Pith. sign in

REVIEW 5 major objections 3 minor 114 references

This paper argues that pruning diffusion models should be guided by spatial importance maps—where the prompt matters—rather than uniform reconstruction error, and shows that reweighting the OBS Hessian with such maps preserves subject fidel

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 10:55 UTC pith:2JY23BPS

load-bearing objection A genuinely new tweak to OBS pruning for diffusion models—weight the Hessian by a spatial importance map—with consistent but small gains, and a causal story that needs a semantic-content control before it fully lands. the 5 major comments →

arxiv 2607.20048 v1 pith:2JY23BPS submitted 2026-07-22 cs.CV

Importance-Aware OBS Pruning for Diffusion Models

Ba-Thinh Lam , Srijan Das , Hieu Le This is my paper
classification cs.CV
keywords diffusion modelsone-shot pruningOptimal Brain Surgeonclassifier-free guidancespatial importance mapsmodel compressiontext-to-image generationsaliency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Diffusion models are expensive, and one-shot pruning is appealing, but standard second-order pruning (OBS) treats every image region equally, so at high sparsity it visibly damages the subjects the prompt cares about. This paper proposes importance-aware OBS: weight the reconstruction objective by a spatial importance map—typically the magnitude of classifier-free guidance—so that the Hessian is accumulated from importance-filtered activations. Parameters whose effects concentrate in prompt-salient regions then get larger saliency and are retained. The authors show across two text-to-image models and several sparsity levels that this simple substitution improves subject fidelity, structural coherence, and human preference scores versus OBS-Diff, while adding no retraining or extra inference cost. The broader claim is that pruning should optimize preservation of what matters perceptually, not average reconstruction loss.

Core claim

The paper's central claim is that injecting external spatial importance signals into the OBS saliency computation produces a pruning criterion aligned with perceptual relevance. Formally, for token-wise linear layers, replacing the reconstruction objective with a spatially weighted version is equivalent to a standard OBS problem on importance-filtered activations, giving a Hessian accumulated from these filtered activations. Parameters with activations concentrated in high-importance regions accumulate larger curvature and are preferentially retained. The paper reports consistent improvements over OBS-Diff on two text-to-image backbones at 40–60 percent sparsity, with the largest gains on hu

What carries the argument

The central object is the spatial importance map M_t, computed as the channel-averaged magnitude of the classifier-free guidance delta, normalized per sample and timestep, and scaled into a weighting map A_t. The argument turns on the identity that for token-wise linear layers, weighting the reconstruction error by A_t is equivalent to replacing activations X with importance-filtered X' in the OBS Hessian. This one substitution changes which parameters are deemed salient—those whose influence concentrates in important spatial regions—while preserving the closed-form OBS weight-update structure and its computational complexity.

Load-bearing premise

The argument assumes the spatial importance map—typically the CFG delta—reliably marks where the content of the image matters; if the map is noisy, incomplete, or mislocalized, as the authors concede for abstract prompts, small objects, and cluttered scenes, the reweighted Hessian can direct pruning toward the wrong parameters.

What would settle it

Find or construct a prompt whose CFG map is concentrated on a distractor region (e.g., a busy background) while human-rated importance lies on the subject; if pruning guided by that map degrades the subject more than uniform OBS does, the central claim fails. More directly, measure per-region reconstruction error after pruning: the paper's mechanism predicts errors concentrate in low-importance regions, and a violation of that prediction would refute the claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim is right, one-shot, training-free pruning of text-to-image models can be steered toward preserving prompt-relevant content, reducing the visible collapse of subjects at sparsity levels where uniform-error pruning fails.
  • Because the importance signal is plug-in, the same pipeline can preserve edge structure, object-focused content, or user-specified categories just by swapping the map or the calibration set.
  • Category-targeted calibration plus spatial guidance offers a practical way to specialize a compressed model for a domain—such as preserving people or vehicles—at the cost of other content.
  • The principle generalizes beyond pruning: any objective that weights reconstruction by perceptual importance should shift compression artifacts away from salient regions.
  • The method inherits the OBS pipeline, so it can be combined with other acceleration techniques like token merging or quantization for further speedups.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The exact equivalence between spatially weighted reconstruction error and importance-filtered activations holds for token-wise linear layers only; for convolutions and patch-embedding layers it is an approximation, so the theoretical grounding is strongest for transformer blocks.
  • A natural stress test is to measure where reconstruction error actually concentrates after pruning: the mechanism implies errors shift into low-importance regions, and a violation of that prediction would refute the mechanism.
  • Because CFG maps are computed during sampling, the approach could be extended to per-prompt or per-image adaptive pruning, re-ranking weights for each generation at some extra cost.
  • The framework suggests a broader principle for generative-model compression—objectives should be content-aware—which might also be applied to quantization or distillation, not just pruning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 3 minor

Summary. The paper proposes importance-aware OBS pruning for text-to-image diffusion models. It extends the OBS-Diff framework by reweighting the per-layer Hessian with a spatial importance map derived from classifier-free guidance (CFG) deltas (Eqs. 3–8). The resulting saliency scores (Eq. 9) are claimed to align pruning with perceptual relevance rather than uniform reconstruction error. Experiments on SD3-Medium and PixArt-Σ report improvements over OBS-Diff in CLIP Score, ImageReward, and MUSIQ for unstructured, structured, and category-targeted pruning, plus a small user study in Appendix A. The derivation in Appendix C is a standard OBS derivation with the Hessian computed on importance-filtered activations.

