- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper in inscribed in a growing body of works that attempts to improve FM by pairing targets with an ‘optimal’ source distribution. Instead of simultaneously learning a flow and an unsupervised transformation on the initial distribution, the authors propose to decouple the process. The goal is to obtain condition dependent sources to improve prediction alignment with the guiding condition.
Critical strengths
Experiments are consistent with claims, and the idea is original and fills a gap in the existing literature. Decoupling is shown to improve both image-text consistency and, somewhat surprisingly, overall image quality.
Conditioning modality transferability shown in section 5 is an interesting application, with great potential to domains outside image/text.
Presentation is good. The submission is clearly written (except for section 4) and structured, and the progression of ideas is natural and readable.
Critical weaknesses
The motivation is rather light. The advantages of decoupling are shown with a toy experiment, but a theoretical analysis is missing.
Ablations are missing, and section 4 is hard to follow, with unstated assumptions and unjustified design choices. Why concatenating aligned and unaligned image representations in section 4? How can you guarantee that the better performance shown in table 1 is not due to increased expressivity of the higher dimensional space you are using? In my opinion this is the weakest point of the manuscript, and I'm willing to raise my score if this point is addressed.
Questions for authors
Consider the naïve decoupling approach, where we use the Representation Autoencoder paradigm with CLIP or SigCLIP as encoder. Given a text conditioning signal, we encode it, revert the flow, and obtain an initial sample z0. Have you compared against this or a similar baseline?
Can you show FID, precision and recall for figure 1b? What does D_x stand for in equation (9)?
I’m confused as to what motivates the factorized image representation in section 4. In Reg a single global token is added, but here its many more right? This raises the question, given that decoupling as explained in section 3 could be achieved with a single representation, why increse the number of dimensions? Is it possible to train FlowSeed without an unaligned component? Do you have experiments on this?
Any intuition on what does the push-forward measure under your method look like?
Response to Reviewer EHkA.
Thank you for recognizing the originality and conditioning-transfer potential of FlowSeed. We directly address your main concern: whether the gain comes from decoupled source learning or increased representation capacity.
W2 / Q3: Dimensionality, factorization, and component contributions.
In DINOv2/SigLIP, we add one global semantic token alongside the original 256 local DINO tokens, as in REG. The global component provides an aligned interface for source transfer; local tokens retain spatial detail and stochastic variation. The following controls separate their roles. A/B use the same data, backbone design, and training protocol; B-D share one frozen image-flow checkpoint, representation dimensions, decoder, and conditioning. All entries are re-evaluated for this control.
Adding the global representation gives +1.4 GenEval points (B-A). Learned initialization alone adds +3.9 GenEval/+1.5 DPG (C-B), before FlowEdit adds +1.5/+0.7 (D-C). Thus, D improves over B by 5.4/2.2 points at fixed dimensionality and with the same image generator. FID changes from 2.77 to 2.81, showing a small image-quality trade-off for the semantic improvement. Each FID uses its indicated sampler, including the full sampler for D.
The unaligned component is also not required for applicability. Table 1's single-space CLIP setting gives Standard FM/FlowSeed GenEval 58.4/64.7, DPG 61.2/70.8, and FID 3.45/3.10, without concatenating an unaligned representation. Both methods share the DALL-E 2-style decoder within that setting. These results establish single-space feasibility; the matched controls above isolate contributions in the main setting.
Q1: Comparison with the proposed naive decoupling baseline.
We tested your baseline in the single-space CLIP setting. Both variants start from the same text embedding and use the same forward image flow, decoder, and forward sampling budget, without FlowEdit. Only the inverse mapping changes:
Independent text-flow inversion improves GenEval/DPG by 4.1/6.1 points and reduces FID by 0.40. This isolates its benefit from FlowEdit and representation expansion. Both use 50 inverse and 50 forward steps, totaling 100 velocity-field evaluations; this matches evaluation counts, not the computational cost of the differently sized flows.
Q2: FID, precision/recall, and the meaning of Dx.
We supplemented the capacity study at 80M and 160M, spanning the CSFM optimum and subsequent degradation. Sample counts, reference images, and evaluators are identical.
CSFM loses both fidelity and coverage beyond 80M. FlowSeed instead improves FID, precision, and recall while its prompt-following scores increase (Table 2). This supports scalable source learning beyond semantic metrics alone. In Eq. (9), Dx is the x-modality reconstruction decoder used during representation alignment.
W1 / Q4: Mechanism, assumptions, and push-forward.
Jointly updating a learned source changes both interpolated states and sample-level velocity targets x1−z0. We hypothesize that this interaction explains the observed scaling limitation. FlowSeed removes this optimization coupling: source learning leaves the standard image-FM objective and learned image vector field fixed.
For explicit two-stage initialization, let zc=Tc(Etext(c)). Then q0(⋅∣c)=δzc⊗N(0,Ilocal) and p_{\rm FS}(\cdot\mid c)=(G_c)_{\\#}q_0(\cdot\mid c). Semantic initialization is condition dependent while local randomness remains. FlowEdit uses a separate velocity-difference sampler; exact equivalence to inversion is not assumed.
For the toy's unconditional flows, identical aligned marginals, Gaussian priors, independent couplings, and FM paths define the same population objective. This motivates compatible transports; practical alignment is supported empirically, not guaranteed pointwise. Together, the matched controls and naive baseline directly support a source-learning contribution independent of added dimensions or FlowEdit.