One cell, step by step.

Follow one generated cell through 50 denoising steps. Compare its noisy state and expression estimate with the real-cell average for its type.

ctx2048 · E12 · EMA · 50 steps · 2,980 genes · One cell per type · Recorded run · no new inference

Loading the recorded run…

Genes

Noise in, cell out

What the model seesNoisy state x_t

What the model predictsExpression estimate x̂₀

Real-cell average

0 · log1p(CP10k)

One tile = one gene. Hover or tap a tile for its name. Blocks and gene order come from real cells of this donor, never from the generated cell.

Shared = genes detected in at least half of the real cells in each of the five broad cell types in this donor, including ribosomal, mitochondrial and early-response genes such as FOS and JUN. The prediction changes at every step because the model recomputes it from the updated noisy state. In these examples, the leading cell type stays the same after the early steps, while expression estimates keep changing.

Cell type

Which cell type is it becoming?

PredictionNoisy state1 = typical real cell

0 = the average of all five types. A type counts as settled once it stays on top above 0.5.

Map

Where the prediction sits

Real cellsGenerated endpoints · final typeNowStep 0

Same fixed UMAP as the main recording.

Similarity

How close to real cells?

PredictionNoisy state

Cosine to the real average of this type, over all 2,980 genes. For scale: a real cell against its nearest other real cell scores (median), with an observed maximum of . These observed values provide context; they are not a fixed upper bound for similarity to a cell-type average.

Selected genes

Each bar is the selected cell at this step. The gray bar with a black edge is the real average for its type, from the same donor.

All generated cells

The same measurement for every recorded cell, not only the one shown above.

Prediction settles: Step median · p10–p90
Noisy state settles: Step median · p10–p90
Gauge vs scANVI: agreement on real cells · on generated endpoints
Never settle: cells whose final type stays below 0.5

PredictionNoisy state

Methods & data

Recorded run: PBISC-Diffusion ctx2048 E12 (EMA weights), DDIM with 50 steps and no added noise, one pseudobulk condition from GSE135779, generated cells. This page runs no new model inference.
Shown cells: for each type, the first generated cell whose final scANVI label is that type. The rule was fixed before looking at any result. The Monocyte is the same cell as on the main page.
Gene blocks come from the real cells of the same donor. A marker is at least 0.5 log1p(CP10k) above the next type and detected in at least 20% of its type. IFN-α = the measured genes of MSigDB HALLMARK_INTERFERON_ALPHA_RESPONSE. Shared = detected in at least half the real cells of each of the five broad cell types in this donor.
Gauge: marker genes only. 0 = the average of the five type averages, 1 = a typical real cell of that type. It agrees with scANVI on of real cells and of generated endpoints.
The noisy state is the decoded diffusion state. Its early steps are not biological counts. All values are log1p(CP10k) on the 2,980-gene panel, stored to 1/255 of the colour range.
Page data (JSON) ↗UMAP points (JSON) ↗Main recording ↗