Significance. If the central claim were established, the contribution would be a simple, training-free, modular extension to a state-of-the-art pruning method: one only needs to substitute X_{l,t} with A_t⊙X_{l,t} in the Hessian. The paper is honest about limitations (Sec. 5), includes a user study, and the algebraic derivation is internally consistent for token-wise linear layers. However, the significance is currently limited by the lack of a control for the semantic content of the importance map and by weak statistical evidence: all headline tables report single-run means, and the ablation (Table 6) does not support the λ value used in the main experiments.

major comments (5)
  1. [§4.2, Tables 1–5] All headline quantitative results are single-run means with no error bars, confidence intervals, or significance tests. ImageReward deltas are typically 0.03–0.09 (e.g., Table 1: 0.71→0.76 at 50%; Table 2: 0.49→0.52 at 60%), which is within the noise one would expect from the stochastic sampling and calibration procedure. Without repeated runs or paired statistics, the claimed consistent improvements are not established. Please add multiple seeds/runs and report standard errors or confidence intervals, ideally with paired tests across the same prompts.
  2. [§4.6, Table 6 vs. §4.1] The main experiments set λ=1.0 for all results, but the ablation in Table 6 shows that λ∈[0.1,0.3] gives the best ImageReward (0.54–0.55) while λ=1.0 gives 0.52. Thus the reported numbers use a suboptimal hyperparameter relative to the authors' own ablation. This also affects the interpretation: a strong non-uniform mask at λ=1.0 can produce large effects even if its spatial layout is semantically meaningless. Please either tune λ on a validation split and report results at the tuned value, or justify why λ=1.0 is the appropriate operating point despite the ablation.
  3. [§3.2, §3.3, Eqs. (3)–(8)] The central scientific claim is that the semantic content of the CFG-delta map M_t, not merely the act of non-uniformly weighting the Hessian, drives the improvements. The paper only compares against uniform OBS-Diff. To support the semantic-content interpretation, the authors should include a non-semantic control: e.g., a random permutation of the same M_t per sample/timestep, or a mask with identical spatial statistics but shuffled locations. If any non-uniform reweighting improves over OBS-Diff, the reported gains could instead arise from reweighting as a regularizer or an effective calibration-distribution shift. This control is essential to the abstract's claim that the method produces saliency estimates 'aligned with perceptual relevance rather than uniform reconstruction error.'
  4. [§3.3, Eq. (8), and Appendix C] The identity ‖A_t⊙(ΔW X_{l,t})‖² = ‖ΔW(A_t⊙X_{l,t})‖² is exact only for token-wise linear layers, as the paper notes in passing. However, the experimental sections do not specify how A_t is handled in convolutional layers, patch embeddings, or other non-token-wise operations in SD3-Medium and PixArt-Σ. If pruning is applied to such layers, the derivation does not directly apply. Please state explicitly which layer types are pruned and how the importance map is aligned or resampled for each, or restrict the method's claim to the layers where Eq. (8) is exact.
  5. [§4.4, Table 4] The category-targeted results are not uniformly consistent with the paper's claim of improvement. For example, in the Airplane row, Ours (target) has CLIP 28.76 vs. OBS-Diff (general) 29.35, and in the Cat row, Ours (target) has ImageReward 0.34 vs. Ours (general) 0.41. The MUSIQ gains are also mixed (e.g., Cat: Ours general 69.52 vs. Ours target 71.64, but Airplane: Ours general 68.19 vs. Ours target 70.81, with reversed CLIP/ImageReward trends). The text says these results 'support' the hypothesis, but the table shows trade-offs rather than a consistent win. Please discuss these trade-offs and add significance testing for the category-targeted setting.
minor comments (3)
  1. [Throughout] Repeated typos: 'resutls' in several figure captions (e.g., Figs. 7–41), 'ImgRw.' in Table 6, and inconsistent naming ('GCC3M' vs. the referenced Conceptual Captions dataset [108]).
  2. [Appendix A, Table 8] The per-sample alignment table shows ImageReward agrees with human preference in only 50% of the 20 samples, and MUSIQ in 70%. This is a useful honesty check, but it also undermines the paper's emphasis on ImageReward as a 'human-aligned perceptual quality metric.' Please contextualize these alignment rates and avoid over-relying on ImageReward in the main claims.
  3. [§4.1] The hyperparameter setting says CFG scale during pruning is 7.0 and 4.5 for SD3-Medium and PixArt-Σ, respectively, and 7.0 for both during testing. This asymmetry is not motivated; please explain why the pruning-time CFG scale differs from the test-time scale, or use a consistent setting.

Circularity Check

0 steps flagged

No significant circularity: the importance-weighted Hessian is an explicit algebraic extension of OBS, and the empirical claims are tested against an external baseline.

full rationale

The derivation chain is self-contained. The method modifies the layer-wise reconstruction objective of Eq. (1) by inserting a spatial importance weighting At, giving Eq. (7). For token-wise linear layers the key step Wl(At ⊙ Xl,t) = At ⊙ (Wl Xl,t) is stated in §3.3 and derived in Appendix C, yielding the importance-filtered Hessian of Eqs. (8) and the saliency of Eq. (9). This is an explicit algebraic substitution, not a quantity fitted from the evaluation metrics (CLIP Score, ImageReward, MUSIQ), so there is no fitted-input-called-prediction reduction. The only scalar hyperparameter λ is set to 1.0 across all main experiments (§4.1), and the ablation in Table 6 actually shows that value is not the tuned optimum (λ ∈ [0.1, 0.3] is better), which works against cherry-picking rather than constituting circularity. The paper contains self-citations ([12], [14], [90] include an author of the present paper), but they appear only in background and motivation; they are not used to derive the saliency formula, to forbid alternatives, or to supply a load-bearing uniqueness claim. The paper also openly acknowledges in §5 that CFG-based importance maps can be noisy or incomplete for abstract prompts and cluttered scenes; this is an honest validity caveat, not a circular step. The central empirical claim is evaluated against OBS-Diff [1] on MS-COCO and with human-aligned metrics, so the reported improvements are not equivalent by construction to the method's inputs. Overall, no circularity is found; at most there are non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The important signal M_t is a derived quantity from existing model outputs (CFG), not a new mechanism. The key free parameters are λ and the inherited timestep schedule; the axioms are the standard OBS layer-wise assumption plus the domain heuristic that CFG magnitude equals perceptual importance.

