acceptodds
Under review as a conference paper at ICLR 2027

TILT: MODEL-INTRINSIC REWARD ALIGNMENT FOR COMPOSITIONAL DIFFUSION

Abstract

Consider conditional generation where is a prompt composed of multiple concepts . Diffusion models often struggle with compositional prompts, producing samples in which some concepts dominate while others are missing or weakly represented. Prior work attributes these failures to *mode collision*, where single-concept modes of overlap with modes of the joint . To seek out collision-free modes of , or *pure modes*, corrector-based approaches have attempted to suppress collisions at intermediate diffusion times. However, local corrections are often heuristic and do not necessarily steer the generation to a *pure mode* in the final data space. Derived from a principled formulation, we present TILT (**T**est-time model-**I**ntrinsic reward a**L**ignment via **T**ilting), a training-free framework that poses eventual pure mode sampling as a reward for intermediate-time alignment. This reward offers valuable advantages: (1) it is intrinsic to the model, hence external reward models need not be trained by modality-specific datasets, (2) it yields a closed-form target under a variational approximation, which makes it realizable through standard diffusion sampling, and (3) it is interpretable, hence amenable to preference-based modifications. Project page: [https://tinyurl.com/yr5j3bwe](https://tinyurl.com/yr5j3bwe)

Primary area:
generative models
Keywords:
Text-to-image generation, compositional generation, diffusion model, reward alignment, diffusion steering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_cUVbOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies concept dominance in compositional diffusion generation, which it attributes to mode overlap between the joint distribution p(x∣C)p(x\mid C) and the single-concept distributions p(x∣ci)p(x\mid c_i). The authors argue that corrector-based methods such as CO3 take this ratio at intermediate timesteps, so they do not actually target the pure-mode distribution defined at t=0t=0.

The proposed method, TILT, instead defines pure-mode sampling as a reward at t=0t=0, Rw(x0)=log⁡p(x0∣C)/∏ip(x0∣ci)wiR_w(x_0)=\log p(x_0\mid C)/\prod_i p(x_0\mid c_i)^{w_i}, and solves a KL-regularized reward alignment problem. The reward is estimated from diffusion classifier energies, so no external reward model is needed. Sampling uses DPS-style guidance on the Tweedie estimate x^0\hat x_0, applied only in the early denoising steps.

Experiments on SDXL (GenEval, T2I-CompBench) and AudioLDM (AudioCaps) show gains over CFG and prior compositional methods. Plugged into FKS or SCORE in place of CLIP or CLAP, the reward is competitive with these external rewards.

Critical strengths
  1. The paper identifies a real problem. Corrector-based methods such as CO3 take the ratio of noised distributions at each intermediate step, while the mode overlap they aim to fix actually lives at t=0t=0. Noising and then dividing is not the same as dividing and then noising, so these methods do not actually sample from the pure-mode distribution they target. The observation is simple, but I think it is correct, and it is the most valuable part of the paper.

  2. The reward and the sampling derivation are clear. The authors first define the target in data space, then use the KL-regularized formulation and posterior sampling to derive guidance at intermediate steps. Because the reward is a likelihood ratio, the intractable terms cancel under the ELBO approximation, and no external reward model is needed. The argument is easy to follow, and it is clear where each approximation enters.

Critical weaknesses
  1. The step from the target distribution to the actual algorithm relies on two approximations, and I don't think their assumptions hold here. Eq. 17 follows DPS and replaces the posterior expectation in p(O∣xt)p(O\mid x_t) with a point estimate at x^0\hat x_0. This works reasonably in DPS when the likelihood is smooth and the posterior is concentrated, but here the likelihood is exp⁡(βR)\exp(\beta R) with RR built from UNet denoising errors, and there is no reason to expect it to be smooth. Eq. 21 then follows MPGD and drops the Tweedie Jacobian, updating x^0\hat x_0 directly. MPGD's argument depends on the gradient lying in the tangent space of the data manifold, and even then it only says samples stay on the manifold, not that they follow the target distribution. The gradient here comes from a UNet-based energy, and the authors never check whether it satisfies this. What makes it worse is that guidance is only applied in the first 5 steps, at high noise, where x^0\hat x_0 is basically a blurry mean image and both approximations are at their weakest. The central argument of the paper is that CO3 fails to sample from the pure-mode distribution while TILT does it properly. After these approximations, though, it is just as unclear what distribution TILT ends up sampling from, so I don't find this contrast convincing.

