SALD: Self-Referenced Advantage Learning for Diffusion Models

Aryan Das; Surjo Dey; Koushik Biswas; Swalpa Kumar Roy; Moloud Abdar; Arnab Bhattacharya; Vinay Kumar Verma

cs.CVOct 1, 2026

SALD: Self-Referenced Advantage Learning for Diffusion Models

Authors: Aryan Das, Surjo Dey, Koushik Biswas, Swalpa Kumar Roy, Moloud Abdar, Arnab Bhattacharya, Vinay Kumar Verma

Organizations: Indian Institute of Technology Kanpur, India · Rajiv Gandhi Institute of Petroleum Technology, India · IIIT Delhi, India · Tezpur University, India · The University of Queensland, Australia

Abstract

Recent work on language-model adaptation has shown that single models can obtain informative training signals by evaluating their behavior in demonstrationor feedback-augmented contexts, with the help of a teacher network, which is driven by the student's learned parameters. Inspired by this internal-reference principle, we investigate how diffusion models can identify self-referenced training signals without external demonstrations or teacher networks. We introduce SALD, a self-referenced training framework that evaluates each image-caption pair at two noise levels using the same model. The easier, lower-noise path is evaluated without gradient tracking to provide a reference, while the harder, higher-noise path provides the training gradient. Rather than directly distilling the easy-path prediction, SALD uses the difference between two path errors to adapt the hardpath objective. The proposed Advantage-Guided Diffusion (AGD) converts this relative error into a differentiable sample-level weight. Temporal Advantage Memory (TAM) accumulates relative difficulty across training and adapts the future gap between the two noise levels. Spectral Advantage Decomposition (SAD) further compares the residual power spectra of the two paths and constructs a differentiable, frequency-derived latent-element weight. All components share a single set of model parameters, requiring neither an external teacher network nor additional trainable parameters during training or inference, and no modification to the inference procedure. Experiments across multiple architectures and datasets demonstrate consistent improvements in generation quality, while component-wise ablations quantify the contributions of the proposed components.

Figures & tables

Figure 1: Images generated by SALD using the SANA 0.6B model at a resolution of $1024\times 1024$ .

Figure 2: SALD framework. A frozen VAE encodes the input latent, processed through two paths: Low-Noise Path A (w/o gradient) and High-Noise Path B (with gradient). AGD computes an advantage for sample reweighting, while TAM stabilizes it via per-sample EMA, forming a persistent curriculum across timestep gaps. SAD compares FFT power spectra of velocity residual maps, applies a sigmoid gate, and reconstructs a spatial weight map via IFFT to focus gradients on high-frequency error regions. Together, these define $\mathcal{L}_{\text{SALD}}$ for adaptive gradient allocation across samples.

Model	Method	COCO (2017)			MultiGen-20M			Flickr8k			CUB-200			Oxford 102 Flowers
Model	Method	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$
SANA (1.6B)	Base	26.49	26.44	0.7411	26.44	26.25	0.7541	27.22	28.49	0.7336	21.63	26.01	0.7748	33.11	26.35	0.7896
	SpeeD	25.38	26.83	0.7409	26.11	26.26	0.7558	26.87	28.52	0.7324	21.53	26.28	0.7753	32.04	26.51	0.7782
	Temporal Diff	26.36	26.87	0.7405	26.31	26.29	0.7539	26.85	28.85	0.7325	20.72	26.25	0.7766	33.10	26.54	0.7801
	SRA	26.89	26.32	0.7461	26.05	25.89	0.7554	26.87	28.78	0.7341	20.78	26.18	0.7722	32.75	26.40	0.7781
	SALD	22.76	26.88	0.7391	23.86	26.29	0.7533	26.75	28.88	0.7316	18.16	26.29	0.7537	27.37	26.57	0.7781

Table 1: Quantitative results on (

512\times 512

) resolution datasets. Best values per metric are in Blue .

Model	Method	MAGICK			DALL $\cdot$ E 3
Model	Method	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$
SANA 0.6B	Base	38.59	26.68	0.7470	14.68	29.96	0.7882
	SpeeD	38.27	26.65	0.7303	14.18	30.05	0.7878
	Temporal Diff	37.95	26.69	0.7294	13.84	30.38	0.7854
	SRA	38.15	26.70	0.7297	13.96	30.73	0.7839
	SALD	36.74	26.85	0.7230	11.98	31.06	0.7826

Table 2: Quantitative results on

1024\!\times\!1024

resolution images. Best values per metric within each model are in Blue .