free parameters (3)
  • λ (importance mask weight) = 1.0 (main experiments); ablation optimum 0.1-0.3
    Controls the strength of spatial importance weighting in A_t=λM_t. Fixed at 1.0 in §4.1, but Table 6 shows λ=0.1-0.3 is best on PixArt-Σ at 60% sparsity; a tuned hyperparameter.
  • α_t (timestep weighting schedule) = Logarithmically decreasing, per OBS-Diff [1]
    Timestep reweighting of the layer-wise reconstruction objective (Eq. 1); inherited from the baseline and affects saliency magnitudes.
  • CFG scale during pruning = 7.0 (SD3-Medium), 4.5 (PixArt-Σ)
    Sets the magnitude of the CFG delta used to build importance maps; chosen per model, not ablated.
axioms (4)
  • domain assumption Layer-wise OBS approximation: each output row is pruned independently and cross-layer error propagation is ignored.
    Inherited from OBS/OBS-Diff; stated in §3.1 and Appendix C; the entire saliency framework depends on this.
  • domain assumption The quadratic surrogate is an accurate proxy for the true generation loss of the diffusion model.
    Motivated in §1: the model is not at a local optimum, so OBS saliency is approximate; the paper's premise is that this approximation can be steered by importance maps.
  • ad hoc to paper CFG delta magnitude |ϵθ(x,t,c)−ϵθ(x,t,∅)| measures perceptual importance of spatial regions.
    Eqs. (3)-(4) define the importance map directly from CFG; the paper cites [46,47] for the related claim that CFG highlights semantic regions, but the map itself is a heuristic adopted for this method.
  • standard math For linear layers, spatial weighting can be absorbed into activations (X' = A⊙X).
    ‖A⊙(ΔWX)‖² = ‖ΔW(A⊙X)‖² holds for any linear map acting on the channel axis; exact for token-wise linear layers, approximate for convolutions/attention with spatial kernels.

pith-pipeline@v1.3.0-alltime-deepseek · 21146 in / 15254 out tokens · 106998 ms · 2026-08-01T10:55:22.834549+00:00 · methodology

0 comments
read the original abstract

We propose importance-aware pruning for diffusion models, a training-free framework that prioritizes preserving parameters critical to semantically salient image regions. To do so, we incorporate spatial importance maps -- derived from conditioning signals or model attention -- into the pruning objective. This produces parameter rankings aligned with perceptual relevance rather than uniform reconstruction error. On MS-COCO dataset, our proposed approach consistently retains subject fidelity and structural correctness at high compression ratios where conventional pruning causes visible degradation. These results demonstrate that content-aware objectives are key to perceptually faithful compression of generative models.

Figures

Figures reproduced from arXiv: 2607.20048 by Ba-Thinh Lam, Hieu Le, Srijan Das.