  2. The image experiments use only SDXL, which is fairly old at this point, and a lot of its compositional failures come from its CLIP text encoders. Newer models like FLUX or Qwen-Image use much stronger text encoders and handle composition noticeably better, so I'm not sure the problem is as severe there, or that TILT would still help. These models are also mostly flow-matching based, and some, such as FLUX.1-dev, are guidance-distilled, so it's not obvious the ϵ\epsilon-based energy carries over directly. Without results on at least one recent model, it's hard to judge how much the method matters in practice today.

Questions for authors
  1. What happens if you keep the Jacobian, i.e., use J⊤∇x^0RJ^\top\nabla_{\hat x_0}R, or precondition with some approximation of the posterior covariance, instead of the direct update in Eq. 21? Since computing ∇x^0R\nabla_{\hat x_0}R already requires backpropagating through the UNet, one extra VJP shouldn't add much cost. If the results are similar, the approximation is probably harmless. If they differ noticeably, the gap between the derivation and what is actually run is larger than the paper suggests. (Experiment request)

  2. Why is guidance applied only in the first 5 steps? Did you try guiding at later steps, and what happened? Is this choice motivated by the derivation, or is it purely empirical? (Clarification)

  3. Does the final algorithm actually approximate p∗p^*? A low-dimensional example like the one in Fig. 1, where p∗p^* can be computed exactly, would make this easy to check: compare samples from TILT and from the CO3 corrector against the true p∗p^*. I think this is the most direct evidence for the paper's main claim that TILT targets the right distribution while correctors do not. (Experiment request)

  4. When computing ∇x^0Ec\nabla_{\hat x_0}E_c, do you backpropagate through the UNet? If not, and you use an SDS-style approximation instead, the update reduces to roughly ∑iwi(ϵθ(xτ∣C)−ϵθ(xτ∣ci))\sum_i w_i\big(\epsilon_\theta(x_\tau\mid C)-\epsilon_\theta(x_\tau\mid c_i)\big), which is essentially guidance with the sub-concepts used as negative prompts. Could you discuss how TILT relates to this kind of negative-prompt guidance, and ideally add a simple baseline that uses the sub-concepts as negative prompts directly? That would show whether TILT gives anything beyond this simpler approach. (Clarification and experiment request)

Overall, I like the idea behind this paper, and my current rating of weak reject mainly reflects the practical issues above rather than doubts about the approach itself. If the authors can clearly address the questions and the weaknesses I raised, in particular the two approximations and the lack of results on more recent models, and the results support the method, I would be happy to raise my score to weak accept or higher.

6 comments

W1/Q3: The step from the target distribution to the actual algorithm relies on two approximations, and their assumptions don't hold here. Does the final algorithm actually approximate p∗p^{\ast}? Comparison of samples from CO3 and TILT on a low-dimensional toy example would make this easy to check.

It's true that the Tweedie and the Jacobian-free updates introduce approximations in the proposed algorithm. But we also want to highlight the following: while DPS's likelihood is smooth, it has a data consistency term which is much stricter than the soft likelihood based reward used in TILT. TILT only needs to encourage the intermediate latents to move towards higher likelihood regions under the target distribution, which is much weaker than DPS's data consistency term, thus less sensitive towards approximation in Tweedie.

As for the Jacobian-free update, we observe empirical gains as presented in the new ablations in the tables below. The with-Jacobian and without-Jacobian updates approaches are comparable, so dropping it saves some compute cost. We will include this ablation in the revised version of the paper.

 

Table: TILT with and without Jacobian update on GenEval prompts.

MethodT2V ↑\uparrowHPSv3 ↑\uparrow
CFG0.3934 ± 0.0195.5029 ± 0.10
TILT(w/o Jac)0.5192 ± 0.0186.1219 ± 0.09
TILT(w/ Jac)0.5057 ± 0.0145.9021 ± 0.076

 

Table: TILT with and without Jacobian update on T2I-CompBench prompts.

