Research map / Image generation / Fast diffusion sampling
Diffusion Adversarial Post-Training for One-Step Video Generation
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.
Proposes Adversarial Post-Training (APT), which turns a pretrained 8B video/image DiT into a one-step generator by adversarial training on real data with architectural discriminator changes and an approximated R1 penalty.
Problem
Iterative diffusion sampling is slow and expensive, and existing one-step distillation methods lose quality, especially for high-resolution video.
Approach
Skip the diffusion teacher: continue training a consistency-distilled DiT adversarially against real data, using a discriminator initialized from the diffusion weights with multi-layer cross-attention logit heads, timestep-ensemble inputs, and an approximated R1 loss based on perturbed real samples that avoids double backward.
Key results
- One-step APT is preferred +37.2% over the 25-step base diffusion model for visual fidelity in image evaluation, but is weaker on structural integrity (-13.1%) and text alignment (-8.1%).
- The one-step model generates 1280x720 24fps 2-second videos; one H100 takes 6.03s for the full pipeline, and 8 H100s reach 1.97s (real-time).
- On VBench, one-step APT scores 82.00 total vs 82.15 for the 25-step diffusion teacher and 67.05 for 1-step consistency distillation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use adversarial post-training on real data instead of one-step diffusion distillation | APT image generator vs its 25-step diffusion model · Image generation | visual fidelity preference: +37.2% | Helped |
| Use adversarial post-training on real data instead of one-step diffusion distillation | APT image generator vs its 25-step diffusion model · Image generation | structural integrity preference: -13.1% | Hurt |
| Use adversarial post-training on real data instead of one-step diffusion distillation | APT image generator vs its 25-step diffusion model · Image generation | text alignment preference: -8.1% | Hurt |
| Use adversarial post-training for one-step generation | APT video generator vs 25-step diffusion · Video generation, 1 step | visual fidelity preference: +10.4% | Helped |
| Use adversarial post-training for one-step generation | APT video generator vs 25-step diffusion · Video generation, 1 step | structural integrity preference: -38.5% | Hurt |
| Use adversarial post-training for one-step generation | APT video generator vs 25-step diffusion · Video generation, 1 step | text alignment preference: -8.3% | Hurt |
| Use APT rather than consistency distillation | Video generation model · VBench | total score: 67.05 -> 82.00 | Helped |
| Add approximated R1 regularization (ablation) | APT discriminator · Adversarial training stability | training collapse | Helped |
| Add discriminator depth using the full pre-trained network (ablation) | APT discriminator · Image generation | image quality | Helped |
| Use multilayer discriminator features (ablation) | APT discriminator · Image generation | structural integrity | Helped |
| Train the entire discriminator network rather than freezing its backbone (ablation) | APT discriminator · Image generation | image quality | Helped |
| Use a large training batch instead of batch size 256 (ablation) | APT video model · Video generation | mode collapse | Helped |
Surprising findings
- Automatic metrics (FID, PFID, CLIP) disagree with human evaluation; e.g., FLUX 1-step beats FLUX 25-step on all metrics, contrary to human judgments.
- The one-step APT model beats its 25-step diffusion base on visual fidelity, mainly by removing CFG-induced over-exposure and over-saturation.
Limitations
- Videos limited to two seconds due to compute constraints.
- One-step generation degrades structural integrity and text alignment relative to 25-step diffusion, more so for video due to motion.
- Automatic metrics are unreliable proxies for quality, so evaluation relies mainly on user studies.
Experimental setup
Datasets: proprietary image and video training data (re-captioned) PartiPrompt DrawBench COCO 2014 validation VBench prompts
Benchmarks: PartiPrompt DrawBench COCO5K COCO10K VBench
Metrics: user preference score (visual fidelity, structural integrity FID PFID CLIP score VBench total/quality/semantic score
Baselines: FLUX-Schnell SD3.5-Turbo SDXL-DMD2 SDXL-Hyper SDXL-Lightning SDXL-Nitro-Realism SDXL-Turbo Consistency distillation (1/2/4 NFE)
Builds on: consistency distillation DiT MMDiT DMD2 UFO-Gen
Closest papers by meaning
- SF-V: Single Forward Video Generation Model (NeurIPS 2024 · similarity 0.93)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.92)
- Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation (NeurIPS 2025 · similarity 0.91)
- Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (ICML 2025 · similarity 0.89)
- Bridging Data Gaps in Diffusion Models with Adversarial Noise-Based Transfer Learning (ICML 2024 · similarity 0.88)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.87)
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.