Figure 1
Figure 1. Figure 1: Qualitative comparison on PixArt-Σ for Structured Pruning. Our method (bottom) preserves subject integrity substantially better than the OBS-Diff baseline (top) by retaining semanti￾cally important regions, highlighted by red boxes. We achieve this by injecting importance signals such as classifier-free guidance (CFG) [2] into the pruning objective. Abstract We propose importance-aware pruning for diffusio… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison across different sparsity levels on SD3-Medium. The top row shows OBS-Diff, which exhibits artifacts and loss of subject coherence as sparsity increases. The bottom row shows our Importance-Aware OBS, which better preserves the structural and semantic integrity of the generated image by prioritizing critical feature regions [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison across different sparsity levels on PixArt-Σ for Unstructured Pruning. The top row shows OBS-Diff, which exhibits artifacts and loss of subject coherence as sparsity increases. The bottom row shows our Importance-Aware OBS, which better preserves the structural and semantic integrity of the generated image by prioritizing critical feature regions. PixArt-Σ. As shown in [PITH_FULL_IM… view at source ↗
Figure 4
Figure 4. Figure 4: Pruned Models with Different Important Signals. Comparison of the OBS-Diff baseline against our importance-aware variants (CFG, CFG+Detector, and Canny). The second row visualizes the corresponding saliency masks. Prompt: "black Honda motorcycle parked in front of a garage." Tab. 4 shows that category-specific calibration already steers pruning toward the target concept. At 60% sparsity, Ours (target) outp… view at source ↗
Figure 5
Figure 5. Figure 5: Target-category pruning at 60% sparsity. Category-targeted pruning improves preserva￾tion of the specified category, and our method further improves target fidelity and visual quality. 4.5 Structured Pruning Our formulation can be directly extended to structured pruning by following the aggregation strategy in OBS-Diff [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparisons highlighting the misalignment between automated metrics and human [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A young man and his cute cat enjoy a nap together." 21 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A kitten sitting in a sink with a green brush with green bristles." 45% Sparsity 50% Sparsity 55% Sparsity 60% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A black Honda motorcycle parked in front of a garage." D Broader Impact This work contributes to making text-to-image diffusion models more efficient, practical, and accessi￾ble. By reducing model parameters while preserving percep… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A motorcycle is parked on a dirt road in a forest." 45% Sparsity 50% Sparsity 55% Sparsity 60% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A customized motorcycle with a large rear and skinny front tire." semantically important regions is often more valuable than uniformly minimizing reconstruction error everywhere. The proposed framework is also model-agnostic and e… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "a shop a car traffic lights and buildings." 45% Sparsity 50% Sparsity 55% Sparsity 60% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A beautiful dessert waiting to be shared by two people." carelessly. More generally, controllable compression creates a new axis of model specialization, which should be evaluated not only by average metrics but also by subgroup r… view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A cat eating a bird it has caught." 45% Sparsity 50% Sparsity 55% Sparsity 60% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A cat at attention between two parked cars." 25 [PITH_FULL_IMAGE:figures/full_fig_p025_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A dog sitting between its masters feet on a footstool watching tv." 30% Sparsity 40% Sparsity 45% Sparsity 50% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative resutls on SD3-Medium. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 30%, 40%, 45%, and 50% unstructured sparsity. Prompt: "A man walking beside sheep on a country road." 26 [PITH_FULL_IMAGE:figures/full_fig_p026_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Qualitative resutls on SD3-Medium. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 30%, 40%, 45%, and 50% unstructured sparsity. Prompt: "An American Airlines plane is in the sky." 30% Sparsity 40% Sparsity 45% Sparsity 50% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p027_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Qualitative resutls on SD3-Medium. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 30%, 40%, 45%, and 50% unstructured sparsity. Prompt: "This new fridge goes great in this clean kitchen." 27 [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Qualitative resutls on SD3-Medium. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 30%, 40%, 45%, and 50% unstructured sparsity. Prompt: "Two people riding a motorcycle to the beach." 30% Sparsity 40% Sparsity 45% Sparsity 50% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Qualitative resutls on SD3-Medium. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 30%, 40%, 45%, and 50% unstructured sparsity. Prompt: "A silver car in the street next to a metal railing." 28 [PITH_FULL_IMAGE:figures/full_fig_p028_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Target-Category Pruning: woman. Comparison of general and targeted pruning baselines (OBS-Diff) against our importance-aware variants at 60% sparsity level. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Target-Category Pruning: cat. Comparison of general and targeted pruning baselines (OBS-Diff) against our importance-aware variants at 60% sparsity level. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Target-Category Pruning: motorcycle. Comparison of general and targeted pruning baselines (OBS-Diff) against our importance-aware variants at 60% sparsity level. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Target-Category Pruning: airplane. Comparison of general and targeted pruning baselines (OBS-Diff) against our importance-aware variants at 60% sparsity level. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A black Honda motorcycle parked in front of a garage." 20% Sparsity 30% Sparsity 40% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p033_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A Honda motorcycle parked in a grass driveway." 33 [PITH_FULL_IMAGE:figures/full_fig_p033_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A shot of an elderly man inside a kitchen." 20% Sparsity 30% Sparsity 40% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p034_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A cat in between two cars in a parking lot." 34 [PITH_FULL_IMAGE:figures/full_fig_p034_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "a man sleeping with his cat next to him" 20% Sparsity 30% Sparsity 40% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p035_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A cat stands between two parked cars on a grassy sidewalk." 35 [PITH_FULL_IMAGE:figures/full_fig_p035_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A long haired cat eating a dead bird." 20% Sparsity 30% Sparsity 40% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p036_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "a stripped cat sitting near a brick wall" 36 [PITH_FULL_IMAGE:figures/full_fig_p036_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A giraffe and a zebra checking each other out." 20% Sparsity 30% Sparsity 40% Sparsity OBS-Diff Ours [PITH_FULL_IMAGE:figures/full_fig_p037_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "A red truck has a black dog in the drivers chair." 37 [PITH_FULL_IMAGE:figures/full_fig_p037_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: Qualitative resutls on PixArt-Σ. Comparison between OBS-Diff and our importance￾aware pruning guided by CFG [2] signals at 20%, 30%, and 40% structured sparsity. Prompt: "An Egyptian airlines plane landing at an airport." 38 [PITH_FULL_IMAGE:figures/full_fig_p038_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: Different importance signals. Comparison between OBS-Diff and our importance-aware variants guided by CFG [2], CFG+ object-detection[16], and Canny-edge[15] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A black Honda motorcycle parked in front of a garage." 39 [PITH_FULL_IMAGE:figures/full_fig_p039_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: Different importance signals. Comparison between OBS-Diff and our importance-aware variants guided by CFG [2], CFG+ object-detection[16], and Canny-edge[15] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A large plant is in the corner of a small bathroom." 40 [PITH_FULL_IMAGE:figures/full_fig_p040_38.png] view at source ↗
Figure 39
Figure 39. Figure 39: Different importance signals. Comparison between OBS-Diff and our importance-aware variants guided by CFG [2], CFG+ object-detection[16], and Canny-edge[15] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A dirt bike parked near a tent in the woods." 41 [PITH_FULL_IMAGE:figures/full_fig_p041_39.png] view at source ↗
Figure 40
Figure 40. Figure 40: Different importance signals. Comparison between OBS-Diff and our importance-aware variants guided by CFG [2], CFG+ object-detection[16], and Canny-edge[15] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A gray and white kitten in a white bathroom sink." 42 [PITH_FULL_IMAGE:figures/full_fig_p042_40.png] view at source ↗
Figure 41
Figure 41. Figure 41: Different importance signals. Comparison between OBS-Diff and our importance-aware variants guided by CFG [2], CFG+ object-detection[16], and Canny-edge[15] signals at 45%, 50%, 55%, and 60% unstructured sparsity. Prompt: "A large truck is parked beside an old, wooden bench." 43 [PITH_FULL_IMAGE:figures/full_fig_p043_41.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

