acceptodds
Under review as a conference paper at ICLR 2027

OmniMorph: A Channel-Adaptive Unified Generative Model for Cell Painting

Abstract

Cell Painting uses multiplexed fluorescent dyes to label multiple subcellular structures of cells and is a standard assay for large-scale chemical and genetic screening. Due to the vast combinatorial space of chemical and genetic perturbations, using generative models to replace part of these experiments has become a promising avenue. However, the cell lines and perturbations covered by any single dataset are limited, so a model trained on that dataset alone struggles to generalize beyond them. Each dataset is also set up differently depending on its research purpose, such as whether it includes brightfield images or same-batch controls. Therefore, a unified generative model for Cell Painting that supports multiple datasets and multiple tasks is needed. Unlike natural images with a fixed RGB channel space, Cell Painting datasets vary in both the number of channels and the subcellular structures they label, which makes such a unified model difficult to train across datasets. To address this, we propose OmniMorph, to our knowledge the first unified generative model for Cell Painting. OmniMorph is a flow-matching model with a DiT backbone that encodes each fluorescence channel separately. We disentangle the semantics of a channel from its role with two embeddings. The channel embedding marks the subcellular structure the channel represents and is shared across datasets, and the role embedding marks the part it plays in the current task. The model is further conditioned on perturbation embeddings from pretrained encoders and on a natural-language instruction that specifies which task to perform. We train OmniMorph jointly on six datasets and on the four tasks of morphology generation, perturbation prediction, missing-channel completion and brightfield-to-fluorescence translation. It mostly outperforms task-specific methods on image-distribution and biological metrics. Ablations show that the channel embedding is critical and that joint training shares knowledge across datasets and yields gains across tasks, providing a scalable foundation for datasets of heterogeneous channel composition and further generative tasks. Code is available at https://anonymous.4open.science/r/OmniMorph-2EBB/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cell Painting, morphology generation, perturbation prediction, unified generative model, channel-adaptive, flow matching, multi-task learning, high-content screening

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_LLuxOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OmniMorph, reportedly the first unified generative model for Cell Painting, built on flow matching with a DiT backbone and disentangling channel semantics from task role via channel and role embeddings. It is trained jointly in two stages on six public datasets and four tasks (morphology generation, perturbation prediction, missing-channel completion, brightfield-to-fluorescence), conditioned on Uni-Mol/ESM-2 representations and Qwen2.5-14B-encoded instruction text. It leads in most settings and ranks first on all eight unseen-perturbation (OOD) morphology-generation settings.

Critical strengths
  • Clear and elegant problem formulation: the paper identifies heterogeneity in channel count and labeled subcellular structures as the central obstacle, and uniformly formalizes the four tasks as "experimental condition + role assignment". The formalization is clean and leaves room for extension.

  • Unusually rigorous evaluation: frozen target lists and biology panels, ID/OOD splits with a zero-overlap audit, and all baselines retrained by the authors under identical splits and training exposure (Table 20) rather than quoting published numbers.

  • Evaluation beyond image-distribution metrics: beyond FID/KID, MoA/perturbation balanced accuracy is measured with classifiers trained on real images only, plus effect-direction and dose-response analyses. The evaluation spans distribution, discrimination, and biological direction.

  • Thorough and honest ablations: removing the channel embedding raises the mean FID from 50.4 to 152.9 (Table 5a), directly supporting the core design. The text ablation (Table 10) honestly reveals that the caption carries dataset identity.

  • Strong generalization to unseen perturbations: OmniMorph ranks first on all eight OOD morphology-generation settings (Table 1), whereas MorphGen and MDTv2 rely on a closed perturbation vocabulary. This is the strongest evidence for cross-dataset knowledge sharing.

Critical weaknesses
  • Limited methodological novelty, with the text-condition framing overstated: every component is existing technology, and the core novelty—disentangling channel and role embeddings—is closer to systematization than a breakthrough. The claimed "natural-language instruction" is actually template field filling (Table 15), weakening the persuasiveness of the method section.

  • Fairness concerns in the comparison: OmniMorph's total budget is (200k+200k)×8 while each baseline gets only 200k×8 (Table 20), giving the unified model twice the exposure and all six datasets. In perturbation prediction, IMPA is the only baseline actually conditioned on a control image, and CellFlux (ICML 2025), cited as representative, is not compared.

  • Substantive gaps in biological fidelity: dose response is recovered only for the strongest responders (e.g., 3 of 16 RxRx3 compounds reach a trajectory cosine above 0.3). Despite motivating the work as replacing part of screening experiments, no experiment demonstrates the downstream utility of the generated data.

Questions for authors
  • Could you report each baseline at 400k steps (matching the total budget), or OmniMorph trained for only 200k steps? On which metric and which validation set were the reported checkpoints selected?

  • Table 8 shows joint training clearly worse than per-dataset specialists on several FID settings (e.g., CPG0001 OOD 164.4 vs 25.5)—why is this degradation not discussed? Which model version does Table 5(b) actually use, given its numbers match Table 2 while the text says first-stage?

  • If the semantic channel embedding were replaced by a simple dataset-ID embedding (one learnable vector per dataset), how would the metrics change? This is the key control separating channel-semantic sharing from caption-carried dataset identity.

12 comments
AuthorsOpenReview ↗

Response to W1: novelty and the text condition

W1. Limited methodological novelty, with the text-condition framing overstated: every component is existing technology, and the core novelty – disentangling channel and role embeddings – is closer to systematization than a breakthrough. The claimed "natural-language instruction" is actually template field filling (Table 15), weakening the persuasiveness of the method section.

We agree that the individual components exist and that the text is templated, and the revision says so. Our contribution is the formulation that lets one model cover heterogeneous panels and tasks, and the templated text carries real conditioning signal.

