Research map / Video / Video generation
Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion
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.
Ouroboros-Diffusion improves FIFO-Diffusion for tuning-free long video generation by adding low-frequency tail latent sampling, subject-aware cross-frame attention, and gradient-based self-recurrent guidance from the queue head.
Problem
Tuning-free long video generation with a pre-trained text-to-video model suffers from long-range temporal inconsistency (subject/background drift and flickering) in FIFO/diagonal denoising, due to independent Gaussian tail noise and lack of cross-frame correspondence.
Approach
Builds the enqueued tail latent from the low-frequency component of the second-to-last latent plus high-frequency Gaussian noise, adds subject-masked cross-frame attention using subject tokens from a CLIP-encoded prompt, and guides tail denoising with a moving-average subject memory from the queue head via gradient descent.
Key results
- Single-scene 128-frame generation: Ouroboros-Diffusion reaches 96.06 subject consistency and 96.12 temporal flickering, vs. 94.04 and 93.38 for FIFO-Diffusion.
- Multi-scene 256-frame generation: subject consistency 95.73 vs. 95.07 for FreeNoise (best baseline) and motion smoothness 97.77 vs. 96.57.
- Ablation: coherent tail latent sampling alone raises subject consistency from 94.04 to 95.56 and motion smoothness from 95.88 to 97.61.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace FIFO-Diffusion's tail Gaussian noise with coherent tail latent sampling (ablation) | FIFO-Diffusion / Ouroboros-Diffusion · VBench single-scene | Subject Consistency: 94.04 -> 95.56 | Helped |
| add SACFA after coherent tail latent sampling (ablation) | Ouroboros-Diffusion variant B · VBench single-scene | Subject Consistency: 95.56 -> 95.71 | Helped |
| add self-recurrent guidance after coherent tail latent sampling and SACFA (ablation) | Ouroboros-Diffusion variant C · VBench single-scene | Subject Consistency: 95.71 -> 96.06 | Helped |
| use second-to-last frame rather than independent Gaussian noise for tail sampling (ablation) | Ouroboros-Diffusion · VBench single-scene | Motion Smoothness: 95.88 -> 97.73 | Helped |
| use head frame rather than second-to-last frame for tail latent sampling (ablation) | Ouroboros-Diffusion · VBench single-scene | Motion Smoothness: 95.88 -> 97.51 | Helped |
| use moving-average subject feature-bank updates rather than λ=1 historic-only guidance (ablation) | Ouroboros-Diffusion · VBench single-scene | Subject Consistency: 95.87 -> 96.06 | Helped |
| use moving-average subject feature-bank updates rather than λ=0 current-only guidance (ablation) | Ouroboros-Diffusion · VBench single-scene | Subject Consistency: 96.00 -> 96.06 | Helped |
| apply Ouroboros-Diffusion rather than FreeNoise | VideoCrafter2 long-video generation · VBench multi-scene | Aesthetic Quality: 61.26 -> 61.17 | Hurt |
| apply Ouroboros-Diffusion rather than FIFO-Diffusion | VideoCrafter2 long-video generation · VBench single-scene | Temporal Flickering: 93.38 -> 96.12 | Helped |
| apply Ouroboros-Diffusion rather than FreeNoise | VideoCrafter2 long-video generation · VBench multi-scene | Subject Consistency: 95.07 -> 95.73 | Helped |
Surprising findings
- Using the head frame's low-frequency component for tail sampling yields lower gains than the second-to-last frame, and simply re-noising the second-to-last latent produced limited dynamics.
- Aesthetic quality is slightly lower than FreeNoise in multi-scene generation (61.17 vs. 61.26), attributed to the mismatch between diagonal and parallel denoising noise levels.
Limitations
- Aesthetic quality slightly trails FreeNoise in multi-scene generation.
- Discrepancy between parallel and diagonal denoising noise levels may limit visual quality; authors suggest training could address it.
Experimental setup
Datasets: VBench prompts (93 single-scene prompts) GPT-4o-extended multi-scene prompt groups (78 groups)
Benchmarks: VBench
Metrics: Subject Consistency Background Consistency Motion Smoothness Temporal Flickering Aesthetic Quality Dynamic Degree
Baselines: StreamingT2V StreamingT2V-VideoTetris FIFO-Diffusion FreeNoise
Builds on: FIFO-Diffusion VideoCrafter2 FreeDoM ConsiStory
Closest papers by meaning
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.91)
- FIFO-Diffusion: Generating Infinite Videos from Text without Training (NeurIPS 2024 · similarity 0.91)
- FreeNoise: Tuning-Free Longer Video Diffusion via Noise Rescheduling (ICLR 2024 · similarity 0.90)
- Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.89)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.89)
- AR-Diffusion: Asynchronous Video Generation with Auto-Regressive Diffusion (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.