Diffusion Mental Averages

Phonphrm Thawatdamrongkit; Sukit Seripanitkarn; Supasorn Suwajanakorn

cs.CVMar 31, 2026

Diffusion Mental Averages

Authors: Phonphrm Thawatdamrongkit, Sukit Seripanitkarn, Supasorn Suwajanakorn

Organizations: VISTEC, Thailand

Abstract

Can a diffusion model produce its own "mental average" of a concept-one that is as sharp and realistic as a typical sample? We introduce Diffusion Mental Averages (DMA), a model-centric answer to this question. While prior methods aim to average image collections, they produce blurry results when applied to diffusion samples from the same prompt. These data-centric techniques operate outside the model, ignoring the generative process. In contrast, DMA averages within the diffusion model's semantic space, as discovered by recent studies. Since this space evolves across timesteps and lacks a direct decoder, we cast averaging as trajectory alignment: optimize multiple noise latents so their denoising trajectories progressively converge toward shared coarse-to-fine semantics, yielding a single sharp prototype. We extend our approach to multimodal concepts (e.g., dogs with many breeds) by clustering samples in semantically-rich spaces such as CLIP and applying Textual Inversion or LoRA to bridge CLIP clusters into diffusion space. This is, to our knowledge, the first approach that delivers consistent, realistic averages, even for abstract concepts, serving as a concrete visual summary and a lens into model biases and concept representation.

Figures & tables

Figure 2: Overview. Multiple noise latents are jointly optimized so that their denoising trajectories converge toward shared semantics. At each timestep $t$ , their $h$ -space activations are averaged to form a semantic target, and each latent is optimized to match it before denoising to the next step. Repeating this process across timesteps aligns coarse-to-fine semantics, yielding a single “mental average” of the concept.

Panel of Figure 3 — Figure 3: Qualitative comparison across methods: GANgealing [ 70 ] , Avg. VAE , D 4 M [ 89 ] , MGD 3 [ 13 ] , and DMA (Ours) . Rows show concepts: astronaut , dog , bicycle , freedom . GANgealing cannot process abstract concepts, so its freedom cell is intentionally left blank.

	Consistency (↓)			Representativeness (↓)			ImageReward (↑)
Method	CLIP	DreamSim	LPIPS	CLIP	DreamSim	LPIPS	ImageReward (↑)
GANgealing † [ 70 ]	0	0	0	0.386	0.477	0.851	$-0.684$
Avg VAE	0	0	0	0.473	0.806	0.805	$-2.262$
D 4 M [ 89 ]	0.168	0.274	0.572	0.197	0.363	0.672	0.823
MGD 3 [ 13 ]	0.180	0.319	0.643	0.195	0.364	0.687	0.755
DMA (Ours)	0.031	0.032	0.129	0.179	0.341	0.655	1.002

Table 1: Average scores across four categories ( Animal , Person , Object , Abstract ) on Consistency (↓), Representativeness (↓), and ImageReward (↑). GANgealing † is evaluated only on three non-abstract categories, as it lacks pretrained GANs on abstract concepts

Panel of Figure 4 — Figure 4: Quality Trade-off. DMA achieves a superior balance of representativeness, consistency, and perceptual quality.

Panel of Figure 5 — Figure 5: DMA prototypes of unsupervised modes. Top row: overall average. Bottom rows: averages of discovered modes.

Panel of Figure 6 — Figure 6: DMA prototypes of grounded modes. We cluster astronaut by ethnicity, dog by breed, and car by model or color.

Panel of Figure 7 — Figure 7: LoRA & Textual Inversion bridge the gap between the clustering and the $h$ -spaces, enabling cluster-specific prototypes.

Panel of Figure 8 — Figure 8: DMA prototypes of various SD Variants. Soldier inherits model-specific gender bias, while Italy exhibits a persistent static bias toward a Venice-like scene across all variants.

Panel of Figure 9 — Figure 9: DMA generalizes to the DiT architecture [ 69 ] .