Novelty. As stated in the Introduction, our contribution is that, for one model to cover several datasets and tasks, we disentangle the semantics of a channel from its role in the task with two embeddings added to all tokens of the channel. Public Cell Painting datasets differ in their panels and in whether brightfield images and controls are acquired, so a generator for all of them has to accept varying channel combinations and several tasks. With the channel tokens arranged along the sequence and tagged by the structure they stain, a change of panel becomes a change of sequence length, and a channel is identified by what it stains rather than by its position in a dataset. The role embedding then marks each channel as an observed channel, a brightfield input, a control reference or a generation target, so that the four tasks share one model. As the reviewer notes, "the formalization is clean and leaves room for extension".

What the formulation enables. Both embeddings are needed. The channel embedding lowers the mean FID of morphology generation from 152.9 to 50.4 (Table 5(a)), and removing the role embedding raises FID in 10 of the 12 morphology-generation settings, by up to 19.8 (new role-embedding ablation in the appendix). The formulation also gives capabilities that separate models do not have. BBBC021 has only DNA, tubulin and actin, and OmniMorph generates RNA, ER and mitochondria channels for it, learned from the other datasets (new appendix subsection). On JUMP plates unseen in training, the jointly trained OmniMorph separates the compounds best, reaching 85% of the real-image ceiling (new appendix subsection).

Text condition. We agree that the text is produced by templates (Table 15) and that "natural-language instruction" overstates it. The revision calls it a templated text instruction (abstract, Introduction and Section 3). The templates are a deliberate choice. They give every dataset the same format, so that the same field means the same thing across datasets. The text is encoded by a frozen language model (Qwen2.5-14B), so the format can later be extended to free text without changing the model.

The templated text carries real conditioning signal. We trained a model in which the four structured conditions (perturbation type, perturbation identity, dose and cell line) are removed while the text and the role embedding are kept, and compared it with the full-condition model. Both are initialised from the first-stage model and then trained jointly on the four tasks for 100k steps at an effective batch of 8, half of the 200k×8 joint training of the final model (Table 20), so only the comparison between the two is meaningful. On the text-only generations, the frozen real-image classifier of Table 2 identifies the MoA or perturbation above chance on every panel, at 2.5 to 5 times chance on BBBC021, CPG0001 and RxRx3 compounds (Table R17).

Generations ofBBBC021 (MoA, 6)CPG0001 (MoA, 10)CPG0004 (cmpd, 10)RxRx3-C (cmpd, 15)RxRx3-G (gene, 15)
Random0.1670.1000.1000.0670.067
Text only0.4100.3370.1160.3380.101
Real images (ceiling)0.9540.8150.5710.6200.760

Table R17. Balanced accuracy (↑) of the frozen real-image classifier of Table 2 on the generations of the text-only model.

Conversely, the text-condition ablation of Table 10 removes the description of the sample and keeps the task instruction and the structured conditions. The Pearson correlation of channel completion then drops by 0.04 to 0.08 on all four datasets, and morphology generation collapses on CPG0000 and CPG0036, with FID rising from 51–61 to 116–168 and from 93 to over 200. Since the text alone carries perturbation information, a perturbation type for which no dedicated encoder exists can also enter through its text description, without attaching a further foundation model such as ESM-2 or Uni-Mol.

In summary, the components are known, and our contribution is the formulation that disentangles channel semantics from task roles, which the ablations support and which enables channel generation across datasets and transfer to unseen plates. The text condition is templated, as the revision now states, and it carries real conditioning signal.

AuthorsOpenReview ↗

Response to W2 (1/3): training budget and comparison with CellFlux

W2. Fairness concerns in the comparison: OmniMorph's total budget is (200k+200k)x8 while each baseline gets only 200k x8 (Table 20), giving the unified model twice the exposure and all six datasets. In perturbation prediction, IMPA is the only baseline actually conditioned on a control image, and CellFlux (ICML 2025), cited as representative, is not compared.

We thank the reviewer. Joint training does not give OmniMorph more exposure per dataset, and we added CellFlux. Under the budgets of Table 20, OmniMorph sees each training image 2 to 40 times less often than every baseline in all four tasks (Table R2), its first stage alone, without multi-task training, already outperforms the baselines (Tables R1 and R1b), and it is ahead of CellFlux and IMPA in 13 of the 16 perturbation-prediction settings (Table R8).

Data exposure. Each baseline spends its whole budget on the data covered by one set of its weights (Table 20), whereas OmniMorph shares its budget among six datasets and four tasks (Section 3). Table R2 divides the training samples of each method by the number of training images it covers (Table 16). While preparing it, we found a typo in the Models column of Table 20, which the revision corrects.

(a) Morphology generation

BBBC021CPG0-CCPG0-GCPG0001CPG0004CPG0036RxRx3-CRxRx3-G
SD / MorphoDiff / MorphGen / MorphGen*2157710488821701812
Mol2Image21577–888217018–
MDTv2 / MDTv2*215606088821709.29.2
OmniMorph19.610.410.412.612.117.54.14.1
OmniMorph, first stage17.99.49.411.411.015.93.73.7

(b) Channel completion

BBBC021CPG0000CPG0001CPG0004CPG0036RxRx3
U-Net / Pix2Pix / MCD–24.824.824.8–9.2
U-Net / Pix2Pix (validation plate held out)–28.028.028.0–9.3
OmniMorph5.42.83.43.34.81.1

(c) Brightfield-to-fluorescence, per training pair (CPG0000 and CPG0001)

CPG0000 and CPG0001
cWGAN-GP / ISL / CellResDM43.0
cWGAN-GP / ISL (validation plate held out)50.4
OmniMorph12.9

(d) Perturbation prediction

BBBC021CPG0-CCPG0-GCPG0001CPG0004CPG0036RxRx3-CRxRx3-G
IMPA / CellFlux†215606088821709.29.2
OmniMorph5.42.82.83.43.34.81.11.1

Table R2. Average number of times each training image is seen at the budgets of Table 20. A view counts its perturbation type and all control images of the dataset, which gives the baselines the lowest repetition. Where one set of weights covers several columns (scope "source" or "channel count" in Table 20), its value is repeated in each of them. "–" means not covered. First stage is the first-stage morphology-generation training alone. *Retrained with OmniMorph's frozen condition encoders. †Trained on the SD-VAE latents shared by all latent-space methods, with the same frozen Uni-Mol / ESM-2 embeddings as IMPA.