MethodColor T2V ↑\uparrowColor HPSv3 ↑\uparrowShape T2V ↑\uparrowShape HPSv3 ↑\uparrowTexture T2V ↑\uparrowTexture HPSv3 ↑\uparrowComplex T2V ↑\uparrowComplex HPSv3 ↑\uparrow
CFG0.5183 ± 0.0155.1216 ± 0.100.3565 ± 0.0126.1377 ± 0.090.5529 ± 0.0146.5021 ± 0.080.6238 ± 0.0157.5702 ± 0.10
TILT (w/o Jac)0.5688 ± 0.0155.8002 ± 0.100.4019 ± 0.0137.0523 ± 0.080.5932 ± 0.0147.2136 ± 0.080.6350 ± 0.0158.2156 ± 0.10
TILT (w/ Jac)0.5626 ± 0.0145.5672 ± 0.1060.3994 ± 0.0137.0262 ± 0.0780.5800 ± 0.0147.2039 ± 0.070.6351 ± 0.0158.2160 ± 0.096

 

The toy experiment is a remarkable idea; thank you. We setup a 2-D mixtute of 6 Gaussians to model p(x∣C)p(x | C), p(x∣c1)p(x | c_1), and p(x∣c2)p(x | c_2). Each distribution is of the form p(x∣c)∝∑kπkP(c∣k)N(x;μk,σk2I)p(x\mid c)\propto\sum_k\pi_k P(c\mid k) \mathcal N(x;\mu_k,\sigma_k^2 I) and C=C= "c1c_1 and c2c_2" collides with each single concept. In this case, all scores can be computed analytically. We use CFG++ as a base sampler and compare TILT against CO3. Each method draws 8000 samples per seed over 10 seeds. We draw exact samples of pβ∗∝qeβRwp^{\ast}_\beta \propto q e^{\beta R_w} by importance resampling 400400k CFG++ samples. We then calculate the KL divergence DKL(p^∥pβ∗)D_{\mathrm{KL}}(\hat p\Vert p^{\ast}_\beta) between the estimated distribution p^\hat p and the target pβ∗p^{\ast}_\beta.

The table below reports the results: for both β=1\beta=1 and β=2\beta=2, TILT achieves the lowest KL divergence, while maitaining a competitive expected reward with CO3. Higher KL divergence and higher reward for CO3 indicate that CO3 sacrifices diversity by over-optimizing the reward (which is a form of reward hacking). This imbalance is likely to get far more pronounced in higher dimensional spaces. TILT better balances reward optimization with target distribution-fidelity compared to CO3. We are adding these additional figures, implementation details, and results in the Appendix (Sec A.7.1, page 26).

 

Table: Distance to the exact target and expected reward attained by the samplers.

β\betaMethodKL(p^∥pβ∗)↓\mathrm{KL}(\hat p \Vert p^{\ast}_\beta)\downarrowE[Rw]\mathbb{E}[R_w]
11Target pβ∗(β=1)p^{\ast}_\beta (\beta=1)
CFG++ (base)0.163±0.0100.163 \pm 0.0101.7481.748
CO30.405±0.0130.405 \pm 0.0132.0842.084
TILT (ours)0.036±0.0030.036 \pm 0.0031.9511.951
22Target pβ∗(β=2)p^{\ast}_\beta(\beta=2)
CFG++ (base)0.468±0.0200.468 \pm 0.0201.7481.748
CO30.372±0.0130.372 \pm 0.0132.0842.084
TILT (ours)0.167±0.0090.167 \pm 0.0092.0912.091

   

W2: The image experiments use only SDXL. it's not obvious the ϵ\epsilon-based energy carries over to rectified flow & guidance distilled models directly.

In any Rectified flow model,

xt=(1−t)x0+tϵ,  v=ϵ−x0 and x0=xt−tvx_t = (1-t)x_0 + t\epsilon, \ \ v = \epsilon - x_0 \text{ and } x_0 = x_t - tv

we can represent the noise as ϵ=xt+(1−t)v\epsilon = x_t + (1-t)v. Since this relation is linear, we get ∣∣ϵ−ϵθ∣∣=(1−t)∣∣v−vθ∣∣|| \epsilon - \epsilon_\theta|| = (1-t)||v-v_\theta||. So the current diffusion energy expression in TILT can be converted to a velocity-based energy and an equivalent energy based reward can be used for Rectified-flow-based models.

In this short time, we rapidly implemented a small-scale comparison to see the ballpark results. What we have is not conclusive and significant. Nonetheless, we see TILT outperforms the base model in both metrics. We feel there is room to extract out more performance. So if the paper gets accepted, we will make an earnest effort to build out on these preliminary experiments (including additional comparisons with baselines on more complex prompts).

