State of video generation in Diffusers

_blog.yml

@@ -5296,3 +5296,13 @@
     - open-source-collab
     - nlp
     - cv
+
+- local: video_gen
+  title: "State of open video generation models in Diffusers"
+  author: sayakpaul
+  thumbnail: /blog/assets/video_gen/thumbnail.png
+  date: Jan 23, 2025
+  tags:
+    - diffusers
+    - guide
+    - video_gen

video_gen.md

@@ -0,0 +1,242 @@
+---
+title: "State of open video generation models in Diffusers"
+thumbnail: /blog/assets/video_gen/thumbnail.png
+authors:
+- user: sayakpaul
+- user: a-r-r-o-w
+- user: dn6
+---
+
+# State of open video generation models in Diffusers
+
+OpenAI’s Sora demo marked a striking advance in AI-generated video last year and gave us a glimpse of the potential capabilities of video generation models. The impact was immediate and since that demo, the video generation space has become increasingly competitive with major players and startups producing their own highly capable models such as Google’s Veo2, Haliluo’s Minimax, Runway’s Gen3, Alpha Kling, Pika, and Luma Lab’s Dream Machine.   
+
+Open-source has also had its own surge of video generation models with CogVideoX, Mochi-1, Hunyuan, Allegro, and LTX Video. Is the video community having its “Stable Diffusion moment”?  
+
+This post will provide a brief overview of the state of video generation models, where we are with respect to open video generation models, and how the Diffusers team is planning to support their adoption at scale. 
+
+Specifically, we will discuss:
+
+- Capabilities and limitations of video generation models
+- Why video generation is hard
+- Open video generation models
+    - Licensing
+    - Memory requirements
+- Video generation with Diffusers
+    - Inference and optimizations
+    - Fine-tuning
+- Looking ahead
+
+## Today’s Video Generation Models and their Limitations
+
+As of today, the below models are amongst the most popular ones. 
+
+| **Provider** | **Model** | **Open/Closed** |
+| --- | --- | --- |
+| Meta | [MovieGen](https://ai.meta.com/research/movie-gen/) | Closed (with a detailed [technical report](https://ai.meta.com/research/publications/movie-gen-a-cast-of-media-foundation-models/)) |
+| OpenAI | [Sora](https://sora.com/) | Closed |
+| Google | [Veo 2](https://deepmind.google/technologies/veo/veo-2/) | Closed |
+| RunwayML | [Gen 3 Alpha](https://runwayml.com/research/introducing-gen-3-alpha) | Closed |
+| Pika Labs | [Pika 2.0](https://pika.art/login) | Closed |
+| KlingAI | [Kling](https://www.klingai.com/) | Closed |
+| Haliluo | [MiniMax](https://hailuoai.video/) | Closed |
+| THUDM | [CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox) | Open |
+| Genmo | [Mochi-1](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox) | Open |
+| RhymesAI | [Allegro](https://huggingface.co/docs/diffusers/main/en/api/pipelines/allegro) | Open |
+| Lightricks | [LTX Video](https://huggingface.co/docs/diffusers/main/en/api/pipelines/ltx_video) | Open |
+| Tencent | [Hunyuan Video](https://huggingface.co/docs/diffusers/main/api/pipelines/hunyuan_video) | Open |
+
+**Limitations**: Despite the continually increasing number of video generation models, their limitations are also manifold:
+
+- **High Resource Requirements:** Producing high-quality videos requires large pretrained models, which are computationally expensive to develop and deploy. These costs arise from dataset collection, hardware requirements, extensive training iterations and experimentation. These costs make it hard to justify producing open-source and freely available models. Even though we don’t have a detailed technical report that shed light into the training resources used, [this post](https://www.factorialfunds.com/blog/under-the-hood-how-openai-s-sora-model-works) provides some reasonable estimates.
+- Several open models suffer from limited generalization capabilities and underperform expectations of users. Models may require prompting in a certain way, or LLM-like prompts, or fail to generalize to out-of-distribution data, which are hurdles for widespread user adoption. For example, models like LTX-Video often need to be prompted in a very detailed and specific way for obtaining good quality generations.