AuthorsOpenReview ↗

Response to W2 (2/3): training budget and comparison with CellFlux

W2 (continued).

Matched data and budget. The first-stage models train morphology generation only, with no more than 200k×8 samples (Table 20). Trained on one dataset at a time, on the same data as a single-dataset baseline, the per-dataset model is ahead of every baseline in 11 of the 16 settings, including 6 of the 8 OOD settings. Since MorphGen* and MDTv2* use the same frozen encoders, this gain comes from the architecture. Trained on all six datasets within the same budget, the multi-dataset model sees each training image 2 to 12 times less often than a baseline and is ahead of every baseline of the original submission in 11 of the 16 settings (Table R1). These two first-stage models are the two rows of Table 8, whose multi-dataset row we correct in our answer to Q2.

MethodBBBC021 IDCPG1 IDCPG4 IDCPG36 IDCPG0-C IDCPG0-G IDRxRx3-C IDRxRx3-G IDBBBC021 OODCPG1 OODCPG4 OODCPG36 OODCPG0-C OODCPG0-G OODRxRx3-C OODRxRx3-G OOD
Stable Diffusion214.9141.0193.7264.8218.1200.1108.7197.7235.6150.9221.5260.0222.9195.9112.6181.9
Mol2Image139.0216.960.8138.0119.9–190.4–128.1211.052.7142.0111.6–186.3–
MorphoDiff142.2136.3128.2277.8202.3192.3116.5209.2180.1117.3150.0272.0208.0190.3120.7192.6
MorphGen60.549.447.782.2110.1106.240.075.8––––––––
MorphGen*52.348.941.586.938.161.628.765.642.342.133.787.734.951.029.162.5
MDTv246.974.826.5104.771.560.353.146.1––––––––
MDTv2*60.4101.032.7118.141.062.9102.690.542.287.633.4111.642.953.7100.389.4
OmniMorph (Table 1)43.950.432.793.552.961.129.525.732.390.027.093.151.250.929.126.3
OmniMorph, per-dataset (first stage)43.838.143.099.036.556.429.929.227.325.529.7100.232.645.731.128.1
OmniMorph, multi-dataset (first stage)47.751.338.3102.256.556.230.730.735.3118.030.4100.348.246.031.329.9

Table R1. Morphology-generation FID (↓) on the targets of Table 1: the revised Table 1 with the two first-stage models added. *With OmniMorph's frozen open encoders. "–": a closed-vocabulary model cannot generate unseen perturbations. Bold: best of each column; the counts in the text compare each first-stage row with the baselines only.

Pretraining and encoders. Neither first-stage model has multi-task training, and the per-dataset model sees no other dataset. The benefit of pretraining for the other three tasks is measured in Table 5(c), where the single-task model trained from scratch is the weakest on every metric. Frozen condition encoders are also used by IMPA and CellFlux† (Uni-Mol / ESM-2) and by MorphGen* and MDTv2* (those of OmniMorph).

Biological fidelity. Without a dataset identifier, the effect of a compound is learned from every dataset in which it occurs. On the classification panels of Table 2, the multi-dataset first-stage model is ahead of every baseline on four of the five panels (Table R1b) and ahead of the per-dataset models on four of the five (Table 5(b)).

BBBC021 (MoA, 6)CPG0001 (MoA, 10)CPG0004 (cmpd, 10)RxRx3-C (cmpd, 15)RxRx3-G (gene, 15)
Real images (ref.)0.9540.8150.5710.6200.760
Stable Diffusion0.1570.0880.0830.0650.081
Mol2Image0.2620.0950.1950.099–
MorphoDiff0.2530.0920.1080.0870.085
MDTv20.2500.2990.1750.2370.242
MDTv2*0.2060.2060.1340.1080.107
MorphGen0.6700.5320.2360.3660.304
MorphGen*0.5760.4620.2170.3720.148
OmniMorph (Table 2)0.7030.6490.2470.4760.526
OmniMorph, multi-dataset (first stage)0.6420.6330.2370.4730.460

Table R1b. Morphology generation, balanced accuracy (↑) of the classifier trained on real images only (Table 2), on RGB renderings. *Retrained with OmniMorph's frozen open encoders. Bold: best generated entry of each column.

Reviewer_avvBOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes OmniMorph, a unified generative model for Cell Painting. The generator is a DiT trained with flow matching, and the central design is a pair of embeddings added to each channel token: a channel embedding that encodes the biological semantics of the stain, and a role embedding that encodes whether the channel is a reference or a generation target, so that heterogeneous datasets and four tasks train jointly under one architecture and one set of weights. For conditioning, four structured encodings (perturbation type, perturbation identity, quantitative attributes, cell line) are combined with a task instruction encoded by a frozen Qwen2.5-14B. With a single model, OmniMorph performs four generation tasks on six datasets, outperforming task-specific models on most of them; on completion and brightfield, the margin over the strongest baseline is small. All baselines were retrained under one shared protocol, and one checkpoint covers every dataset and task.

Critical strengths
  1. The dual embedding decouples channel semantics from task role, which is what lets heterogeneous data and multiple tasks share one architecture.
  2. Evaluation goes beyond image metrics. Beyond FID and Pearson correlation, the paper reports perturbation identity, MoA, and retrieval, and checks the direction and magnitude of generated perturbation effects against negative controls.
  3. The evaluation protocol itself is carefully designed.
Critical weaknesses

The role embedding is the mechanism that makes four tasks one model, and it is the only component without an ablation. §4.4 states that the ablations cover channel embedding, multi-dataset training, and multi-task training. The task instruction in the text condition also names the task, so with both present there is no way to tell whether multi-task behaviour comes from the role embedding or from the text encoder.

