From 681a3dacf12e0c8a171df7b758a8003026785c37 Mon Sep 17 00:00:00 2001 From: nicolaus Date: Wed, 19 Mar 2025 18:39:52 +0800 Subject: [PATCH] add inference scaling doc --- README.md | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/README.md b/README.md index 29a5755..3e7bd4a 100644 --- a/README.md +++ b/README.md @@ -267,14 +267,14 @@ torchrun --nproc_per_node 1 --standalone scripts/diffusion/inference.py configs/ We implemented an inference scaling sampling method inspaired by [Inference-Time Scaling for Diffusion Models beyond Scaling Denoising Steps](https://inference-scale-diffusion.github.io). You can spent more computational resources to get better results. Use it by specifying the sampling option. ``` -torchrun --nproc_per_node 4 --standalone scripts/diffusion/inference.py configs/diffusion/inference/768px_t2i2v_inference_scaling.py --save-dir samples --dataset.data-path assets/texts/sora.csv +torchrun --nproc_per_node 4 --standalone scripts/diffusion/inference.py configs/diffusion/inference/t2i2v_768px_inference_scaling.py --save-dir samples --dataset.data-path assets/texts/sora.csv ``` -| Orignal |
num_subtree=3
num_scaling_steps=5
num_noise=1
time=16min |
num_subtree=7
num_scaling_steps=8
num_noise=1
time=1h | +| Original |
num_subtree=3
num_scaling_steps=5
num_noise=1
time=16min |
num_subtree=7
num_scaling_steps=8
num_noise=1
time=1h | |----------------------|----------------------------------------------------------------|----------------------------------------------------------------| -| [Video Placeholder 1] | [Video Placeholder 2] | [Video Placeholder 3] | -| [Video Placeholder 1] | [Video Placeholder 2] | [Video Placeholder 3] | - +| | | | +| | | | +| | | | ### Reproductivity @@ -297,7 +297,7 @@ We test the computational efficiency of text-to-video on H100/H800 GPU. For 256x ## Evaluation -On [VBench](https://huggingface.co/spaces/Vchitect/VBench_Leaderboard), Open-Sora 2.0 significantly narrows the gap with OpenAI’s Sora, reducing it from 4.52% → 0.69% compared to Open-Sora 1.2. +On [VBench](https://huggingface.co/spaces/Vchitect/VBench_Leaderboard), Open-Sora 2.0 significantly narrows the gap with OpenAI's Sora, reducing it from 4.52% → 0.69% compared to Open-Sora 1.2. ![VBench](https://github.com/hpcaitech/Open-Sora-Demo/blob/main/readme/v2_vbench.png)