity Basins

The latent space in diffusion models is highly structured: certain entities, like frequently-photographed celebrities, form distinctive basins, each defined not by a label but by intersections of constituent visual features. Morphological descriptors act as high-dimensional coordinates that, when intersected, uniquely localize an attractor associated with the corresponding identity. This framework is formalized through “morphological addressing,” with the intersection of visual descriptors targeting specific attractor regions instead of relying on explicit names or training examples.

Figure 1: Morphological addressing via descriptor intersection; overlapping natural-language regions define expressive local basins.

Prompting models directly with target names (e.g., “Marilyn Monroe”) yields high-fidelity reproduction of memorized imagery:

Figure 2: Name-based prompting produces outputs closely resembling archival photographs of the memorized target.

Crucially, morphological descriptors lacking the identity token yet including features such as hairstyle, facial attributes, and period-specific aesthetics, drive the model toward the same semantically dense region---delivering plausible outputs within the correct category, but not direct data regurgitation.

Figure 3: Morphological descriptors alone, intersected, induce consistent yet synthetic outputs constrained by the defined feature set.

Self-Distillation: Iterative Morphological Optimization

The effectiveness of morphological addressing is amplified by a self-distillation protocol leveraging LoRA adapters. Starting with an initial synthetic dataset derived purely from morphological descriptors, the model iteratively curates high-fidelity outputs, fine-tunes a LoRA on these selections, and refines descriptors, yielding marked improvements in addressability and clustering strength.

Figure 4: Self-distillation loop---morphological descriptors are used to iteratively curate outputs, distill via LoRA, and refine addressability; accuracy increases substantially with each round.

The hit rate (proportion of outputs landing in the correct identity basin) surges from 8% in the naive initial stage to 70% post-distillation. Each round of self-distillation further polarizes the output space, sharpens basin boundaries, and increases binary decision-making---samples either converge completely or are ejected fully.

Figure 5: Across self-distillation rounds, the model transitions from high variance to distinct binary behavior in addressing the identity basin.

Push-Pull Conditioning and Inverse Navigation

The authors introduce a conditioning protocol to test both attraction and repulsion relative to identity basins. Three regimes are analyzed:

• Push: Morphological “shadow” descriptors pushing toward the identity’s antonymic features.
• Pull: Negative conditioning with target descriptors.
• Push+Pull: Combined, maximizing directional force away from the trained attractor.

On base SD1.5, push conditioning alone produces aesthetically coherent but non-anomalous “shadow” portraits, while pull-only disperses outputs randomly. The combined regime, however, overshoots traditional attractors and reliably induces “eldritch” outputs---structurally degraded faces in sparse latent regions.

Figure 6: Three-arm comparison---push yields coherent shadows, pull alone is scattered, but push+pull enters sparse "eldritch" territory.

Employing a LoRA adapter trained on the target basin fundamentally alters these behaviors: clustering tightens, and the “eldritch” region becomes an “uncanny valley”---anatomically plausible yet clearly deviating from the Platonic ideal.

Figure 7: With LoRA, clustering improves in all conditions; push+pull lands outputs in the uncanny valley, not total breakdown.

A direct visual juxtaposition highlights the qualitative distinction: the LoRA stabilizes even the anomalous regions, enforcing anatomical but subtly erroneous outputs.

Figure 8: Comparison of base SD1.5 and LoRA under maximal inverse conditioning; LoRA suppresses catastrophic collapse in favor of structured "uncanny" errors.

Stability and Topology of Latent Space Navigation

Extensive sweeping of classifier-free guidance (CFG) values reveals that identity basin addressability is invariant under varying degrees of guidance. Outputs remain consistently clustered around the target attractor, confirming the existence of highly stable, CFG-robust attractor geometry.

Figure 9: Identity stability across CFG scales indicates topological robustness of the addressed basin.

Varying the LoRA mixing weight reveals phase transitions: rather than continuous morphing, outputs “snap” between attractors at discrete thresholds. This indicates that the latent space is partitioned by sharp decision boundaries and that LoRA-induced semantic directions correspond to region-specific control rather than mere soft modulation.

Figure 10: As LoRA weight increases, identity exhibits abrupt transitions instead of smooth interpolation, highlighting hardness of basin boundaries.

Phonestheme-Driven Prompt Construction (“Crungus Hunt”)

Moving from visual feature descriptors to linguistic morphological structure, the authors systematically generate 200 nonsense prompt candidates built from English phonesthemes (onsets/suffixes with established semantic associations). They demonstrate that such structured “nonsense” prompts generate significantly more coherent and self-clustered outputs (mean Purity@1 = 0.371) than random controls (0.209), with p < 0.00001 and Cohen’s d = 0.55.

The methodology identifies three “cryptid” prompts---snudgeoid, crashax, and broomix---each yielding Purity@1 = 1.0 without evidence of training data contamination. These nonsense words, by virtue of their compositional structure, define new attractor basins in the model’s output space.

Figure 11: The three confirmed cryptids (snudgeoid, crashax, broomix); each prompt creates a distinct and tight visual cluster from phonesthetic patterns alone.

Theoretical and Practical Implications

The results have profound implications for both interpretability and control in generative models:

• Latent space structure: Identity basins are not diffuse or arbitrary; they are discrete, bounded, and can be topologically mapped via intersecting morphological descriptors or phonological constraints.
• Personalization without exemplars: Fine-tuning for creative or privacy-sensitive targets is possible without explicit examples, circumventing common limitations in DreamBooth/Textual Inversion-like approaches.
• Sound symbolism as navigational signal: Sub-lexical composition in prompts is decoded and mapped to semantically meaningful regions, evidencing that foundational tokenization processes (e.g., BPE in CLIP) play a critical, underexplored role.
• Inverse shaping: LoRA fine-tuning induces bidirectional structure---not only does it attract outputs toward specific basins, but it also modifies behavior in sparse regions, elucidating new perspectives on failure modes and model generalization.

These findings offer new directions for robust model auditing, preventative overfitting detection, data leakage assessment, and creative prompt design.

Limitations and Future Directions

Several limitations are notable: all experiments are conducted on SD1.5/CLIP, raising questions of generalizability; the phonestheme inventory is English-centric; non-exhaustive contamination checks may miss obscure data overlaps; and identity basin granularity across less-memorized entities is not systematically probed. Extensions should investigate cross-linguistic phonesthemes, alternative encoder architectures, and broader concept domains.

Conclusion

Morphological pressure---at the training level through feature descriptors and at the prompt level via phonesthemes---systematically defines navigable gradients in text-to-image diffusion model latent spaces. These gradients manifest as robust, discrete attractor basins, which may be mapped, addressed, and manipulated without reliance on explicit naming or exemplar conditioning. The results provide both a new theoretical understanding of latent space topology and practical schemes for controlled generation via structured prompt engineering and fine-tuning (2602.18533).