Questions for authors
  1. If the role embedding were removed while the task instruction stayed in place, would the four tasks still train jointly? And conversely, if the four structured conditions were removed and only the text description kept — which already states the dataset, the task, and the perturbation in natural language — how much of the conditioning is left?
  2. In Table 1, MDTv2 and MorphGen are better on some datasets under in-distribution perturbations, while OmniMorph leads throughout on unseen perturbations, which the paper attributes to closed-vocabulary models. Is the OOD advantage due to the multi-dataset, multi-task architecture, or to the open encoders (Uni-Mol / ESM-2)? Could the two be separated, for instance by training a single-dataset model with the same open encoders, or by swapping in a closed-vocabulary encoder inside OmniMorph?
  3. What is the parameter count of each component (SD-VAE, Qwen2.5-14B, Uni-Mol, ESM-2), and what is the per-sample inference time for each task? I could not find any parameters or FLOPs in the paper; Table 20 gives only steps × batch and the sampler.
7 comments
AuthorsOpenReview ↗

Response to W1: role-embedding ablation

W1. The role embedding is the mechanism that makes four tasks one model, and it is the only component without an ablation. §4.4 states that the ablations cover channel embedding, multi-dataset training, and multi-task training. The task instruction in the text condition also names the task, so with both present there is no way to tell whether multi-task behaviour comes from the role embedding or from the text encoder.

We added the missing ablation (Table R15). In joint four-task training, the role embedding protects morphology generation. Without it, FID rises in 10 of the 12 settings, by up to 19.8, while the three tasks with an image input are unaffected.

Setup. Both models are initialised from the first-stage model and then trained jointly on the four tasks for 100k steps at an effective batch of 8, half of the 200k×8 joint training of the final model (Table 20). In one of them the role embedding is removed (zeroed and frozen), while the task instruction in the text condition is kept. Both are evaluated on the same frozen targets with the protocols of Tables 1, 3 and 4. Because of the shorter training, only the comparison between the two rows is meaningful.

CPG0001CPG0004CPG0-CCPG0-GRxRx3-CRxRx3-G
With role embedding47.8 / 118.534.4 / 28.253.2 / 48.559.2 / 47.532.2 / 31.824.9 / 26.6
Without role embedding56.9 / 138.334.9 / 31.051.5 / 46.963.4 / 53.137.7 / 37.227.6 / 28.1

Table R15. FID (↓) of morphology generation (ID / OOD), with the task instruction kept in both models. Bold marks the better of each pair. The perturbation-prediction rows are in the appendix table.

Morphology generation. The role embedding assigns each channel one of four roles, three reference roles (observed, brightfield and control) and the generation target (Table 19). It thereby keeps the generation targets of morphology generation apart from the reference channels of the other three tasks. Without it, joint training with those tasks interferes with morphology generation, most strongly on CPG0001 OOD (+19.8).

Tasks with an image input. Channel completion and brightfield-to-fluorescence differ by at most 0.002 in Pearson correlation and SSIM on every dataset of Tables 3 and 4, and perturbation prediction stays within 4.3 FID with no consistent direction. In these tasks the inputs themselves differ (reference channels, brightfield planes or a control image), and together with the instruction they identify the task.

In summary, the instruction tells the model which task it is solving, and the role embedding keeps the channels of the different tasks apart. Together they keep all four tasks intact in one model. The ablation, including the perturbation-prediction results, is added to the appendix of the revision.

AuthorsOpenReview ↗

Response to Q1: role embedding and text-only conditioning

Q1. If the role embedding were removed while the task instruction stayed in place, would the four tasks still train jointly? And conversely, if the four structured conditions were removed and only the text description kept – which already states the dataset, the task, and the perturbation in natural language – how much of the conditioning is left?

We thank the reviewer for these two questions and ran both ablations.

Setup. As in our answer to W1, all models are initialised from the first-stage model and trained jointly on the four tasks for 100k steps at an effective batch of 8, half of the 200k×8 joint training of the final model (Table 20), so only the comparison between rows is meaningful.

Without the role embedding, the four tasks still train jointly. As shown in our answer to W1 (Table R15), the three tasks with an image input are unaffected, while morphology generation loses quality, with FID rising in 10 of the 12 settings.

The text contains no dataset name or identifier. It consists of a task instruction, which names the task and the channels, and a biological context, which states the cell line, the perturbation and the dose (Table 15). No dataset name or identifier appears in the text or anywhere else in the condition. The cell line does not always tell the datasets apart either. CPG0000 and CPG0001 share the cell line (U2OS) and the five Cell Painting channels, so the same perturbation receives the same condition in both.

The text alone carries perturbation information, and the structured conditions add precision. In the text-only model, the four structured conditions are removed, while the text and the role embedding are kept. Since the first-stage model was trained with the full condition, the text-only model has to adapt to the missing input within the shorter training.

CPG0001CPG0004CPG0-CCPG0-GRxRx3-CRxRx3-G
Morphology generation (ID / OOD)
Text + structured47.8 / 118.534.4 / 28.253.2 / 48.559.2 / 47.532.2 / 31.824.9 / 26.6
Text only87.5 / 130.535.8 / 31.574.0 / 68.466.6 / 54.133.2 / 32.426.8 / 32.6
Perturbation prediction (ID)
Text + structured41.235.739.861.430.725.9
Text only44.035.472.863.832.726.5

Table R16. FID (↓) with the full condition and with the text alone. Bold marks the better of each pair.

Generations ofBBBC021 (MoA, 6)CPG0001 (MoA, 10)CPG0004 (cmpd, 10)RxRx3-C (cmpd, 15)RxRx3-G (gene, 15)
Random0.1670.1000.1000.0670.067
Text only0.4100.3370.1160.3380.101
Real images (ceiling)0.9540.8150.5710.6200.760

Table R17. Balanced accuracy (↑) of the frozen real-image classifier of Table 2 on the generations of the text-only model.