+- The high computational and memory demands of video generation result in significant generation latency. For local usage, this is often a roadblock. Most new open video models are inaccessible to community hardware without extensive memory optimizations and quantization approaches that affect both inference latency and quality of the generated videos.
+
+## Why is Video Generation Hard?
+
+There are several factors we’d like to see and control in videos:
+
+- Adherence to Input Conditions (such as a text prompt, a starting image, etc.)
+- Realism
+- Aesthetics
+- Motion Dynamics
+- Spatio-Temporal Consistency and Coherence
+- FPS
+- Duration
+
+With image generation models, we usually only care about the first three aspects. However, for video generation we now have to consider motion quality, coherence and consistency over time, potentially with multiple subjects. Finding the right balance between good data, right inductive priors, and training methodologies to suit these additional requirements has proved to be more challenging than other modalities. 
+
+## Open Video Generation Models
+
+Text-to-video generation models have similar components as their text-to-image counterparts:
+
+- Text encoders for providing rich representations of the input text prompt
+- A denoising network
+- An encoder and decoder to convert between pixel and latent space
+- A non-parametric scheduler responsible for managing all the timestep-related calculations and the denoising step
+
+The latest generation of video models share a core feature where the denoising network processes 3D video tokens that capture both spatial and temporal information. The video encoder-decoder system, responsible for producing and decoding these tokens, employs both spatial and temporal compression. While decoding the latents typically demands the most memory, these models offer frame-by-frame decoding options to reduce memory usage.
+
+Text conditioning is incorporated through either joint attention (introduced in [Stable Diffusion 3](https://arxiv.org/abs/2403.03206)) or cross-attention. T5 has emerged as the preferred text encoder across most models, with HunYuan being an exception in its use of both CLIP-L and LLaMa 3.
+
+The denoising network itself builds on the DiT architecture developed by [William Peebles and Saining Xie](https://arxiv.org/abs/2212.09748), while incorporating various design elements from [PixArt](https://arxiv.org/abs/2310.00426).
+
+### Licensing
+
+The table below provides a list of the checkpoints of the most popular open video generation models, along with their licenses. Mochi-1, despite being a large and high-quality model, comes with an Apache 2.0 license! 
+
+| **Model Name** | **License** |
+| --- | --- |
+| [`THUDM/CogVideoX1.5-5B`](https://huggingface.co/THUDM/CogVideoX1.5-5B) | [Link](https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE) |
+| [`THUDM/CogVideoX1.5-5B-I2V`](https://huggingface.co/THUDM/CogVideoX1.5-5B-I2V) | [Link](https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE) |
+| [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b) | [Link](https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE) |
+| [`THUDM/CogVideoX-5b-I2V`](https://huggingface.co/THUDM/CogVideoX-5b-I2V) | [Link](https://huggingface.co/THUDM/CogVideoX-5b/blob/main/LICENSE) |
+| [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b) | Apache 2.0 |
+| [`genmo/mochi-1-preview`](https://huggingface.co/genmo/mochi-1-preview) | Apache 2.0 |
+| [`rhymes-ai/Allegro`](https://huggingface.co/rhymes-ai/Allegro) | Apache 2.0 |
+| [`tencent/HunyuanVideo`](https://huggingface.co/tencent/HunyuanVideo) | [Link](https://huggingface.co/tencent/HunyuanVideo/blob/main/LICENSE) |
+| [`Lightricks/LTX-Video`](https://huggingface.co/Lightricks/LTX-Video) | [Link](https://huggingface.co/Lightricks/LTX-Video/blob/main/License.txt) |
+
+### **Memory requirements**
+
+The memory requirements for any model can be computed by adding the following:
+
+- Memory required for weights
+- Maximum memory required for storing intermediate activation states
+
+Memory required by weights can be lowered via - quantization, downcasting to lower dtypes, or offloading to CPU. Memory required for activations states can also be lowered but is usually a more involved process, which is out of the scope of this blog. 
+
+It is possible to run any video model with extremely low memory, but it comes at the cost of time required for inference. If the time required by an optimization technique is more than what a user considers reasonable, it is not feasible to run inference. Diffusers provides many such optimizations that are opt-in and can be chained together.
