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

Andrew Fraser

cs.CVFeb 20, 2026

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

Authors: Andrew Fraser

Abstract

We demonstrate that morphological pressure creates navigable gradients at multiple levels of the text-to-image generative pipeline. In Study1, 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 Study2, we extend this to prompt-level morphology. Drawing on phonestheme theory, we generate 200 novel nonsense words from English sound-symbolic clusters (e.g., cr-, sn-, -oid, -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 -- snudgeoid, crashax, and 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.

Figures & tables

Figure 1: Morphological addressing via descriptor intersection. Each natural-language descriptor (e.g., “platinum blonde,” “beauty mark,” “1950s glamour”) defines a region in latent space. Their intersection addresses a specific identity basin without requiring the target’s name or reference photographs.

Figure 2: Grok name-based generation. Direct prompting with “Marilyn Monroe” produces outputs closely resembling archival photographs.

Figure 3: Morphological descriptors—“platinum blonde curled hair, beauty mark, 1950s glamour, white halter dress”—navigate to the same identity basin but generate synthetic outputs within that aesthetic space rather than reproducing training images.

Figure 4: The self-distillation training loop. Starting from morphological descriptors alone, the model iteratively generates images, curates outputs, trains a LoRA adapter on its own successful outputs, and refines prompts. No target name or reference photographs are used at any stage. Hit rate improved from 8% to 70% across four rounds.

Figure 5: Training progression across four rounds of self-distillation. Round 1 (top) shows high variance with only 8.1% of outputs approximating the target. By Round 4 (bottom), outputs exhibit binary behavior—landing clearly in the target basin or ejecting entirely.

Parameter	Value
Base model	SD 1.5
LoRA rank	32
Training epochs	10
Training software	Kohya_ss
Final training set size	52 images
Refinement rounds	4

Table 1: Training hyperparameters for the self-distillation LoRA.

Round	Hit Rate	Notes
1	9/111 (8.1%)	High variance, competing attractors
2	$\sim$ 60%	Blonde anchor added
3	$\sim$ 70%	Self-reinforcing convergence
4	53%	Binary behavior, basin ejection

Table 2: Hit rate progression across self-distillation rounds.

Figure 6: Base SD1.5 three-arm comparison. Arm A (Push) produces coherent shadow aesthetics. Arm B (Pull) scatters randomly. Arm C (Push+Pull) accesses sparse regions with structural anomalies.

Figure 7: LoRA-equipped three-arm comparison. All conditions show improved clustering. Arm C shifts from “eldritch” (base) to “uncanny valley”—coherent but precisely wrong.

Figure 8: Direct comparison of Arm C outputs. Left: Base SD1.5 produces “eldritch” structural breakdown. Right: LoRA-equipped model produces “uncanny valley” outputs—anatomically plausible but unsettling.

Condition	Base Avg Sim	LoRA Avg Sim	Base Max Sim	LoRA Max Sim
Arm A (Push)	0.245	0.465	0.460	0.644
Arm B (Pull)	0.071	0.181	0.149	0.254
Arm C (Push+Pull)	0.287	0.331	0.412	0.603

Table 3: ArcFace pairwise similarity across conditioning arms. The LoRA improved clustering in every condition, with the largest absolute gain in Arm A (+0.220 average similarity) and the most significant qualitative shift in Arm C (eldritch

\to

uncanny).

Figure 9: CFG sweep across values 5–11 with three seeds. Identity remains stable across guidance strengths, indicating robust basin structure rather than parameter-dependent behavior.

Figure 10: LoRA weight sweep showing phase transitions. Rather than smooth interpolation, outputs snap between identity attractors at specific weight thresholds.

Onset	Semantic Association	Examples
cr-	Impact, breaking, rough texture	crash, crush, crumble, crisp
gl-	Light, vision, smooth surfaces	glow, gleam, glitter, gloss
sn-	Nasal/oral, sneaky, quick motion	snout, sniff, snap, sneak
sl-	Sliding, slippery, low/negative	slide, slime, sloth, sludge
gr-	Grasping, grinding, discomfort	grip, grind, groan, grit
thr-	Violent motion, penetration	throw, thrust, thrash, throttle

Table 4: Phonestheme onsets and their semantic associations.