On the text-only generations, the classifier identifies the MoA or perturbation above chance on every panel, at 2.5 to 5 times chance on BBBC021, CPG0001 and RxRx3 compounds (Table R17). With the structured conditions, FID is lower in all 12 morphology-generation settings, by up to 39.7 (CPG0001 ID), and in 5 of the 6 perturbation-prediction settings (Table R16). Channel completion is unchanged (Pearson correlation equal to three decimals on all four datasets), and brightfield-to-fluorescence changes by at most 0.004 in SSIM, because both are driven by the observed images.

This also opens a path for extension. Since the text alone carries perturbation information, a perturbation type for which no dedicated encoder exists can enter through its text description, without attaching a further foundation model such as ESM-2 or Uni-Mol. A structured encoder can be added where it brings precision.

The converse ablation shows that the text is needed as well. Table 10 removes the description of the sample from the text and keeps the task instruction and the structured conditions. The Pearson correlation of channel completion then drops by 0.04 to 0.08 on all four datasets, and morphology generation collapses on CPG0000 and CPG0036, with FID rising from 51–61 to 116–168 and from 93 to over 200.

In summary, the two kinds of condition are complementary. The text gives the task and a description of the sample, and the structured conditions pin down the perturbation.

AuthorsOpenReview ↗

Response to Q2: source of the OOD advantage

Q2. In Table 1, MDTv2 and MorphGen are better on some datasets under in-distribution perturbations, while OmniMorph leads throughout on unseen perturbations, which the paper attributes to closed-vocabulary models. Is the OOD advantage due to the multi-dataset, multi-task architecture, or to the open encoders (Uni-Mol / ESM-2)? Could the two be separated, for instance by training a single-dataset model with the same open encoders, or by swapping in a closed-vocabulary encoder inside OmniMorph?

We thank the reviewer for this question and separated the two factors with three new baselines. The open encoders give the OOD ability. With the encoders matched, OmniMorph's channel-wise architecture still gives better image fidelity, and training it across datasets gives better biology.

New baselines. We retrained MorphGen and MDTv2 with OmniMorph's frozen open encoders (MorphGen* and MDTv2*, with the same data, training budget and sampler as MorphGen and MDTv2, Table 20) and added CellFlux (ICML 2025) to perturbation prediction. All new rows are added to Tables 1 and 2 in the revision.

Open encoders. The original MorphGen and MDTv2 cannot generate any unseen perturbation. With OmniMorph's routing of conditions to frozen open encoders, MorphGen* and MDTv2* generate all eight unseen-perturbation settings and improve on Stable Diffusion, Mol2Image and MorphoDiff in every one of them, for example 51.0 and 53.7 against 190.3 for MorphoDiff on CPG0000-G OOD (Table R13). As the reviewer suggests, a large part of the OOD advantage therefore comes from this routing, which is one of the contributions of OmniMorph. Swapping a closed-vocabulary encoder into OmniMorph would make OOD generation impossible by construction, so we answer the question from the other side by giving the open encoders to the baselines.

Architecture. The per-dataset first-stage models of Table 8 use the same frozen encoders and are trained on one dataset at a time, the same data as MorphGen* and MDTv2*. They are ahead of both in 11 of the 16 settings, including 6 of the 8 OOD settings (Table R13).

BBBC021CPG1CPG4CPG36CPG0-CCPG0-GRxRx3-CRxRx3-G
MorphGen60.5 / –49.4 / –47.7 / –82.2 / –110.1 / –106.2 / –40.0 / –75.8 / –
MorphGen*52.3 / 42.348.9 / 42.141.5 / 33.786.9 / 87.738.1 / 34.961.6 / 51.028.7 / 29.165.6 / 62.5
MDTv246.9 / –74.8 / –26.5 / –104.7 / –71.5 / –60.3 / –53.1 / –46.1 / –
MDTv2*60.4 / 42.2101.0 / 87.632.7 / 33.4118.1 / 111.641.0 / 42.962.9 / 53.7102.6 / 100.390.5 / 89.4
OmniMorph, per-dataset (first stage)43.8 / 27.338.1 / 25.543.0 / 29.799.0 / 100.236.5 / 32.656.4 / 45.729.9 / 31.129.2 / 28.1
OmniMorph43.9 / 32.350.4 / 90.032.7 / 27.093.5 / 93.152.9 / 51.261.1 / 50.929.5 / 29.125.7 / 26.3

Table R13. FID (↓) on the targets of Table 1 (ID / OOD). *Retrained with OmniMorph's frozen open encoders. "–" means that a closed-vocabulary model cannot generate unseen perturbations.

Biology. FID measures whether generated images look like real images of the dataset, not whether they carry the right perturbation effect. With the classifier trained on real images only, OmniMorph is ahead of the open-encoder baselines on every panel (Table R14). This gain comes from training across datasets. The multi-dataset first-stage model of Table 5(b) is already ahead of MorphGen* and MDTv2* on every panel, whereas the per-dataset models are not. In perturbation prediction, OmniMorph is ahead of CellFlux in 13 of the 16 FID settings and on 4 of the 5 panels.

BBBC021 (MoA, 6)CPG0001 (MoA, 10)CPG0004 (cmpd, 10)RxRx3-C (cmpd, 15)RxRx3-G (gene, 15)
Real images (ref.)0.9540.8150.5710.6200.760
Morphology generation
MorphGen0.6700.5320.2360.3660.304
MorphGen*0.5760.4620.2170.3720.148
MDTv20.2500.2990.1750.2370.242
MDTv2*0.2060.2060.1340.1080.107
OmniMorph0.7030.6490.2470.4760.526
Perturbation prediction
IMPA0.1640.1350.2110.1260.074
CellFlux0.4080.3380.2410.1900.169
OmniMorph0.7300.5650.2310.4880.467

Table R14. Balanced accuracy (↑) of the classifier trained on real images only, RGB renderings, as in Table 2.

In summary, the open encoders give the OOD ability, the channel-wise architecture gives better image fidelity with the encoders matched, and training it across datasets gives better biology.