114 extracted references · 43 linked inside Pith

  1. [1]

    OBS-diff: Accurate pruning for diffusion models in one-shot,

    J. Zhu, H. Wang, M. Su, Z. Wang, and H. Wang, “OBS-diff: Accurate pruning for diffusion models in one-shot,” inThe Fourteenth International Conference on Learning Representations,

  2. [2]

    Classifier-free diffusion guidance,

    J. Ho and T. Salimans, “Classifier-free diffusion guidance,”arXiv preprint arXiv:2207.12598, 2022

    Pith/arXiv arXiv 2022

  3. [3]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  4. [5]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boeselet al., “Scaling rectified flow transformers for high-resolution image synthesis,” in Forty-first international conference on machine learning, 2024

    2024

  5. [6]

    Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding,

    Z. Li, J. Zhang, Q. Lin, J. Xiong, Y . Long, X. Deng, Y . Zhang, X. Liu, M. Huang, Z. Xiaoet al., “Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding,”arXiv preprint arXiv:2405.08748, 2024

    Pith/arXiv arXiv 2024

  6. [7]

    Efficient scaling of diffusion transformers for text-to-image generation,

    H. Li, S. Lal, Z. Li, Y . Xie, Y . Wang, Y . Zou, O. Majumder, R. Manmatha, Z. Tu, S. Ermon et al., “Efficient scaling of diffusion transformers for text-to-image generation,”arXiv preprint arXiv:2412.12391, 2024

    Pith/arXiv arXiv 2024

  7. [8]

    Exploring the deep fusion of large language models and diffusion transformers for text-to-image synthesis,

    B. Tang, B. Zheng, S. Paul, and S. Xie, “Exploring the deep fusion of large language models and diffusion transformers for text-to-image synthesis,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 28 586–28 595

    2025

  8. [9]

    Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer,

    E. Xie, J. Chen, Y . Zhao, J. Yu, L. Zhu, C. Wu, Y . Lin, Z. Zhang, M. Li, J. Chenet al., “Sana 1.5: Efficient scaling of training-time and inference-time compute in linear diffusion transformer,”arXiv preprint arXiv:2501.18427, 2025

    Pith/arXiv arXiv 2025

  9. [10]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,

    B. F. Labs, S. Batifol, A. Blattmann, F. Boesel, S. Consul, C. Diagne, T. Dockhorn, J. English, Z. English, P. Esser, S. Kulal, K. Lacey, Y . Levi, C. Li, D. Lorenz, J. Müller, D. Podell, R. Rombach, H. Saini, A. Sauer, and L. Smith, “Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,” 2025. [Online]. Available: htt...

    Pith/arXiv arXiv 2025

  10. [11]

    Second order derivatives for network pruning: Optimal brain surgeon,

    B. Hassibi and D. G. Stork, “Second order derivatives for network pruning: Optimal brain surgeon,” inNeural Information Processing Systems, 1992. [Online]. Available: https://api.semanticscholar.org/CorpusID:7057040

    1992

  11. [12]

    Learning to weight parameters for training data attribution,

    S. Li, H. Le, J. Xu, and M. Salzmann, “Learning to weight parameters for training data attribution,” inThe Fourteenth International Conference on Learning Representations, 2026. [Online]. Available: https://openreview.net/forum?id=EhUkQp9Yah

    2026

  12. [13]

    A model of saliency-based visual attention for rapid scene analysis,

    L. Itti, C. Koch, and E. Niebur, “A model of saliency-based visual attention for rapid scene analysis,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, pp. 1254–1259, 1998. [Online]. Available: https://api.semanticscholar.org/CorpusID:3108956

    1998

  13. [14]

    Few-shot personalized scanpath prediction,

    R. Xue, J. Xu, S. Mondal, H. Le, G. J. Zelinsky, M. Hoai, and D. Samaras, “Few-shot personalized scanpath prediction,”2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 13 497–13 507, 2025. [Online]. Available: https://api.semanticscholar.org/CorpusID:277627903 10

    2025

  14. [15]

    A computational approach to edge detection,

    J. F. Canny, “A computational approach to edge detection,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, pp. 679–698, 1986. [Online]. Available: https://api.semanticscholar.org/CorpusID:13284142

    1986

  15. [16]

    Yolov8: A novel object detection algorithm with enhanced perfor- mance and robustness,

    R. Varghese and S. M., “Yolov8: A novel object detection algorithm with enhanced perfor- mance and robustness,” in2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS), 2024, pp. 1–6

    2024

  16. [17]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” inInternational conference on machine learning. PMLR, 2015, pp. 2256–2265

    2015

  17. [18]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  18. [19]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,”arXiv preprint arXiv:2010.02502, 2020

    Pith/arXiv arXiv 2010

  19. [20]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 10 684–10 695

    2022

  20. [21]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopeset al., “Photorealistic text-to-image diffusion models with deep language understanding,”arXiv preprint arXiv:2205.11487, 2022

    Pith/arXiv arXiv 2022

  21. [22]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4195–4205

    2023

  22. [23]

    Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis,

    J. Chen, J. Yu, C. Ge, L. Yao, E. Xie, Y . Wu, Z. Wang, J. Kwok, P. Luo, H. Lu, and Z. Li, “Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis,” 2023

    2023

  23. [24]

    Hierarchical text-conditional image generation with clip latents,

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with clip latents,”ArXiv, vol. abs/2204.06125, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:248097655

