Research map / Image generation / Fast diffusion sampling
SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Scales the linear diffusion transformer SANA from 1.6B to 4.8B via depth-growth initialization, prunes blocks by input-output similarity, uses an 8-bit CAME optimizer, and boosts GenEval to 0.96 with VLM-judged inference-time sampling.
Problem
Scaling text-to-image diffusion transformers is costly in training and inference; the paper asks how to scale a linear DiT efficiently and reduce training cost while keeping quality under varied compute budgets.
Approach
Grow a pre-trained linear DiT by adding identity-initialized blocks instead of training from scratch, prune low-importance blocks using input-output similarity with brief fine-tuning, and use a fine-tuned VLM (VILA-Judge) in a tournament to select the best of many samples at inference.
Key results
- Model growth raises SANA 4.8B pre-trained GenEval from 0.66 (1.6B) to 0.72, and post-training on curated SFT data reaches 0.81, with ~60% fewer training steps than training from scratch.
- Pruning the 4.8B model to 1.6B and fine-tuning for 100 steps gives GenEval 0.672, above SANA-1.0 1.6B at 0.664.
- Inference scaling with 2048 samples and VILA-Judge lifts SANA-1.5 4.8B GenEval from 0.81 to 0.96, exceeding Playground v3 (0.76).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Scale SANA-1.0 1.6B to SANA-1.5 4.8B Pre using model growth | SANA-1.0 1.6B · GenEval | GenEval: 0.66 to 0.72 | Helped |
| Use model growth instead of training the 4.8B model from scratch | SANA-1.5 4.8B · GenEval training curve | training steps to reach 0.70 GenEval: 60% fewer training steps | Helped |
| Post-train on high-quality data | SANA-1.5 4.8B Pre · GenEval | GenEval: 0.72 to 0.81 | Helped |
| Post-train on high-quality data | SANA-1.5 4.8B Pre · MJHQ-30K | FID: 5.42 -> 5.99 | Hurt |
| Generate 2048 candidates and select with VILA-Judge | SANA-1.5 4.8B Ours · GenEval | Overall: 0.81 to 0.96 | Helped |
| Prune the 4.8B model to 3.2B without fine-tuning (ablation) | SANA-1.5 4.8B · GenEval | GenEval: 0.693 -> 0.673 | Hurt |
| Fine-tune the 3.2B pruned model (ablation) | SANA-1.5 3.2B pruned · GenEval | GenEval: 0.673 -> 0.684 | Helped |
| Prune the 4.8B model to 1.6B without fine-tuning (ablation) | SANA-1.5 4.8B · GenEval | GenEval: 0.693 -> 0.571 | Hurt |
| Fine-tune the 1.6B pruned model (ablation) | SANA-1.5 1.6B pruned · GenEval | GenEval: 0.571 -> 0.672 | Helped |
| Prune to 1.6B and fine-tune instead of using SANA-1.0 1.6B (ablation) | SANA-1.5 4.8B · GenEval | GenEval: 0.664 to 0.672 | Helped |
| Use CAME-8bit instead of AdamW | SANA-1.6B training · SANA-1.6B training | GPU memory: 57GB to 43GB | Helped |
| Use CAME-8bit instead of AdamW | SANA-1.6B training · SANA-1.6B training | training convergence speed | No clear effect |
Surprising findings
- Naively appending new blocks after all pre-trained blocks led to stuck local minima; removing the last two pre-trained blocks was needed for learning.
- GPT-4o and Gemini-1.5-pro showed inconsistent ratings and strong bias toward first-presented images when used as judges.
Limitations
- Inference scaling with VLM judging adds large compute cost (N x 49,140G FLOPs for generation plus 2N x 4,518G FLOPs for judging).
- Struggles with text rendering and human details, per the authors.
- Pruned models lose high-frequency detail and rely on fine-tuning to recover quality.
Experimental setup
Datasets: MJHQ-30K GenEval prompts-style synthetic data CLIP-filtered image-text pairs
Benchmarks: GenEval DPG-Bench MJHQ-30K
Metrics: GenEval overall DPG score FID CLIP score throughput latency
Baselines: SANA-1.0 1.6B PixArt-Sigma SDXL SD3-medium FLUX-dev FLUX-schnell Playground v3 Hunyuan-DiT DALLE 3 Playground v2.5 LUMINA-Next
Builds on: SANA CAME AdamW-8bit Minitron NVILA
Closest papers by meaning
- SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers (ICLR 2025 · similarity 0.91)
- On the Scalability of Diffusion-based Text-to-Image Generation (CVPR 2024 · similarity 0.91)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.90)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.89)
- Towards Precise Scaling Laws for Video Diffusion Transformers (CVPR 2025 · similarity 0.89)
- Stretching Each Dollar: Diffusion Training from Scratch on a Micro-Budget (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.