Reviewer_M7PyOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OmniMorph, a unified latent flow-matching model for Cell Painting images that addresses heterogeneity across datasets and tasks. The model incorporates channel-aware representations and experimental conditioning to support four tasks: morphology generation, perturbation prediction, missing-channel completion, and brightfield-to-fluorescence translation.

Critical strengths

The paper is generally well-written, well-motivated, and easy to follow. It presents strong empirical results across six datasets and four generative tasks, particularly in morphology generation and perturbation prediction, while also demonstrating biological fidelity through mechanism-of-action (MoA) classification and perturbation-specific morphological features.

Critical weaknesses
  1. Insufficient evidence of generalization. Although the paper emphasizes generalizability across Cell Painting datasets, the current evaluation does not fully support this claim. The model is trained on all six evaluation datasets, and the OOD experiments consider only unseen perturbations within previously observed datasets. A more convincing evaluation would use a leave-one-dataset-out setting, training on five datasets and evaluating on the held-out dataset (e.g., CPG0001 or CPG0004) to assess cross-dataset generalization. Moreover, Table 8 shows that multi-dataset training can substantially degrade FID compared with dataset-specific training, suggesting potential negative transfer. These results raise questions about the generalizability as the authors claimed.

  2. Fairness of baseline comparisons. Related to 1, OmniMorph is trained jointly across datasets and tasks, while most baselines are trained on each dataset separately. Although this demonstrates the advantages of unified modeling, OmniMorph also benefits from broader training data and pretraining. More controlled comparisons under matched data exposure and compute budgets would help isolate the architectural contributions.

  3. Computational efficiency and scalability. OmniMorph processes channel-wise latent tokens as a long sequence, resulting in quadratic self-attention complexity O(n2)O(n^2), where nn is the sequence length. This raises concerns about computational efficiency and scalability, particularly as the number of fluorescence channels increases. The authors should provide comparisons of training and inference time, and model parameters against existing baselines.

  4. Limited methodological novelty. OmniMorph main novelty lies in the design of latent tokens. However, the proposed channel-wise latent tokenization and channel embeddings closely resemble existing approaches. For example, ChannelViT [1] and ChAda-ViT [2] already introduce channel-aware tokenization for heterogeneous microscopy images. Similarly, Lu et al. [3] employ channel-specific embeddings and natural-language conditioning to represent experimental perturbations. Furthermore, [4] already uses independent fluorescence-channel encoding with a pretrained VAE. The authors should better clarify the methodological novelty of OmniMorph.

Reference


  1. Bao, Yujia, Srinivasan Sivanandan, and Theofanis Karaletsos. "Channel vision transformers: An image is worth 1 x 16 x 16 words." ICLR 2024.
  2. Bourriez, Nicolas, et al. "ChAda-ViT: Channel adaptive attention for joint representation learning of heterogeneous microscopy image." 2024 CVPR 2024.
  3. Lu et al., "CellCLIP: Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning," NeurIPS 2025.
  4. Demirel, Berker, et al. "MorphGen: Controllable and morphologically plausible generative cell-imaging." arXiv preprint arXiv 2025
Questions for authors
  1. Why does multi-dataset training degrade FID on certain datasets in Table 8, despite improving biological classification performance?
  2. How does inference cost scale with the number of fluorescence channels?
12 comments
AuthorsOpenReview ↗

Response to W1: generalization to a held-out dataset

W1. Insufficient evidence of generalization. Although the paper emphasizes generalizability across Cell Painting datasets, the current evaluation does not fully support this claim. The model is trained on all six evaluation datasets, and the OOD experiments consider only unseen perturbations within previously observed datasets. A more convincing evaluation would use a leave-one-dataset-out setting, training on five datasets and evaluating on the held-out dataset (e.g., CPG0001 or CPG0004) to assess cross-dataset generalization. Moreover, Table 8 shows that multi-dataset training can substantially degrade FID compared with dataset-specific training, suggesting potential negative transfer. These results raise questions about the generalizability as the authors claimed.

We added a dataset that no model saw in training, the TARGET2 plates of JUMP source_3 (Chandrasekaran et al., 2023; U2OS, the five Cell Painting channels, 25 plates from a different site). This is stricter than leave-one-dataset-out, since none of its plates is in any training set. Applied zero-shot, OmniMorph generates unseen compounds with the lowest FID and identifies compounds best (Table R9, added to the appendix).

All baselines use their CPG0000 models, which share cell line and channels with source_3, and OmniMorph is shown both trained on CPG0000 alone and as the joint model. MorphGen* and MDTv2* are retrained on CPG0000 with OmniMorph's frozen condition encoders, so that they can generate unseen compounds. Each cell holds 512 frozen targets, for seen compounds (104, with CPG0000 training images) or unseen compounds (162, absent from every training set). Compound identification applies a classifier trained on real source_3 images (OpenPhenom features, as in Table 2) to the generations of five compounds with a strong real phenotype.

MethodFID ↓ SeenFID ↓ UnseenCompound ID ↑
Stable Diffusion213.1206.925.3
MorphoDiff183.2186.425.3
MorphGen88.5–24.4
MDTv281.6–27.0
MorphGen*89.775.727.2
MDTv2*93.682.626.9
OmniMorph, CPG0000 only (first stage)84.775.239.0
OmniMorph93.484.443.4
Real images (reference)84.367.551.3

Table R9. Held-out source_3 plates, zero-shot. FID (↓) and compound identification (balanced accuracy, %, ↑, five compounds). Real source_3 images are rescaled per channel to the CPG0000 intensity range. Bold: best of each column.

Trained on CPG0000 alone, like the baselines, OmniMorph has the lowest FID on unseen compounds. On seen compounds its FID (84.7) matches that of real CPG0000 images (84.3), so the remaining distance is the acquisition shift between sites. Both OmniMorph models identify the compounds best (43.4 and 39.0, against at most 27.2), and the joint model reaches 85% of the real-image ceiling.