Suffix	Semantic Association	Examples
-oid	Resembling, robotic, biological	android, humanoid, asteroid
-ax	Tool, sharp, decisive	pickax, Ajax, thorax
-us	Latin biological/taxonomic	fungus, cactus, octopus
-um	Latin element/material	uranium, podium, museum
-or	Agent, doer	predator, terminator
-ix	Comic/magical, feminine-coded	Asterix, matrix, phoenix

Table 5: Phonestheme suffixes and their semantic associations.

Group	$n$	Mean Purity@1	Std Dev	Median	Pass Threshold
Phonestheme	200	0.3709	0.2997	0.3125	7 (3.5%)
Random pronounceable	100	0.2087	0.2837	0.0625	2 (2.0%)
Positive control	4	0.6719	0.3202	0.6875	1 (25.0%)
Negative control	50	0.1412	0.2556	0.0000	1 (2.0%)

Table 6: Purity@1 scores by candidate group.

Comparison	$t$ -statistic	$p$ -value	Cohen’s $d$
Phonestheme vs. Random	4.497	0.000010	0.551 (medium)
Phonestheme vs. Negative	4.983	0.000001	0.788 (medium-large)
Random vs. Negative	1.418	0.158	—

Table 7: Statistical comparisons between candidate groups.

Candidate	Type	Purity@1	Contaminated?
drudgea	phonestheme	1.0	YES — Drudge Report
crashor	phonestheme	1.0	YES — literal “crash” + agentive -or
broomix	phonestheme	1.0	NO
crashax	phonestheme	1.0	NO
snudgeoid	phonestheme	1.0	NO
skogum	phonestheme	1.0	YES — Swedish “skog” = forest

Table 8: All candidates achieving Purity@1 = 1.0.

Figure 11: The three confirmed cryptids, each achieving Purity@1 = 1.0 with zero training data contamination. Each row shows 4 of 16 outputs generated from the candidate string alone.

Onset	Mean Purity@1	$n$
cr-	0.473	21
sk-	0.409	20
gr-	0.410	18
thr-	0.404	24
dr-	0.384	20
sn-	0.352	16

Table 9: Mean Purity@1 by phonestheme onset.

Suffix	Mean Purity@1	$n$
-ling	0.452	13
-us	0.445	24
-oid	0.445	26
-a	0.398	25
-or	0.346	28
-ax	0.345	31

Table 10: Mean Purity@1 by phonestheme suffix.

T2I Generation

Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.

See all 299 papers

Jul 27, 2026

Diff-ID: Identity Consistent Facial Image Generation and Morphing via Diffusion Models

Generative diffusion models have revolutionized facial image synthesis, yet robust identity preservation in high resolution outputs remains a critical challenge. This issue is especially vital for security systems, biometric authentication, and privacy sensitive applications, where any drift in identity integrity can undermine trust and functionality. We introduce Diff-ID, a diffusion based framework that enforces identity consistency while delivering photorealistic quality. Central to our approach is a custom 210K image dataset synthesized from CelebA-HQ, FFHQ, and LAION-Face and captioned via a fine tuned BLIP model to bolster identity awareness during training. Diff-ID integrates ArcFace and CLIP embeddings through a dual cross attention adapter within a fine tuned Stable Diffusion UNet. To further reinforce identity fidelity, we propose a pseudo discriminator loss based on ArcFace cosine similarity with exponential timestep weighting. Experiments on held out and unseen faces show that Diff-ID does not exceed InstantID in raw ArcFace Face Similarity, but achieves substantially lower FID and the strongest FIQ based identity--realism trade off among the evaluated methods. We also present a unified DDIM based morphing pipeline that enables qualitative facial interpolation without per identity fine tuning. We further argue that identity preservation and photorealism should be evaluated jointly rather than in isolation, as high identity similarity alone does not guarantee realistic outputs. To make this trade off explicit, we report Face Image Quality (FIQ) as a complementary ratio based score that combines identity similarity and perceptual realism while keeping FS and FID as the primary metrics.