 

Table: Quantitative results on GenEval prompts using SD3 as the backbone.

MethodT2V ↑\uparrowHPSv3 ↑\uparrow
CFG0.778 ± 0.037.21 ± 0.121
TILT0.793 ± 0.027.51 ± 0.12

   

Q1: What happens if you keep the Jacobian?

Keeping the Jacobian gives comparable results on GenEval and T2I-CompBench (see the tables in our response to W1/Q3 above): the Jacobian-free formulation remains competitive or slightly better, indicating that the approximation does not degrade performance, while preserving the inference speed-up by avoiding the costly JVP operation for extra gradient propagation.

   

Q2: Why is guidance applied only in the first 5 steps?

This is an empirical observation from our evaluation results. The table below shows TILT's performance over different correction-steps and performance drops beyond 55 across most categories. We observe that gradient from the reward becomes stronger and sharper as diffusion denoising moves toward the clean sample, which leads to over-correction. We suspect this stems from the increasingly ill-conditioned behavior of the denoiser gradient at lower time steps, but we are not sure. We adopt an early stopping mechanism to preserve image quality.

 

Table: TILT's performance over different correction steps.

# Correction StepsColor T2V ↑\uparrowColor HPSv3 ↑\uparrowShape T2V ↑\uparrowShape HPSv3 ↑\uparrowTexture T2V ↑\uparrowTexture HPSv3 ↑\uparrowComplex T2V ↑\uparrowComplex HPSv3 ↑\uparrow
0 (CFG)0.5183 ± 0.0155.1216 ± 0.100.3565 ± 0.0126.1377 ± 0.090.5529 ± 0.0146.5021 ± 0.080.6238 ± 0.0157.5702 ± 0.10
50.5688 ± 0.0155.8002 ± 0.100.4019 ± 0.0137.0523 ± 0.080.5932 ± 0.0147.2136 ± 0.080.6350 ± 0.0158.2156 ± 0.10
100.5598 ± 0.0145.7102 ± 0.1020.3834 ± 0.0116.9086 ± 0.070.5861 ± 0.0147.1448 ± 0.0790.6368 ± 0.0158.0945 ± 0.10

   

Reviewer_zWt9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is addressing the concept dominance phenomenon in compositional diffusion models, where one of the conditional concepts in a multi-concept prompt is dominating in the final outcome. It is building on top of the CO3, which is steering the diffusion trajectory to the samples favoring the entire prompt rather than any single concept in it. This paper is arguing that CO3 is applying these corrections at noisy intermediate steps, and they instead apply the same guidance into a reward on the final clean image by framing the sampling as a KL-regularized reward maximization problem. They estimate the reward from model's own denoising, where the idea is comparing the full prompt to each of the condition concepts. Then they apply this to the predicted clean image by gradient corrections. There are also multiple experiments where TILT has slightly better performance compared to baselines.

Critical strengths
  • Estimating the reward objective from the generator model's own denoising errors is a good idea. It doesn't need any external reward model and uses the score model itself.
  • It could be used as a reward for the other methods as well, where paper showed that it can replace the CLIP in FKS. Paper also showed the benefits on the audio similarly. It empirically proves that this reward has good signal.
Critical weaknesses
  • The method combines several previous works, therefore there is a limited novelty. The target reward is coming from the CO3's correction moved to final clean image space. The likelihood ratio trick is coming from the diffusion classifier energies. And the update mechanism of removing the Jacobians is coming from the MPGD (which is the jacobian free version of DPS).
  • Also the guarantee does not consider the sampling choices. In practice, the reward is calculated by the blurry mean image; and also the jacobian is approximated by identity.
  • The improvements over the baselines are also very modest/small. Moreover, it would be nice to see the MPGD/DPS samplings with other rewards than this reward. For the audio, the CLAP baseline is beating the TILT.
  • There is also no variance analysis over multiple seeds to see if the reward always gives a good signal or has a high variance.

I believe the novelty of the work is limited given that it combines previously proposed reward model into a sampling/steering method, which was also established before.

Questions for authors
  • What does TILT-S/C mean in the Appendix A3?
  • Did you try this current sampling (MPGD) with some other learnt rewards?
  • Eq.20 and Algorithm 1 (line 4) has sign mismatch?
  • Can you also report/analyze the runtime and memory consumptions against baselines?
  • How does TILT perform on recent flow based models e.g. SD3, FLUX?
6 comments

W1: The method combines several previous works, therefore there is a limited novelty.