    Pith/arXiv arXiv 2022

  24. [25]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models,

    A. Nichol, P. Dhariwal, A. Ramesh, P. Shyam, P. Mishkin, B. McGrew, I. Sutskever, and M. Chen, “Glide: Towards photorealistic image generation and editing with text-guided diffusion models,”arXiv preprint arXiv:2112.10741, 2021

    Pith/arXiv arXiv 2021

  25. [26]

    Learned representation-guided diffusion models for large-image generation,

    A. Graikos, S. Yellapragada, M.-Q. Le, S. Kapse, P. Prasanna, J. Saltz, and D. Samaras, “Learned representation-guided diffusion models for large-image generation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 8532– 8542

    2024

  26. [27]

    Sdedit: Guided image synthesis and editing with stochastic differential equations,

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.-Y . Zhu, and S. Ermon, “Sdedit: Guided image synthesis and editing with stochastic differential equations,”arXiv preprint arXiv:2108.01073, 2021

    Pith/arXiv arXiv 2021

  27. [28]

    Raphael: Text-to-image generation via large mixture of diffusion paths,

    Z. Xue, G. Song, Q. Guo, B. Liu, Z. Zong, Y . Liu, and P. Luo, “Raphael: Text-to-image generation via large mixture of diffusion paths,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  28. [29]

    ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers,

    Y . Balaji, S. Nah, X. Huang, A. Vahdat, J. Song, Q. Zhang, K. Kreis, M. Aittala, T. Aila, S. Laineet al., “ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers,” arXiv preprint arXiv:2211.01324, 2022

    Pith/arXiv arXiv 2022

  29. [30]

    Guiding a diffusion model with a bad version of itself,

    T. Karras, M. Aittala, T. Kynkäänniemi, J. Lehtinen, T. Aila, and S. Laine, “Guiding a diffusion model with a bad version of itself,”Advances in Neural Information Processing Systems, vol. 37, pp. 52 996–53 021, 2025

    2025

  30. [31]

    Make-a-video: Text-to-video generation without text-video data,

    U. Singer, A. Polyak, T. Hayes, X. Yin, J. An, S. Zhang, Q. Hu, H. Yang, O. Ashual, O. Gafni et al., “Make-a-video: Text-to-video generation without text-video data,”arXiv preprint arXiv:2209.14792, 2022. 11

    Pith/arXiv arXiv 2022

  31. [32]

    Stable video diffusion: Scaling latent video diffusion models to large datasets,

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y . Levi, Z. English, V . V oleti, A. Lettset al., “Stable video diffusion: Scaling latent video diffusion models to large datasets,”arXiv preprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  32. [33]

    Animatediff: Animate your personalized text-to-image diffusion models without specific tuning,

    Y . Guo, C. Yang, A. Rao, Z. Liang, Y . Wang, Y . Qiao, M. Agrawala, D. Lin, and B. Dai, “Animatediff: Animate your personalized text-to-image diffusion models without specific tuning,”arXiv preprint arXiv:2307.04725, 2023

    Pith/arXiv arXiv 2023

  33. [34]

    Open-sora: Democratizing efficient video production for all,

    Z. Zheng, X. Peng, T. Yang, C. Shen, S. Li, H. Liu, Y . Zhou, T. Li, and Y . You, “Open-sora: Democratizing efficient video production for all,” March 2024. [Online]. Available: https://github.com/hpcaitech/Open-Sora

    2024

  34. [35]

    Open-sora-plan,

    P.-Y . Lab and T. A. etc., “Open-sora-plan,” Apr. 2024. [Online]. Available: https: //doi.org/10.5281/zenodo.10948109

    doi:10.5281/zenodo.10948109 2024

  35. [36]

    Video generation models as world simulators. 2024,

    T. Brooks, B. Peebles, C. Holmes, W. DePue, Y . Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhmanet al., “Video generation models as world simulators. 2024,”URL https://openai. com/research/video-generation-models-as-world-simulators, vol. 3, 2024

    2024

  36. [37]

    Dreamfusion: Text-to-3d using 2d diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “Dreamfusion: Text-to-3d using 2d diffusion,” arXiv preprint arXiv:2209.14988, 2022

    Pith/arXiv arXiv 2022

  37. [38]

    Zero-1-to-3: Zero-shot one image to 3d object,

    R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, and C. V ondrick, “Zero-1-to-3: Zero-shot one image to 3d object,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 9298–9309

    2023

  38. [39]

    Wonder3d: Single image to 3d using cross-domain diffusion,

    X. Long, Y .-C. Guo, C. Lin, Y . Liu, Z. Dou, L. Liu, Y . Ma, S.-H. Zhang, M. Habermann, C. Theobaltet al., “Wonder3d: Single image to 3d using cross-domain diffusion,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 9970–9980

    2024

  39. [40]

    Zero123++: a single image to consistent multi-view diffusion base model,

    R. Shi, H. Chen, Z. Zhang, M. Liu, C. Xu, X. Wei, L. Chen, C. Zeng, and H. Su, “Zero123++: a single image to consistent multi-view diffusion base model,”arXiv preprint arXiv:2310.15110, 2023

    Pith/arXiv arXiv 2023

  40. [41]

    3ddesigner: Towards photore- alistic 3d object generation and editing with text-guided diffusion models,

    G. Li, H. Zheng, C. Wang, C. Li, C. Zheng, and D. Tao, “3ddesigner: Towards photore- alistic 3d object generation and editing with text-guided diffusion models,”arXiv preprint arXiv:2211.14108, 2022

    Pith/arXiv arXiv 2022

  41. [42]