+
+In the table below, we provide the memory requirements for three popular video generation models with reasonable defaults:
+
+| **Model Name** | **Memory (GB)** |
+| --- | --- |
+| HunyuanVideo | 60.09 |
+| LTX-Video | 17.75 |
+| CogVideoX (1.5 5B) | 36.51 |
+
+These numbers were obtained with the following settings on an 80GB A100 machine (full script [here](https://gist.github.com/sayakpaul/2bc49a30cf76cea07914104d28b1fb86)):
+
+- `torch.bfloat16` dtype
+- `num_frames`: 121, `height`: 512, `width`: 768
+- `max_sequence_length`: 128
+- `num_inference_dtype`: 50
+
+These requirements are quite staggering, making these models difficult to run on consumer hardware. As mentioned above, with Diffusers, users can enable different optimizations to suit their needs. The following table provides memory requirements for widely used models with sensible optimizations enabled (that do not compromise on quality or time required for inference). We studied this with the HunyuanVideo model as it’s sufficiently large to show the benefits of the optimizations in a progressive manner.
+
+| Base | 60.10 GB |
+| --- | --- |
+| VAE tiling | 43.58 GB |
+| CPU offloading | 28.87 GB |
+| 8Bit | 49.9 GB |
+| 8Bit + CPU offloading* | 35.66 GB |
+| 8Bit + VAE tiling | 36.92 GB |
+| 8Bit + CPU offloading + VAE tiling | 26.18 GB |
+| 4Bit | 42.96 GB |
+| 4Bit + CPU offloading | 21.99 GB |
+| 4Bit + VAE tiling | 26.42 GB |
+| 4Bit + CPU offloading + VAE tiling | 14.15 GB |
+
+*8Bit models in `bitsandbytes` cannot be moved to CPU from GPU, unlike the 4Bit ones.
+
+We used the same settings as above to obtain these numbers. Quantization was performed with the [`bitsandbytes` library](https://huggingface.co/docs/bitsandbytes/main/en/index) (Diffusers [supports three different quantization backends](https://huggingface.co/docs/diffusers/main/en/quantization/overview) as of now). Also note that due to numerical precision loss, quantization can impact the quality of the outputs, effects of which are more prominent in videos than images.
+
+## Video Generation with Diffusers
+
+<div align="center">
+<iframe src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/video_gen/hunyuan-output.mp4" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
+</div>
+
+There are three broad categories of generation possible when working with video models:
+
+1. Text to Video
+2. Image or Image Control condition + Text to Video
+3. Video or Video Control condition + Text to Video
+
+### Suite of optimizations
+
+Video generation can be quite difficult on resource-constrained devices and time-consuming even on beefier GPUs. Diffusers provides a suite of utilities that help to optimize both the runtime and memory consumption of these models. These optimizations fall under the following categories:
+
+- **Quantization**: The model weights are quantized to lower precision data types, which lowers the VRAM requirements of models.
+- **Offloading**: Different layers of a model can be loaded on the GPU when required for computation on-the-fly and then offloaded back to CPU. This saves a significant amount of memory during inference.
+- **Chunked Inference**: By splitting inference across non-embedding dimensions of input latent tensors, the memory overheads from intermediate activation states can be reduced. Common use of this technique is often seen in encoder/decoder slicing/tiling.
+- **Re-use of Attention & MLP states**: Computation of certain denoising steps can be skipped and past states can be re-used, if certain conditions are satisfied for particular algorithms, to speed up the generation process with minimal quality loss.
+
+Note that in the above four options, as of now, we only support the first two. Support for the rest of the two will be merged in soon. If you’re interested to follow along the progress, here are the PRs:
+
+- TODO:
+- TODO:
+
+The list of memory optimizations discussed here will soon become non-exhaustive, so, we suggest you to always keep an eye on the Diffusers repository to stay updated. 
+
+We can also apply optimizations during training. The two most well-known techniques applied to video models include:
+
+- **Timestep distillation**: This involves teaching the model to denoise the noisy latents faster in lesser amount of inference steps, in a recursive fashion. For example, if a model takes 32 steps to generate good videos, it can be augmented to try and predict the final outputs in only 16-steps, or 8-steps, or even 2-steps! This may be accompanied by loss in quality depending on how fewer steps are used. Some examples of timestep-distilled models include [Flux.1-Schnell](https://huggingface.co/black-forest-labs/FLUX.1-schnell/) and [FastHunyuan](https://huggingface.co/FastVideo/FastHunyuan).