Figure 5: Visual comparison of generated samples. Red boxes indicate missing details or artifacts in latest baseline models. Green boxes highlight that our method generates better high-quality details. We use the same seed for all comparisons to ensure fairness. Captions are in Blue Color .

Cases	Setting	COCO ( $512$ )			MAGICK ( $1024$ )
Cases	Setting	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$
Case 1	AGD only	30.95	26.30	0.7462	37.75	26.73	0.7354
	SAD only	31.02	26.49	0.7427	37.86	26.70	0.7314
	TAM only	30.98	26.50	0.7474	37.35	26.74	0.7323
Case 2	AGD + TAM	30.78	26.60	0.7529	37.34	26.72	0.7396
	AGD + SAD	30.56	26.66	0.7375	37.78	26.79	0.7400

Table 3: Ablation study of SALD on SANA 0.6B across COCO 2017 (

512

) and MAGICK (

1024

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Params (B)	FLOPs (G)	Resolution	LR	Guidance	Infer. Steps
SANA (0.6B)	0.6	548	$512\times 512$	1e-4	5	20
SANA (1.6B)	1.6	567	$512\times 512$	1e-4	5	20
PixArt- $\Sigma$	0.6	567	$512\times 512$	2e-5	5	20
SANA (0.6B) †	0.6	1397	$1024\times 1024$	1e-4	5	20
SANA (1.6B) †	1.6	2240	$1024\times 1024$	1e-4	5	20
PixArt- $\Sigma^{\dagger}$	0.6	1399	$1024\times 1024$	2e-5	5	20

Appendix

Table 4: Detailed model configurations used in our experiments, including parameter count, per-forward-pass FLOPs, training resolution, learning rate, guidance scale, and inference steps. † denotes models additionally evaluated at

1024\!\times\!1024

resolution.

Figure 6: Qualitative comparison across baseline models and SALD on 1024 × 1024 datasets.

Figure 7: Qualitative comparisons: samples generated by SALD and other different Baselines on 1024 × 1024 Datasets. Captions are, Column-1: A cute orange cat sleeping curled up, flat vector illustration style. Column-2: An hourglass with golden sand, dramatic cinematic lighting on a wooden table. Column-3: A woman’s face in profile with flowing blue hair made of water, digital art.

Figure 8: Qualitative comparison of inference-time generation across different sampling steps for SALD and the baseline using SANA 1.6B trained on DALL $\cdot$ E 3 1M.

Dataset	Model	Variant	Params (B)	Steps	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$
CUB-200	SANA	Base	1.6	12.5k	21.47	26.01	0.7709
		SpeeD	1.6	12.5k	20.83	26.08	0.7670
		SRA	1.6	12.5k	20.72	26.15	0.7632
		Temporal Diff	1.6	12.5k	20.61	26.22	0.7593
		SALD	1.6	10k	18.16	26.29	0.7537
	SANA	Base	0.6	12.5k	25.54	26.19	0.7802

Appendix

Table 5: Wall-clock time and quantitative comparison on CUB-200 and MultiGen datasets. Best values in Blue

Model	Params (B)	Variant	Steps = 10			Steps = 15			Steps = 20
			FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	CLIP $\uparrow$	LPIPS $\downarrow$
SANA	1.6	Base	55.06	25.35	0.8153	37.49	25.60	0.8055	33.11	26.35	0.7896
		SALD	41.56	25.48	0.8088	30.60	25.72	0.7962	27.37	26.57	0.7781
SANA	0.6	Base	44.05	24.31	0.8059	35.26	25.29	0.7971	30.64	26.43	0.7805
		SALD	33.68	24.48	0.7906	25.54	25.63	0.7865	25.83	26.65	0.7728
PixArt- $\Sigma$	0.6	Base	49.12	25.10	0.8120	43.25	25.25	0.7910	37.88	25.37	0.7687

Appendix

Table 6: Quantitative comparison across different inference steps on the Oxford Flowers dataset. Light blue highlights indicate the metric values for the SALD variant.