    Diffusion probabilistic models for 3d point cloud generation,

    S. Luo and W. Hu, “Diffusion probabilistic models for 3d point cloud generation,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 2837–2845

    2021

  42. [43]

    Point-e: A system for generating 3d point clouds from complex prompts,

    A. Nichol, H. Jun, P. Dhariwal, P. Mishkin, and M. Chen, “Point-e: A system for generating 3d point clouds from complex prompts,”arXiv preprint arXiv:2212.08751, 2022

    Pith/arXiv arXiv 2022

  43. [44]

    Diffwave: A versatile diffusion model for audio synthesis,

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Diffwave: A versatile diffusion model for audio synthesis,”arXiv preprint arXiv:2009.09761, 2020

    Pith/arXiv arXiv 2009

  44. [45]

    Make- an-audio: Text-to-audio generation with prompt-enhanced diffusion models,

    R. Huang, J. Huang, D. Yang, Y . Ren, L. Liu, M. Li, Z. Ye, J. Liu, X. Yin, and Z. Zhao, “Make- an-audio: Text-to-audio generation with prompt-enhanced diffusion models,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 13 916–13 932

    2023

  45. [46]

    Magicfusion: Boosting text-to-image gener- ation performance by fusing diffusion models,

    J. Zhao, H. Zheng, C. Wang, L. Lan, and W. Yang, “Magicfusion: Boosting text-to-image gener- ation performance by fusing diffusion models,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 22 592–22 602

    2023

  46. [47]

    High-fidelity person-centric subject-to-image synthesis,

    Y . Wang, W. Zhang, J. Zheng, and C. Jin, “High-fidelity person-centric subject-to-image synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 7675–7684

    2024

  47. [48]

    Würstchen: An efficient architecture for large-scale text-to-image diffusion models,

    P. Pernias, D. Rampas, M. L. Richter, C. J. Pal, and M. Aubreville, “Würstchen: An efficient architecture for large-scale text-to-image diffusion models,”arXiv preprint arXiv:2306.00637, 2023. 12

    Pith/arXiv arXiv 2023

  48. [49]

    On architectural compression of text-to- image diffusion models,

    B.-K. Kim, H.-K. Song, T. Castells, and S. Choi, “On architectural compression of text-to- image diffusion models,” 2023

    2023

  49. [50]

    Structural pruning for diffusion models,

    G. Fang, X. Ma, and X. Wang, “Structural pruning for diffusion models,” inAdvances in Neural Information Processing Systems, 2023

    2023

  50. [51]

    Mobilediffusion: Subsecond text-to-image generation on mobile devices,

    Y . Zhao, Y . Xu, Z. Xiao, and T. Hou, “Mobilediffusion: Subsecond text-to-image generation on mobile devices,”arXiv preprint arXiv:2311.16567, 2023

    Pith/arXiv arXiv 2023

  51. [52]

    Diffusion probabilistic model made slim,

    X. Yang, D. Zhou, J. Feng, and X. Wang, “Diffusion probabilistic model made slim,” in Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, 2023, pp. 22 552–22 562

    2023

  52. [53]

    Snapfusion: Text-to-image diffusion model on mobile devices within two seconds,

    Y . Li, H. Wang, Q. Jin, J. Hu, P. Chemerys, Y . Fu, Y . Wang, S. Tulyakov, and J. Ren, “Snapfusion: Text-to-image diffusion model on mobile devices within two seconds,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  53. [54]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,”arXiv preprint arXiv:2202.00512, 2022

    Pith/arXiv arXiv 2022

  54. [55]

    On distillation of guided diffusion models,

    C. Meng, R. Rombach, R. Gao, D. Kingma, S. Ermon, J. Ho, and T. Salimans, “On distillation of guided diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 297–14 306

    2023

  55. [56]

    Adversarial diffusion distillation,

    A. Sauer, D. Lorenz, A. Blattmann, and R. Rombach, “Adversarial diffusion distillation,” in European Conference on Computer Vision. Springer, 2025, pp. 87–103

    2025

  56. [57]

    Instaflow: One step is enough for high-quality diffusion-based text-to-image generation,

    X. Liu, X. Zhang, J. Ma, J. Penget al., “Instaflow: One step is enough for high-quality diffusion-based text-to-image generation,” inThe Twelfth International Conference on Learn- ing Representations, 2023

    2023

  57. [58]

    Clockwork diffusion: Efficient generation with model-step distillation,

    A. Habibian, A. Ghodrati, N. Fathima, G. Sautiere, R. Garrepalli, F. Porikli, and J. Petersen, “Clockwork diffusion: Efficient generation with model-step distillation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 8352–8361

    2024

  58. [59]

    Consistency models,

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever, “Consistency models,”arXiv preprint arXiv:2303.01469, 2023

    Pith/arXiv arXiv 2023

  59. [60]

    Latent consistency models: Synthesizing high-resolution images with few-step inference,

    S. Luo, Y . Tan, L. Huang, J. Li, and H. Zhao, “Latent consistency models: Synthesizing high-resolution images with few-step inference,”arXiv preprint arXiv:2310.04378, 2023

    Pith/arXiv arXiv 2023

  60. [61]

    Pseudo numerical methods for diffusion models on manifolds,

    L. Liu, Y . Ren, Z. Lin, and Z. Zhao, “Pseudo numerical methods for diffusion models on manifolds,”arXiv preprint arXiv:2202.09778, 2022

    Pith/arXiv arXiv 2022

  61. [62]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps,”Advances in Neural Information Processing Systems, vol. 35, pp. 5775–5787, 2022

    2022

  62. [63]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models,

    ——, “Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models,” arXiv preprint arXiv:2211.01095, 2022