It is difficult to argue objectively here, but allow us to emphasize the following:

  1. Indeed, the target reward was originally hypothesized by CO3, but there was little grounding for the hypothesis. We provided grounding through empirical measurements and the connection to Total Correlation and diffusion classifier. We think this paper gives much more credence to what was simply a hypothesis.

  2. More importantly, CO3's heuristic was incorrect (as shown in Fig. 1) and TILT recognizes that the idea can be reformulated/solved correctly, by (1) first formulating "pure mode sampling" as an intrinsic reward, and (2) by deriving the correct distribution ∀t>0\forall t >0 that optimizes the intrinsic reward in the data-space (t=0t=0). We believe this is the core contribution where novelty resides. Yes, we borrow likelihood ratio tricks and Jacobian-removal from MPGD, but those are objects needed to realize the sampling from the correctly-designed distribution.

W2: the guarantee does not consider the sampling choices. The reward is calculated by the blurry mean image; and also the jacobian is approximated by identity.

It's true that the Tweedie and the Jacobian-free updates introduce approximations that the guarantee does not account for. But we also want to highlight the following: the Tweedie approximation already works well in DPS, whose data consistency term is much stricter than the soft likelihood based reward used in TILT. TILT only needs to encourage the intermediate latents to move towards higher likelihood regions under the target distribution, thus it is even less sensitive towards the approximation in Tweedie.

As for the Jacobian-free update, it comes from empirical observation. We show the with-Jacobian and without-Jacobian results in the tables below. The two updates are comparable, so dropping the Jacobian saves some compute cost.

 

Table: TILT with and without Jacobian update on GenEval prompts.

MethodT2V ↑\uparrowHPSv3 ↑\uparrow
CFG0.3934 ± 0.0195.5029 ± 0.10
TILT(w/o Jac)0.5192 ± 0.0186.1219 ± 0.09
TILT(w/ Jac)0.5057 ± 0.0145.9021 ± 0.076

 

Table: TILT with and without Jacobian update on T2I-CompBench prompts.

MethodColor T2V ↑\uparrowColor HPSv3 ↑\uparrowShape T2V ↑\uparrowShape HPSv3 ↑\uparrowTexture T2V ↑\uparrowTexture HPSv3 ↑\uparrowComplex T2V ↑\uparrowComplex HPSv3 ↑\uparrow
CFG0.5183 ± 0.0155.1216 ± 0.100.3565 ± 0.0126.1377 ± 0.090.5529 ± 0.0146.5021 ± 0.080.6238 ± 0.0157.5702 ± 0.10
TILT (w/o Jac)0.5688 ± 0.0155.8002 ± 0.100.4019 ± 0.0137.0523 ± 0.080.5932 ± 0.0147.2136 ± 0.080.6350 ± 0.0158.2156 ± 0.10
TILT (w/ Jac)0.5626 ± 0.0145.5672 ± 0.1060.3994 ± 0.0137.0262 ± 0.0780.5800 ± 0.0147.2039 ± 0.070.6351 ± 0.0158.2160 ± 0.096

 

To validate that TILT still approximates the target under these approximations, we ran a toy experiment where everything can be calculated exactly. We setup a 2-D mixture of 6 Gaussians to model p(x∣C)p(x | C), p(x∣c1)p(x | c_1), and p(x∣c2)p(x | c_2). Each distribution is of the form p(x∣c)∝∑kπkP(c∣k)N(x;μk,σk2I)p(x\mid c)\propto\sum_k\pi_k P(c\mid k) \mathcal N(x;\mu_k,\sigma_k^2 I) and C=C= "c1c_1 and c2c_2" collides with each single concept. In this case, all scores can be computed analytically. We use CFG++ as a base sampler and compare TILT against CO3. Each method draws 8000 samples per seed over 10 seeds. We draw exact samples of pβ∗∝qeβRwp^{\ast}_\beta \propto q e^{\beta R_w} by importance resampling 400400k CFG++ samples. We then calculate the KL divergence DKL(p^∥pβ∗)D_{\mathrm{KL}}(\hat p\Vert p^{\ast}_\beta) between the estimated distribution p^\hat p and the target pβ∗p^{\ast}_\beta.