The degradation in Table 8 is a deliberate trade-off, not negative transfer. The condition contains no dataset identifier, so that the effect of a compound is shared across datasets. CPG0001 shares cell line and channels with CPG0000 but is stained 3–8 times brighter, so the joint model renders it with the CPG0000 appearance. The biology is preserved, as the multi-dataset first-stage model scores 0.633 on the CPG0001 MoA panel, above the per-dataset first-stage model (0.544) and every baseline. A same-plate control image closes the gap (CPG0001 OOD, 90.0 → 52.2 in Table 1). Source_3 shows the same pattern, where the joint model has a higher FID than the CPG0000-only model yet identifies the compounds best. The full analysis is in our answer to Q1.

AuthorsOpenReview ↗

Response to W2 (1/2): data exposure and training budget

W2. Fairness of baseline comparisons. Related to 1, OmniMorph is trained jointly across datasets and tasks, while most baselines are trained on each dataset separately. Although this demonstrates the advantages of unified modeling, OmniMorph also benefits from broader training data and pretraining. More controlled comparisons under matched data exposure and compute budgets would help isolate the architectural contributions.

Joint training does not give OmniMorph more exposure per dataset. Under the budgets of Table 20, it sees each training image 2 to 40 times less often than every baseline in all four tasks (Table R2), and its first stage alone, without multi-task training, already outperforms the baselines (Tables R1 and R1b).

Data exposure. Each baseline spends its whole budget on the data covered by one set of its weights (Table 20), whereas OmniMorph shares its budget among six datasets and four tasks (Section 3). Table R2 divides the training samples of each method by the number of training images it covers (Table 16). While preparing it, we found a typo in the Models column of Table 20, which the revision corrects.

(a) Morphology generation

BBBC021CPG0-CCPG0-GCPG0001CPG0004CPG0036RxRx3-CRxRx3-G
SD / MorphoDiff / MorphGen / MorphGen*2157710488821701812
Mol2Image21577–888217018–
MDTv2 / MDTv2*215606088821709.29.2
OmniMorph19.610.410.412.612.117.54.14.1
OmniMorph, first stage17.99.49.411.411.015.93.73.7

(b) Channel completion

BBBC021CPG0000CPG0001CPG0004CPG0036RxRx3
U-Net / Pix2Pix / MCD–24.824.824.8–9.2
U-Net / Pix2Pix (validation plate held out)–28.028.028.0–9.3
OmniMorph5.42.83.43.34.81.1

(c) Brightfield-to-fluorescence, per training pair (CPG0000 and CPG0001)

CPG0000 and CPG0001
cWGAN-GP / ISL / CellResDM43.0
cWGAN-GP / ISL (validation plate held out)50.4
OmniMorph12.9

(d) Perturbation prediction

BBBC021CPG0-CCPG0-GCPG0001CPG0004CPG0036RxRx3-CRxRx3-G
IMPA / CellFlux†215606088821709.29.2
OmniMorph5.42.82.83.43.34.81.11.1

Table R2. Average number of times each training image is seen at the budgets of Table 20. A view counts its perturbation type and all control images of the dataset, which gives the baselines the lowest repetition. Where one set of weights covers several columns (scope "source" or "channel count" in Table 20), its value is repeated in each of them. "–" means not covered. First stage is the first-stage morphology-generation training alone. *Retrained with OmniMorph's frozen condition encoders. †Trained on the SD-VAE latents shared by all latent-space methods, with the same frozen Uni-Mol / ESM-2 embeddings as IMPA.

AuthorsOpenReview ↗

Response to W2 (2/2): data exposure and training budget

W2 (continued).

Matched data and budget. The first-stage models train morphology generation only, with no more than 200k×8 samples (Table 20). Trained on one dataset at a time, on the same data as a single-dataset baseline, the per-dataset model is ahead of every baseline in 11 of the 16 settings, including 6 of the 8 OOD settings. Since MorphGen* and MDTv2* use the same frozen encoders, this gain comes from the architecture. Trained on all six datasets within the same budget, the multi-dataset model sees each training image 2 to 12 times less often than a baseline and is ahead of every baseline of the original submission in 11 of the 16 settings (Table R1). These two first-stage models are the two rows of Table 8, whose multi-dataset row we correct in our answer to Q1.

MethodBBBC021 IDCPG1 IDCPG4 IDCPG36 IDCPG0-C IDCPG0-G IDRxRx3-C IDRxRx3-G IDBBBC021 OODCPG1 OODCPG4 OODCPG36 OODCPG0-C OODCPG0-G OODRxRx3-C OODRxRx3-G OOD
Stable Diffusion214.9141.0193.7264.8218.1200.1108.7197.7235.6150.9221.5260.0222.9195.9112.6181.9
Mol2Image139.0216.960.8138.0119.9–190.4–128.1211.052.7142.0111.6–186.3–
MorphoDiff142.2136.3128.2277.8202.3192.3116.5209.2180.1117.3150.0272.0208.0190.3120.7192.6
MorphGen60.549.447.782.2110.1106.240.075.8––––––––
MorphGen*52.348.941.586.938.161.628.765.642.342.133.787.734.951.029.162.5
MDTv246.974.826.5104.771.560.353.146.1––––––––
MDTv2*60.4101.032.7118.141.062.9102.690.542.287.633.4111.642.953.7100.389.4
OmniMorph (Table 1)43.950.432.793.552.961.129.525.732.390.027.093.151.250.929.126.3
OmniMorph, per-dataset (first stage)43.838.143.099.036.556.429.929.227.325.529.7100.232.645.731.128.1
OmniMorph, multi-dataset (first stage)47.751.338.3102.256.556.230.730.735.3118.030.4100.348.246.031.329.9

Table R1. Morphology-generation FID (↓) on the targets of Table 1: the revised Table 1 with the two first-stage models added. *With OmniMorph's frozen open encoders. "–": a closed-vocabulary model cannot generate unseen perturbations. Bold: best of each column; the counts in the text compare each first-stage row with the baselines only.