Steps	Variant	MultiGen						Oxford 102 Flowers
		SANA 1.6			SANA 0.6			SANA 1.6			SANA 0.6
		FID	CLIP	LPIPS	FID	CLIP	LPIPS	FID	CLIP	LPIPS	FID	CLIP	LPIPS
10k	Baseline	26.44	26.25	0.7541	31.93	26.68	0.7495	33.11	26.35	0.7896	30.64	26.43	0.7805
10k	SALD	23.86	26.29	0.7533	29.15	26.69	0.7472	27.37	26.57	0.7781	25.83	26.65	0.7728
20k	Baseline	26.18	26.31	0.7550	31.42	26.48	0.7605	32.45	26.30	0.7842	30.41	26.38	0.7761
20k	SALD	23.52	26.35	0.7542	28.78	26.52	0.7598	26.80	26.52	0.7745	25.70	26.60	0.7695

Appendix

Table 7: Evaluation metrics across training steps for SANA 1.6 and SANA 0.6 on MultiGen and Oxford 102 Flowers datasets. Lower FID and LPIPS (

\downarrow

) are better, while higher CLIP (

\uparrow

) is better.

Figure 9: Qualitative samples generated by SALD using PixArt- $\Sigma$ at $1024\!\times\!1024$ resolution on the DALL $\cdot$ E 3 1M dataset.

Figure 10: Qualitative samples generated by SALD using Flux.2 klein Base at $1024\!\times\!1024$ resolution on the DALL $\cdot$ E 3 1M dataset.

Figure 11: Qualitative samples generated by SALD across multiple datasets and architectures. Row 1: SANA 1.6B at $1024\!\times\!1024$ resolution trained on DALL $\cdot$ E 3 1M. Row 2: SANA 1.6B generations on CUB-200. Row 3: PixArt- $\Sigma$ generations on Oxford-102 Flowers.

Explore similar work

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models
May 28, 2026Xiaohang Tang, Keyue Jiang, Che Liu +4Diffusion Language Models Large Language Model Reinforcement Learning
UNIST · University of Basel
Forward-Free Diffusion Language Models
Jun 6, 2026Haotian Sun, Rushi Qiang, Yuqian Zheng +1Diffusion Language Models Diffusion Models
Georgia Institute of Technology
Representation-Space MMD for Diffusion Language Models
Oct 5, 2026Ilya Drobyshevskiy, Ilia Sudakov, Maksim Semenov +7Diffusion Language Models
Yandex Research · HSE University · T-Tech +3

May 28, 2026, cs.LGJ/K move · Enter open · S save

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models

Xiaohang Tang, Keyue Jiang, Che Liu, Qifang Zhao+3

UNIST · University of Basel

Diffusion Language Models Large Language Model Reinforcement Learning Unsupervised On-Policy Self-Distillation Denoising Diffusion Probabilistic Model Lower Bounds

Open paper AskPDF arXiv

Reinforcement learning (RL) can be used to improve the policy (denoiser) of diffusion large language models (dLLMs), while being hindered by the intractability of the policy likelihood. A dominant and efficient family of methods replaces the likelihood in standard RL with its evidence lower bound (ELBO), estimated from randomly masked sequences. Despite being well aligned with pre-training, these approaches introduce bias through training--inference mismatch by using the ELBO as a likelihood surrogate, which can degrade performance. In this work, we propose Guided Denoiser Self-Distillation (GDSD) to directly distill the denoiser of dLLMs from an advantage-guided self-teacher, derived from the closed-form optimum of reverse-KL regularized RL. GDSD matches the dLLM's denoiser logits to the teacher's via a normalization-free objective, which reduces RL to likelihood-free self-distillation and thus bypasses the TIM biases. Recent ELBO-based methods emerge as instances of applying different distillation divergences, but with diagnosable pathologies that GDSD avoids. On planning, math, and coding benchmarks with LLaDA-8B and Dream-7B, GDSD consistently outperforms prior state-of-the-art ELBO-based methods with a more stable training reward dynamics, achieving test-accuracy improvements of up to

+19.6\%

. These results suggest that direct denoiser self-distillation, without relying on an ELBO likelihood surrogate, can provide a more stable and effective RL procedure for dLLMs. Code is available at https://github.com/GaryBall/GDSD.

Abstract

Explore similar work

GDSD: Reinforcement Learning as Guided Denoiser Self-Distillation for Diffusion Language Models

Forward-Free Diffusion Language Models

Representation-Space MMD for Diffusion Language Models