Panel of Figure 11 — Figure 11: Comparison with the pre-computed activation baseline. The pre-computed baseline (top) degrades over timesteps, while DMA (bottom) remains stable and consistent.

Panel of Figure 12 — Figure 12: Comparison with the single-timestep optimization baseline. This baseline produces inconsistent results and introduces noticeable artifacts if optimization is done in later timesteps. In contrast, our method maintains both semantic consistency and visual quality.

Panel of Figure 13 — Figure 13: Results of the replacement baseline with different $t_{\text{stop}}$ . This baseline yields inconsistent images, even with high $t_{\text{stop}}$ values.

Panel of Figure 14 — Figure 14: Effects of different $t_{\text{stop}}$ values on visual consistency. $t_{\text{stop}}=10$ yields consistent images with reduced computational cost.

Panel of Figure 15 — Figure 15: Quality trade-off across baseline hyperparameters. Across the full range of hyperparameter settings for each baseline, our method consistently achieves superior representativeness, consistency, and visual quality.

Panel of Figure 16 — Figure 16: Effects of CFG scale on visual quality. Low CFG scales fail to generate faithful structures (as seen in monstera ), while excessively high CFG scales lead to oversaturated colors and unnatural contrast.

Figure 17: Effect of the number of noisy latents $K$ . In the blue box, we sample four distinct sets of $K=100$ noisy latents and compute their averages. For each set, we show eight example averages (in a 4x2 grid), each computed from a random latent in that set, which converge to similar-looking images within each set . However, at $K=100$ , noticeable variations are still observed across the four sets shown in the images with colored borders. As $K$ increases, averages across distinct sets become more similar, consistent with the statistical behavior of larger sample sizes.

Panel of Figure 18 — Figure 18: Ablation on DiT transformer block selection for DMA. See details in Section C.7 .

Panel of Figure 19 — Figure 19: Hyperparameter tuning of D 4 M for fair comparison. We sweep the noise-injection timestep $t\in\{2,4,6,8,10\}$ across four concepts: bird , astronaut , car , and freedom . Larger values of $t$ preserve the averaged latent more strongly but lead to blurrier outputs, while smaller values make the result less consistent.

Panel of Figure 20 — Figure 20: Hyperparameter tuning of MGD 3 for fair comparison. We sweep the mode-guidance weight $\lambda\in\{0.01,0.02,0.05,0.1,0.2\}$ across four concepts: bird , astronaut , car , and freedom . Larger $\lambda$ increases attraction toward the estimated mode, improving structure but sometimes introducing over-sharpening or hallucinations.

Method	Representativeness (↓)	Consistency (↓)	ImageReward (↑)
MGD 3	0.165 / 0.678 / 0.364	0.147 / 0.610 / 0.317	-0.8151
DMA (Ours)	0.151 / 0.626 / 0.318	0.050 / 0.192 / 0.093	-0.7085

Table 2: Quantitative comparison on DiT. (CLIP / LPIPS / DreamSim)

Panel of Figure 21 — Figure 21: Average images of person , house , and cloth from different ethnicities.

Panel of Figure 22 — Figure 22: Average images of a person with different attributes.

Panel of Figure 23 — Figure 23: Average images of additional concepts with two different attributes.

Panel of Figure 24 — Figure 24: Average images of various concepts.