The table below reports the results: for both β=1\beta=1 and β=2\beta=2, TILT achieves the lowest KL divergence, while maintaining a competitive expected reward with CO3. Higher KL divergence and higher reward for CO3 indicate that CO3 sacrifices diversity by over-optimizing the reward. Thus, even under the approximations, TILT approximates the target well while maintaining stronger fidelity than CO3. We have added more figures and results from this experiment to Sec A.7.1 in Appendix (page 26) of the main paper.

 

Table: Distance to the exact target and expected reward.

β\betaMethodKL(p^∥pβ∗)↓\mathrm{KL}(\hat p \Vert p^{\ast}_\beta)\downarrowE[Rw]\mathbb{E}[R_w]
11Target pβ∗(β=1)p^{\ast}_\beta (\beta=1)
CFG++ (base)0.163±0.0100.163 \pm 0.0101.7481.748
CO30.405±0.0130.405 \pm 0.0132.0842.084
TILT (ours)0.036±0.0030.036 \pm 0.0031.9511.951
22Target pβ∗(β=2)p^{\ast}_\beta(\beta=2)
CFG++ (base)0.468±0.0200.468 \pm 0.0201.7481.748
CO30.372±0.0130.372 \pm 0.0132.0842.084
TILT (ours)0.167±0.0090.167 \pm 0.0092.0912.091

   

W3: There is also no variance analysis over multiple seeds

We have now added SE (standard error) along with the means to the tables below. We generate 3 images per prompt and the mean and SE are calculated over the whole dataset on GenEval and T2I-CompBench. We will add the SE values to all the tables in the revised paper.

 

Table: Quantitative results on GenEval prompts.

MethodT2V ↑\uparrowHPSv3 ↑\uparrow
CFG0.3934 ± 0.0195.5029 ± 0.10
CFG++0.5073 ± 0.0185.9806 ± 0.10
R2F0.4799 ± 0.0185.8176 ± 0.10
SuperDiff0.1589 ± 0.0110.0820 ± 0.17
CO30.5014 ± 0.0184.3026 ± 0.11
DOS0.5025 ± 0.0185.5744 ± 0.11
TILT0.5192 ± 0.0186.1219 ± 0.09

 

Table: Quantitative results on T2I-CompBench prompts.

MethodColor T2V ↑\uparrowColor HPSv3 ↑\uparrowShape T2V ↑\uparrowShape HPSv3 ↑\uparrowTexture T2V ↑\uparrowTexture HPSv3 ↑\uparrowComplex T2V ↑\uparrowComplex HPSv3 ↑\uparrow
CFG0.5183 ± 0.0155.1216 ± 0.100.3565 ± 0.0126.1377 ± 0.090.5529 ± 0.0146.5021 ± 0.080.6238 ± 0.0157.5702 ± 0.10
CFG++0.5668 ± 0.0155.7358 ± 0.110.3839 ± 0.0136.9679 ± 0.080.5839 ± 0.0147.1200 ± 0.090.6307 ± 0.0158.1785 ± 0.11
R2F0.5328 ± 0.0155.5146 ± 0.100.3452 ± 0.0126.5545 ± 0.080.5573 ± 0.0146.7431 ± 0.080.6234 ± 0.0158.0638 ± 0.10
CO30.5938 ± 0.0144.4736 ± 0.100.3895 ± 0.0136.1982 ± 0.080.4252 ± 0.0156.2801 ± 0.080.6313 ± 0.0156.6193 ± 0.11
DOS0.5028 ± 0.0154.8979 ± 0.110.3865 ± 0.0126.0388 ± 0.090.6199 ± 0.0146.4046 ± 0.090.6603 ± 0.0156.4046 ± 0.10
SuperDiff0.2420 ± 0.012-0.2098 ± 0.170.1660 ± 0.0092.6294 ± 0.150.2528 ± 0.0110.4163 ± 0.170.3865 ± 0.0152.3976 ± 0.22
TILT (Ours)0.5688 ± 0.0155.8002 ± 0.100.4019 ± 0.0137.0523 ± 0.080.5932 ± 0.0147.2136 ± 0.080.6350 ± 0.0158.2156 ± 0.10

   

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

The paper proposes a training-free diffusion steering method that defines a pure-mode target in the data space with an intrinsic reward from the joint-versus-individual concept likelihood ratio. The reward uses diffusion denoising energies to guide sampling. The method is clearly an extension of the CO3 work from last year, which already introduced the core idea of contrasting the joint prompt distribution with individual-concept distributions. The main new contribution is the formulation of the objective and more principled reward-based steering.