Pretraining and encoders. Neither first-stage model has multi-task training, and the per-dataset model sees no other dataset. The benefit of pretraining for the other three tasks is measured in Table 5(c), where the single-task model trained from scratch is the weakest on every metric. Frozen condition encoders are also used by IMPA and CellFlux† (Uni-Mol / ESM-2) and by MorphGen* and MDTv2* (those of OmniMorph).

Biological fidelity. Without a dataset identifier, the effect of a compound is learned from every dataset in which it occurs. On the classification panels of Table 2, the multi-dataset first-stage model is ahead of every baseline on four of the five panels (Table R1b) and ahead of the per-dataset models on four of the five (Table 5(b)).

BBBC021 (MoA, 6)CPG0001 (MoA, 10)CPG0004 (cmpd, 10)RxRx3-C (cmpd, 15)RxRx3-G (gene, 15)
Real images (ref.)0.9540.8150.5710.6200.760
Stable Diffusion0.1570.0880.0830.0650.081
Mol2Image0.2620.0950.1950.099–
MorphoDiff0.2530.0920.1080.0870.085
MDTv20.2500.2990.1750.2370.242
MDTv2*0.2060.2060.1340.1080.107
MorphGen0.6700.5320.2360.3660.304
MorphGen*0.5760.4620.2170.3720.148
OmniMorph (Table 2)0.7030.6490.2470.4760.526
OmniMorph, multi-dataset (first stage)0.6420.6330.2370.4730.460

Table R1b. Morphology generation, balanced accuracy (↑) of the classifier trained on real images only (Table 2), on RGB renderings. *Retrained with OmniMorph's frozen open encoders. Bold: best generated entry of each column.

In summary, under matched data and budget OmniMorph is ahead of the baselines in most settings, and since its design lets the datasets reinforce one another, joint training makes the MoA and perturbation of its generations more accurately identifiable.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

Dear Area Chairs and Reviewers,

We thank all reviewers for their careful reviews and constructive comments. OmniMorph is, to our knowledge, the first unified generative model for Cell Painting. Each channel is identified by the structure it stains and by its role in the task, so one model with one set of weights performs four generation tasks on six datasets with different channels and takes flexible combinations of input and output channels. We are grateful that the reviewers recognized these strengths.

  • The channel and role embeddings decouple channel semantics from task role, so that heterogeneous data and multiple tasks share one architecture (avvB).
  • The results are strong across six datasets and four tasks, with biological fidelity verified through MoA classification (M7Py).
  • The problem formulation is clean, the evaluation is rigorous, and the generalization to unseen perturbations is strong (LLux).

During the discussion, we added the following.

  • Extensibility and downstream use. A new dataset can join training without changing the architecture, and thanks to the flexible channel design, OmniMorph can generate channels that a dataset never acquired, such as RNA, ER and mitochondria channels for BBBC021, learned from the other datasets. When real data are scarce, adding OmniMorph's generations improves the classification of real images. (To LLux)
  • Cross-dataset and unseen-perturbation generalization. Joint training on six datasets exposes OmniMorph to more perturbations, and together with the open encoders this lets it generalize to unseen perturbations and datasets. Applied zero-shot to a dataset held out from training, it generates unseen compounds with the lowest FID and identifies compounds best (Table R9). (To M7Py)
  • Fair comparison and new baselines. Beyond the extensive comparisons in the paper, we added more experiments on fairness. Joint training does not give OmniMorph more exposure per dataset, and trained on one dataset at a time with the same data as the baselines, OmniMorph still leads in most settings. We also added the baselines the reviewers suggested, CellFlux (ICML 2025), and MorphGen* and MDTv2* with our open encoders. (To M7Py, LLux, avvB)
  • Source of the OOD advantage. The open encoders give the ability to generate unseen perturbations, complemented by the text condition, which alone can already generate them (Table R16). With the same encoders, OmniMorph's architecture, which encodes each channel separately, gives better image fidelity, and training across datasets gives more accurate biology (Tables R13, R14). (To avvB)
  • New ablations. We added two ablations, removing the role embedding and replacing the channel embedding with one learnable vector per dataset. The role embedding protects morphology generation in joint training, because unlike the other three tasks it has no reference image to rely on (Table R15). With one vector per dataset, the model falls far behind, so identifying each channel by the structure it stains is what lets the datasets share knowledge. (To avvB, LLux)
  • Analysis of the CPG0001 case. Our new analysis shows that CPG0001, where multi-dataset training raises FID, is not a failure case. CPG0001 and CPG0000 share the cell line and all five channels but were acquired differently. Since the condition deliberately carries no dataset identifier, the model renders CPG0001 with the appearance of CPG0000, and FID mostly measures this difference. The biology is preserved, with better CPG0001 MoA identification than the per-dataset model. (To M7Py, LLux)
  • Computational cost. We report the parameters of each component and the inference time of each task, analyze how the cost grows with the number of channels, and acknowledge that it exceeds that of the baselines. (To M7Py, avvB)
  • Novelty. The related works the reviewers pointed out inspired our design, but they address different tasks. ChannelViT and ChAda-ViT target image understanding and CellCLIP targets retrieval, so none of them generates images, and the generator of MorphGen is tied to one fixed channel panel. OmniMorph performs four generation tasks on datasets with different channels in one model, which is why it separates the semantics of a channel from its role. (To M7Py, LLux)
  • Evaluation details. We clarified the checkpoint selection and the model version used in Table 5(b). (To LLux)

The revised manuscript adds the new baselines to Tables 1 and 2 and a row to Table 7, corrects Tables 5(b) and 8, and adds to the appendix the held-out-dataset experiment, the generation of unacquired channels, the CPG0001 analysis, the computational cost and the role-embedding ablation. All changes are marked in red.

We have carefully added extensive experiments, our individual responses answer each question in detail, and we welcome further discussion.

Thank you again for your insightful suggestions, which have greatly improved the quality of our paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.