Panel of Figure 25 — Figure 25: Cluster averages for the Bicycle concept grounded by color . Top: Diffusion samples grouped into clusters according to the grounding attribute color . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 26 — Figure 26: Cluster averages for the Cake concept grounded by type . Top: Diffusion samples grouped into clusters according to the grounding attribute type . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 27 — Figure 27: Cluster averages for the Jellyfish concept grounded by color . Top: Diffusion samples grouped into clusters according to the grounding attribute color . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 28 — Figure 28: Cluster averages for the Toy concept grounded by shape . Top: Diffusion samples grouped into clusters according to the grounding attribute shape . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 29 — Figure 29: Cluster averages for the Nurse concept grounded by ethnicity . Top: Diffusion samples grouped into clusters according to the grounding attribute ethnicity . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 30 — Figure 30: Cluster averages for the Place concept grounded by environment . Top: Diffusion samples grouped into clusters according to the grounding attribute environment . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 31 — Figure 31: Cluster averages for the Object concept grounded by shape . Top: Diffusion samples grouped into clusters according to the grounding attribute shape . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 32 — Figure 32: Cluster averages for the Building concept grounded by architecture . Top: Diffusion samples grouped into clusters according to the grounding attribute architecture . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 33 — Figure 33: Cluster averages for the Person with musical instrument concept grounded by instrument . Top: Diffusion samples grouped into clusters according to the grounding attribute instrument . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

Panel of Figure 34 — Figure 34: Cluster averages for the Person with pet concept grounded by pet . Top: Diffusion samples grouped into clusters according to the grounding attribute pet . Bottom: Cluster averages computed by DMA with LoRA. These averages highlight how our method captures the dominant visual characteristics within each cluster.

	Animal			Person			Object			Abstract
Method	CLIP↓	DreamSim↓	LPIPS↓	CLIP↓	DreamSim↓	LPIPS↓	CLIP↓	DreamSim↓	LPIPS↓	CLIP↓	DreamSim↓	LPIPS↓
GANgealing [ 70 ]	0	0	0	0	0	0	0	0	0	–	–	–
Avg VAE	0	0	0	0	0	0	0	0	0	0	0	0
D 4 M [ 89 ]	0.123	0.191	0.615	0.127	0.174	0.510	0.193	0.308	0.627	0.229	0.425	0.534
MGD 3 [ 13 ]	0.122	0.224	0.679	0.131	0.202	0.598	0.215	0.342	0.686	0.251	0.508	0.610
DMA (Ours)	0.015	0.021	0.138	0.022	0.019	0.118	0.040	0.039	0.158	0.047	0.049	0.101

Table 3: Consistency (↓) across concept categories. Lower values indicate higher consistency of prototypes within each seed. Each category includes representative concepts: Animal — Bird, Cat, Dog; Person — Astronaut, Doctor, Firefighter; Object — Bicycle, Car, TV Monitor; Abstract — Anger, Freedom, Poverty.

	Animal			Person			Object			Abstract
Method	CLIP↓	DreamSim↓	LPIPS↓	CLIP↓	DreamSim↓	LPIPS↓	CLIP↓	DreamSim↓	LPIPS↓	CLIP↓	DreamSim↓	LPIPS↓
GANgealing [ 70 ]	0.310	0.394	0.839	0.523	0.599	0.866	0.323	0.439	0.848	–	–	–
Avg VAE	0.505	0.867	0.831	0.504	0.780	0.796	0.464	0.832	0.882	0.420	0.746	0.711
D 4 M [ 89 ]	0.158	0.320	0.708	0.144	0.230	0.613	0.224	0.379	0.714	0.260	0.522	0.652
MGD 3 [ 13 ]	0.155	0.310	0.718	0.140	0.230	0.636	0.223	0.375	0.724	0.262	0.543	0.671
DMA (Ours)	0.143	0.307	0.669	0.142	0.208	0.595	0.202	0.368	0.718	0.229	0.479	0.638

Table 4: Representativeness (↓) across concept categories. Lower values indicate closer alignment between the prototype and the overall concept distribution.

Panel of Figure 35 — Figure 35: Baseline comparison for the astronaut concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 36 — Figure 36: Baseline comparison for the firefighter concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 37 — Figure 37: Baseline comparison for the doctor concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 38 — Figure 38: Baseline comparison for the bird concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 39 — Figure 39: Baseline comparison for the cat concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 40 — Figure 40: Baseline comparison for the dog concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 41 — Figure 41: Baseline comparison for the bicycle concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 42 — Figure 42: Baseline comparison for the car concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Panel of Figure 43 — Figure 43: Baseline comparison for the tvmonitor concept. Top: Generated samples from the diffusion model. Bottom: Average images produced by GANgealing [ 70 ] , D 4 M [ 89 ] , MGD 3 [ 13 ] , and our method.