Critical strengths
  • Principled formulation of compositional steering rather than heuristic correction.
  • Stronger than CO3 on most reported image-generation metrics.
  • Demonstrates the same idea across image and audio generation.
Critical weaknesses
  • The overlap with CO3 should be more clearly described, for instance the core joint-vs-individual concept contrast is already formulated there.
  • The Jacobian-free correction is computationally convenient but introduces another approximation.
  • Inference becomes more expensive because each reward evaluation requires extra diffusion-model passes and gradient computation.
Questions for authors
  • How much performance is affected by the Jacobian-free correction compared with using the full gradient?
  • Could you give a more systematic evaluation of the inference-time compute overhead relative to CO3 and standard CFG?
5 comments

W1: The overlap with CO3 should be clearly described -- core joint vs individual concept contrast is already formulated there.

We do not intend to play down CO3. In the revised version, we will generously attribute credit to CO3 and clearly demarcate with TILT as follows:

  1. TILT shows that CO3's corrector distribution is a heuristic because reducing overlap in latent-space does not result in minimal mode-overlap in data-space (which was CO3's goal). TILT designs the correct distribution ∀t>0\forall t > 0 and shows how it is practical to sample from this distribution. This is correct by design and results reflect this improvement.

  2. CO3 proposed an interesting failure hypothesis. TILT gives credence to this hypothesis by first showing empirical evidence in Fig. 1, and then deriving that "pure mode" sampling is equivalent to maximizing the Total correlation between concepts. The paper also establishes a connection between the "pure mode" reward and diffusion classifier. This theoretical grounding was absent in CO3.

W2/Q2: Jacobian-free formulation introduces another approximation. How much performance is affected by the Jacobian-free approach?

This is a pertinent question and we have now compared TILT with and without the Jacobian on the GenEval and T2I-CompBench. Results are still promising: the Jacobian-free formulation remains competitive or slightly better, indicating that the approximation does not degrade performance, while preserving the inference speed-up by avoiding the costly JVP operation for extra gradient propagation.

 

Table: TILT with and without Jacobian update on GenEval prompts.

MethodT2V ↑\uparrowHPSv3 ↑\uparrow
CFG0.3934 ± 0.0195.5029 ± 0.10
TILT(w/o Jac)0.5192 ± 0.0186.1219 ± 0.09
TILT(w/ Jac)0.5057 ± 0.0145.9021 ± 0.076

 

Table: TILT with and without Jacobian update on T2I-CompBench prompts.

MethodColor T2V ↑\uparrowColor HPSv3 ↑\uparrowShape T2V ↑\uparrowShape HPSv3 ↑\uparrowTexture T2V ↑\uparrowTexture HPSv3 ↑\uparrowComplex T2V ↑\uparrowComplex HPSv3 ↑\uparrow
CFG0.5183 ± 0.0155.1216 ± 0.100.3565 ± 0.0126.1377 ± 0.090.5529 ± 0.0146.5021 ± 0.080.6238 ± 0.0157.5702 ± 0.10
TILT (w/o Jac)0.5688 ± 0.0155.8002 ± 0.100.4019 ± 0.0137.0523 ± 0.080.5932 ± 0.0147.2136 ± 0.080.6350 ± 0.0158.2156 ± 0.10
TILT (w/ Jac)0.5626 ± 0.0145.5672 ± 0.1060.3994 ± 0.0137.0262 ± 0.0780.5800 ± 0.0147.2039 ± 0.070.6351 ± 0.0158.2160 ± 0.096

   

W3/Q3: Inference becomes expensive. Give systematic evaluation of inference time compute overhead relative to CO3 and standard CFG.

We ran new experiments and find that all baselines are expensive compared to CFG (with DDIM) except for DOS. SuperDiff is the most expensive baseline, requiring 147.07 seconds for inference. While TILT takes slightly higher memory than CO3, it is relatively faster, as CO3 takes many more correction steps to achieve comparable performance.

 

Table: Inference time compute comparison of TILT with baselines. The numbers are computed using an NVIDIA A6000 Pro Blackwell GPU on GenEval prompts.

MethodTime (s)Memory (GB)
CFG4.678.5
R2F3.9814.76
SuperDiff147.0711.95
CO316.1212.85
DOS4.2114.92
TILT (Ours)14.2314.71

   

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.