+- **Guidance distillation**: [Classifier-Free Guidance](https://arxiv.org/abs/2207.12598) is a technique widely used in diffusion models that enhances generation quality. This, however, doubles the generation time because it involves two full forward passes through the models per inference step, followed by an interpolation step. By teaching models to predict the output of both forward passes and interpolation at the cost of one forward pass, this method can enable much faster generation. Some examples of guidance-distilled models include [Flux.1-Dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) and [HunyuanVideo](https://huggingface.co/docs/diffusers/main/api/pipelines/hunyuan_video).
+- Architectural compression through distillation as done in [SSD1B](https://huggingface.co/segmind/SSD-1B).
+
+We refer the readers to [this guide](https://huggingface.co/docs/diffusers/main/en/using-diffusers/text-img2vid) for a detailed take on video generation and the current possibilities in Diffusers.
+
+### Fine-tuning
+
+We’ve created [`finetrainers`](https://github.com/a-r-r-o-w/finetrainers) - a repository that allows you to easily fine-tune the latest generation of open video models. For example, here is how you would fine-tune CogVideoX with LoRA:
+
+```bash
+# Download a dataset
+huggingface-cli download \
+  --repo-type dataset Wild-Heart/Disney-VideoGeneration-Dataset \
+  --local-dir video-dataset-disney
+
+# Then launch training
+accelerate launch train.py \
+  --model_name="cogvideox" --pretrained_model_name_or_path="THUDM/CogVideoX1.5-5B" \
+  --data_root="video-dataset-disney" \
+  --video_column="videos.txt" \
+  --caption_column="prompt.txt" \
+  --training_type="lora" \
+  --seed=42 \
+  --mixed_precision="bf16" \
+  --batch_size=1 \
+  --train_steps=1200 \
+  --rank=128 \
+  --lora_alpha=128 \
+  --target_modules to_q to_k to_v to_out.0 \
+  --gradient_accumulation_steps 1 \
+  --gradient_checkpointing \
+  --checkpointing_steps 500 \
+  --checkpointing_limit 2 \
+  --enable_slicing \
+  --enable_tiling \
+  --optimizer adamw \
+  --lr 3e-5 \
+  --lr_scheduler constant_with_warmup \
+  --lr_warmup_steps 100 \
+  --lr_num_cycles 1 \
+  --beta1 0.9 \
+  --beta2 0.95 \
+  --weight_decay 1e-4 \
+  --epsilon 1e-8 \
+  --max_grad_norm 1.0
+
+# ...
+# (Full training command removed for brevity)
+```
+
+For more details, check out the repository [here](https://github.com/a-r-r-o-w/finetrainers).
+
+## Looking ahead
+
+As it has become quite apparent that video generation models will continue to grow in 2025, Diffusers users can expect more optimization-related goodies. Our goal is to also make it easy and accessible to do video model fine-tuning which is why we will continue to grow the `finetrainers` library. LoRA training is just the beginning, but there’s more to come - Control LoRAs, Distillation algorithms, ControlNets, Adapters, and more. We would love to welcome contributions from the community as we go 🤗 
+
+We will also continue to collaborate with model publishers, fine-tuners, and anyone from the community willing to help us take the state of video generation to the next level and bring you the latest and the greatest in the domain. 
+
+## Resources
+
+We cited a number of links throughout the post. To make sure you don’t miss out on the most important ones, we provide a list below:
+
+- [Video generation guide](https://huggingface.co/docs/diffusers/main/en/using-diffusers/text-img2vid)
+- [Quantization support in Diffusers](https://huggingface.co/docs/diffusers/main/en/quantization/overview)
+- [General LoRA guide in Diffusers](https://huggingface.co/docs/diffusers/main/en/tutorials/using_peft_for_inference)
+- [Memory optimization guide for CogVideoX](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cogvideox#memory-optimization) (it applies to other video models, too)
+- [`finetrainers`](https://github.com/a-r-r-o-w/finetrainers) for fine-tuning