Taimoor Rizwan, Sara Atito, Muhammad Awais +2

Centre for Vision, Speech and Signal Processing (CVSSP) University of Surrey Guildford, United Kingdom · School of Artificial Intelligence and Computer Science Jiangnan University Wuxi, China

Image Generation Identity-Preserving Image Generation Diffusion-Based Image Generation

Ask PDF

Oct 13, 2026

Discovering Undisclosed Concepts in LoRA Adapters for Diffusion-Based AIGC Models

Low-Rank Adaptation (LoRA) has made diffusion model personalization affordable and widely shareable, but it also creates moderation risks when harmful or problematic content is not faithfully disclosed. Hence, it is desired to develop a mechanism that, given a LoRA adapter, discovers the concept it was trained to generate, so as to support downstream investigation. We characterize the target concept through the conditioning input that most clearly exposes what the LoRA has learned. Such an input should induce coherent LoRA outputs that are distinctive from those of the base model and strongly align with the learned LoRA subspace. We therefore formulate concept discovery as searching for the conditioning input that jointly maximizes output coherence, LoRA-induced distinctiveness, and subspace alignment. Even with a clear objective, however, optimizing it over an open-ended input space is prohibitively expensive, since reliable evaluation of coherence and distinctiveness requires repeated diffusion rendering. Crucially, we observe that an input need not be rendered to determine whether it is promising. Its alignment with the learned LoRA subspace provides a lightweight signal for localizing promising candidates before visual verification. Based on this insight, we develop a two-stage framework for text-to-image LoRAs. Subspace-Guided Evolutionary Search (SGES) uses render-free subspace alignment to explore candidate prompts and retain a compact shortlist, while Paired Visual Identification (PVI) renders only these candidates and compares the LoRA and base-model outputs to identify those exhibiting coherent and distinctive generation behavior. This design enables efficient and reliable discovery of previously unknown concepts encoded in LoRA adapters.

Yupeng Qiu, Han Fang, Dhananjai Sharma +1

National University of Singapore · University of Science and Technology of China

T2I Diffusion Models Diffusion Models Low-Rank Adaptation

Ask PDF

Oct 19, 2026

LEVI: Benchmarking Visual Identity Fidelity with Verifiable Low-Entropy Constructs

Modern generative models can produce visually plausible images while violating exact identity constraints. Evaluating such failures requires distinguishing violations of the requested identity from permissible variation in appearance. We introduce Low-Entropy Visual Identity (LEVI), which characterizes visual identities determined by compact facts and explicit rules. The low-entropy requirement concerns identity uncertainty, while appearance may remain diverse. We formulate identity fidelity in terms of the probability that a conditional image distribution assigns to the identities permitted by the prompt. Visual Identity Spaces (VIS) formalize this distinction by specifying admissible identity values and which different realizations express the same identity. Building on this formulation, we develop Verifiable Visual Grammar (VVG), a declarative language whose typed primitives and compositional constraints describe identity-bearing structures. Using this grammar, we build a benchmark for image generation. It assesses each output against the task's identity requirements, drawing evidence from the final pixels. The curated benchmark contains 1,375 tasks spanning 120 base identity spaces across six domains. Among ten evaluated model configurations, the highest whole-task pass rate is 58.69%. These results reveal a persistent gap between visually plausible images and faithful realization of their specified identities. Project page: https://maxiao0234.github.io/LEVI/.

Xiao Ma, Zhonghao Zhan, Yi Tang +1

ZODA · Imperial College London

Image Generation Evaluation Identity-Preserving Image Generation

Ask PDF

Abstract

T2I Generation

Explore similar work

Diff-ID: Identity Consistent Facial Image Generation and Morphing via Diffusion Models

Discovering Undisclosed Concepts in LoRA Adapters for Diffusion-Based AIGC Models

LEVI: Benchmarking Visual Identity Fidelity with Verifiable Low-Entropy Constructs