    Pith/arXiv arXiv 2022

  63. [64]

    Deepcache: Accelerating diffusion models for free,

    X. Ma, G. Fang, and X. Wang, “Deepcache: Accelerating diffusion models for free,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 762–15 772

    2024

  64. [65]

    Cache me if you can: Accelerating diffusion models through block caching,

    F. Wimbauer, B. Wu, E. Schoenfeld, X. Dai, J. Hou, Z. He, A. Sanakoyeu, P. Zhang, S. Tsai, J. Kohleret al., “Cache me if you can: Accelerating diffusion models through block caching,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 6211–6220

    2024

  65. [66]

    Delta- dit: A training-free acceleration method tailored for diffusion transformers,

    P. Chen, M. Shen, P. Ye, J. Cao, C. Tu, C.-S. Bouganis, Y . Zhao, and T. Chen, “Delta- dit: A training-free acceleration method tailored for diffusion transformers,”arXiv preprint arXiv:2406.01125, 2024. 13

    Pith/arXiv arXiv 2024

  66. [67]

    Real-time video generation with pyramid attention broadcast,

    X. Zhao, X. Jin, K. Wang, and Y . You, “Real-time video generation with pyramid attention broadcast,”arXiv preprint arXiv:2408.12588, 2024

    Pith/arXiv arXiv 2024

  67. [68]

    Adaptive caching for faster video generation with diffusion transformers,

    K. Kahatapitiya, H. Liu, S. He, D. Liu, M. Jia, M. S. Ryoo, and T. Xie, “Adaptive caching for faster video generation with diffusion transformers,”arXiv preprint arXiv:2411.02397, 2024

    Pith/arXiv arXiv 2024

  68. [69]

    Frdiff: Feature reuse for universal training-free acceleration of diffusion models,

    J. So, J. Lee, and E. Park, “Frdiff: Feature reuse for universal training-free acceleration of diffusion models,”arXiv preprint arXiv:2312.03517, 2023

    Pith/arXiv arXiv 2023

  69. [70]

    Fastercache: Training-free video diffusion model acceleration with high quality,

    Z. Lv, C. Si, J. Song, Z. Yang, Y . Qiao, Z. Liu, and K.-Y . K. Wong, “Fastercache: Training-free video diffusion model acceleration with high quality,”arXiv preprint arXiv:2410.19355, 2024

    Pith/arXiv arXiv 2024

  70. [71]

    Faster diffusion: Rethinking the role of unet encoder in diffusion models,

    S. Li, T. Hu, F. S. Khan, L. Li, S. Yang, Y . Wang, M.-M. Cheng, and J. Yang, “Faster diffusion: Rethinking the role of unet encoder in diffusion models,”CoRR, 2023

    2023

  71. [72]

    Q-diffusion: Quantizing diffusion models,

    X. Li, Y . Liu, L. Lian, H. Yang, Z. Dong, D. Kang, S. Zhang, and K. Keutzer, “Q-diffusion: Quantizing diffusion models,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 17 535–17 545

    2023

  72. [73]

    Q-dit: Accurate post-training quantization for diffusion transformers,

    L. Chen, Y . Meng, C. Tang, X. Ma, J. Jiang, X. Wang, Z. Wang, and W. Zhu, “Q-dit: Accurate post-training quantization for diffusion transformers,”arXiv preprint arXiv:2406.17343, 2024

    Pith/arXiv arXiv 2024

  73. [74]

    Ptqd: Accurate post-training quantization for diffusion models,

    Y . He, L. Liu, J. Liu, W. Wu, H. Zhou, and B. Zhuang, “Ptqd: Accurate post-training quantization for diffusion models,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  74. [75]

    Quest: Low-bit diffusion model quantization via efficient selective finetuning,

    H. Wang, Y . Shang, Z. Yuan, J. Wu, J. Yan, and Y . Yan, “Quest: Low-bit diffusion model quantization via efficient selective finetuning,”arXiv preprint arXiv:2402.03666, 2024

    Pith/arXiv arXiv 2024

  75. [76]

    Vq4dit: Efficient post-training vector quantization for diffusion transformers,

    J. Deng, S. Li, Z. Wang, H. Gu, K. Xu, and K. Huang, “Vq4dit: Efficient post-training vector quantization for diffusion transformers,”arXiv preprint arXiv:2408.17131, 2024

    Pith/arXiv arXiv 2024

  76. [77]

    Temporal dynamic quantization for diffusion models,

    J. So, J. Lee, D. Ahn, H. Kim, and E. Park, “Temporal dynamic quantization for diffusion models,”Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  77. [78]

    Object- centric diffusion for efficient video editing,

    K. Kahatapitiya, A. Karjauv, D. Abati, F. Porikli, Y . M. Asano, and A. Habibian, “Object- centric diffusion for efficient video editing,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 91–108

    2025

  78. [79]

    Token merging for fast stable diffusion,

    D. Bolya and J. Hoffman, “Token merging for fast stable diffusion,”2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 4599–4603, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:257833518

    2023

  79. [80]

    Token merging: Your vit but faster,

    D. Bolya, C.-Y . Fu, X. Dai, P. Zhang, C. Feichtenhofer, and J. Hoffman, “Token merging: Your vit but faster,”ArXiv, vol. abs/2210.09461, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:252968113

    Pith/arXiv arXiv 2022

  80. [81]

    Token fusion: Bridging the gap between token pruning and token merging,

    M. Kim, S. Gao, Y .-C. Hsu, Y . Shen, and H. Jin, “Token fusion: Bridging the gap between token pruning and token merging,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 1383–1392

    2024

Showing first 80 references.