NeurIPS 2026

Few Channels Draw The Whole Picture

Source corgi wearing a pink rain jacket.
Source
Target desert landscape.
Target
Corgi transported into the desert scene.
Transported

Revealing Massive Activations in Diffusion Transformer

Evelyn Turri* Davide Bucciarelli* Sara Sarto Lorenzo Baraldi Marcella Cornia
University of Modena and Reggio Emilia · University of Pisa · *Equal contribution
Paper Code · coming soon

A sparse subspace that organizes, localizes, and transports semantic information inside modern DiT models: critical, spatially organized, and transferable semantic carrier channels.

Core idea

Three views of the same sparse semantic subspace.

Instead of treating massive activations as mere outliers, the paper studies what happens when we disrupt them, visualize them in space, and transplant them between generations.

01 · FUNCTION

Disrupt them, generation collapses.

Zeroing the top-ranked channels sharply degrades generation, while removing equally many low-statistic channels has only a limited effect.

02 · SPACE

They reveal where semantics live.

Restricting image tokens to massive channels produces structured activation maps and masks aligned with salient regions and foreground objects.

03 · CONTROL

Transport them, semantics move.

Injecting source massive activations into a target trajectory enables localized semantic transfer rather than unstructured pixel blending.

Interactive explainer

One canvas, three views of massive activations.

Switch between the sparse channel view, the spatial mask that emerges from those channels, and the masked activation transport pipeline. The three animations share the same corgi example and visual vocabulary.

Single DiT layer · image streamSchematic activations
generated image
Corgi wearing a pink rain jacket standing in a puddle.
16 image tokens
transformer block
DiT layer ℓ AdaLN · attention · MLP
tokens →
channels →
XℓI ∈ ℝNI × D
top-k massive channels · c06 · c18 · c27
start from the generation
Corgi wearing a pink rain jacket standing in a puddle.
image tokens inherit the latent spatial grid
keep only top-k channels
c06
c18
c27
each selected channel becomes a spatial activation map
cluster token features
Binary spatial mask isolating the corgi.
spatial mask p
source + target trajectories
SOURCE Corgi source generation. corgi in a pink jacket
TARGET Desert target generation. sunlit desert dunes
masking at layer ℓ
Masking
SOURCEXSI ⊙ M
⊙
=
Masking
TARGETXTI ⊙ (1 − M)
⊙
=
localized transport
Masked source corgi contribution. SOURCE ⊙ M
+
Masked target desert contribution. TARGET ⊙ (1 − M)
=
Corgi transported into the desert target scene. TRANSPORTED
X̃TI = XTI ⊙ (1 − M) + XSI ⊙ M
01
Tokens pass through one DiT layer.The output is a token-by-channel matrix.
Use cases

From prompt composition to personalization.

Source semantics are transported into a target generation while preserving the target scene and composition.

01

Text-to-text transport

A source concept is composed with a second prompt.

A small dog sleeping in a sunlit patch of grass.
The transported result: the small sleeping dog on the target beach.
The words Love Lives Here written with seashells on a sandy beach.
Source

A small dog dozing in a patch of sunlight.

Target

A sandy beach where ‘Love Lives Here’ is written with seashells.

02

Image-to-text transport

A subject from a reference image is personalized in a new scene.

Reference image of a spaniel wearing a pink hoodie.
The transported result: the reference spaniel at a busy city crosswalk.
A dog sitting at a busy city crosswalk.
Target

A photo of a dog sitting patiently at a busy city crosswalk.

Citation

BibTeX

@inproceedings{turri2026fewchannels,
  title     = {{Few Channels Draw The Whole Picture: Revealing Massive Activations in Diffusion Transformers}},
  author    = {Turri, Evelyn and Bucciarelli, Davide and Sarto, Sara and Baraldi, Lorenzo and Cornia, Marcella},
  booktitle = {{Advances in Neural Information Processing Systems}},
  year      = {2026}
}