Diffusion Models

Diffusion models learn to reverse gradual data corruption, enabling flexible generation, editing, forecasting, and inference.

See all 1,169 papers

Sep 28, 2026

You Can't Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning

Concept unlearning in text-to-image diffusion models aims to suppress a target concept (e.g., horse) while preserving related but distinct content (e.g., donkey), yet existing methods either leak under indirect prompts or visibly degrade other concepts. We show that these failure modes stem from the geometry of concept representations rather than from any particular algorithm. Formalizing concepts as activation-space regions, we prove that the overlap between a target and other concepts lower-bounds the damage any robust erasure must inflict on them, with the trade-off scaling linearly in the degree of overlap. Across thirteen unlearning methods, including methods designed to preserve non-target concepts, no method achieves both strong erasure and strong neighbor preservation: STEREO nearly eliminates indirect leakage but cuts neighbor generation by more than 75%, while sparse inference-time methods preserve neighbors but leak. Damage increases with our overlap measure, monotonically so for STEREO; the

κ

-scaling reproduces on SDXL, and neighbor-selective damage recurs on FLUX. Perfect unlearning is the wrong target for entangled concepts; methods should be evaluated on the Pareto frontier our theorem establishes.

Yian Wang, Ali Ebrahimpour-Boroojeny, Hari Sundaram +1

University of Illinois Urbana-Champaign Urbana, IL, USA

Diffusion Model Unlearning Machine Unlearning T2I Diffusion Models Diffusion Models

Ask PDF

Sep 28, 2026

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are then scored and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.

Vighnesh Subramaniam, Boris Katz, Brian Cheung +3

MIT CSAIL · Google Cloud AI Research · UCSF

Diffusion Model Guidance T2I Generation Diffusion Sampling Video Generation Diffusion Models Diffusion-Based Denoising

Ask PDF

Sep 9, 2026

Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion Models

Although text-to-image diffusion models generally exhibit strong prompt-following ability, we identify a persistent and previously underexplored failure pattern in which a small subset of prompts differing only in the object consistently fails to realize the same target concept under identical generation settings. We term this phenomenon object-dependent concept brittleness. Such cases suggest systematic internal blind spots rather than random sampling noise. In this paper, we present an interpretability-oriented framework to audit and minimally correct these failures. Our key idea is to analyze denoising trajectories in a step-wise sparse autoencoder (SAE) space, where abstract style and attribute concepts become more separable than in the raw denoising representation. This sparse space enables us to compare successful and failed generations, identify concept dimensions whose evidence is missing, weakened, or temporally delayed, and construct class-level concept prototypes from reliable class-consistent samples. Based on this audit process, we introduce a lightweight inference-time correction strategy that interpolates denoising features toward the corresponding prototype in SAE space. Rather than serving as a task-specific retraining method, this intervention acts as a validation of the diagnosed concept deficiency. We evaluate the proposed framework on style and attribute failure cases across multiple diffusion backbones, with significant improvements in concept consistency, text fidelity, and repair success. Further analyses show that deeper denoising representations provide clearer concept structure, while early-stage intervention offers the strongest correction leverage. Code is available at https://github.com/Metecade/Object-Dependent-Concept-Brittleness.

Yifan Yuan, Xiangyu Liu, Hongming Shan +5

School of Artificial Intelligence Shenzhen University Shenzhen, China · School of Mathematical Sciences Shenzhen University Shenzhen, China · Institute of Science and Technology for Brain-inspired Intelligence Fudan University Shanghai, China +3

Neural Network Interpretability T2I Diffusion Models

Ask PDF

Abstract

Diffusion Models

Explore similar work

You Can't Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

Interpreting Object-Dependent Concept Brittleness in Text-to-Image Diffusion Models