[Lumina-Next : Making Lumina-T2X Stronger and Faster with Next-DiT](https://github.com/Alpha-VLLM/Lumina-T2X/blob/main/assets/lumina-next.pdf) from Alpha-VLLM, OpenGVLab, Shanghai AI Laboratory.
The abstract from the paper is:
*Lumina-T2X is a nascent family of Flow-based Large Diffusion Transformers (Flag-DiT) that establishes a unified framework for transforming noise into various modalities, such as images and videos, conditioned on text instructions. Despite its promising capabilities, Lumina-T2X still encounters challenges including training instability, slow inference, and extrapolation artifacts. In this paper, we present Lumina-Next, an improved version of Lumina-T2X, showcasing stronger generation performance with increased training and inference efficiency. We begin with a comprehensive analysis of the Flag-DiT architecture and identify several suboptimal components, which we address by introducing the Next-DiT architecture with 3D RoPE and sandwich normalizations. To enable better resolution extrapolation, we thoroughly compare different context extrapolation methods applied to text-to-image generation with 3D RoPE, and propose Frequency- and Time-Aware Scaled RoPE tailored for diffusion transformers. Additionally, we introduce a sigmoid time discretization schedule to reduce sampling steps in solving the Flow ODE and the Context Drop method to merge redundant visual tokens for faster network evaluation, effectively boosting the overall sampling speed. Thanks to these improvements, Lumina-Next not only improves the quality and efficiency of basic text-to-image generation but also demonstrates superior resolution extrapolation capabilities and multilingual generation using decoder-based LLMs as the text encoder, all in a zero-shot manner. To further validate Lumina-Next as a versatile generative framework, we instantiate it on diverse tasks including visual recognition, multi-view, audio, music, and point cloud generation, showcasing strong performance across these domains. By releasing all codes and model weights at https://github.com/Alpha-VLLM/Lumina-T2X, we aim to advance the development of next-generation generative AI capable of universal modeling.*
**Highlights**: Lumina-Next is a next-generation Diffusion Transformer that significantly enhances text-to-image generation, multilingual generation, and multitask performance by introducing the Next-DiT architecture, 3D RoPE, and frequency- and time-aware RoPE, among other improvements.
Lumina-Next has the following components:
* It improves sampling efficiency with fewer and faster Steps.
* It uses a Next-DiT as a transformer backbone with Sandwichnorm 3D RoPE, and Grouped-Query Attention.
* It uses a Frequency- and Time-Aware Scaled RoPE.
---
[Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers](https://huggingface.co/papers/2405.05945) from Alpha-VLLM, OpenGVLab, Shanghai AI Laboratory.
The abstract from the paper is:
*Sora unveils the potential of scaling Diffusion Transformer for generating photorealistic images and videos at arbitrary resolutions, aspect ratios, and durations, yet it still lacks sufficient implementation details. In this technical report, we introduce the Lumina-T2X family - a series of Flow-based Large Diffusion Transformers (Flag-DiT) equipped with zero-initialized attention, as a unified framework designed to transform noise into images, videos, multi-view 3D objects, and audio clips conditioned on text instructions. By tokenizing the latent spatial-temporal space and incorporating learnable placeholders such as [nextline] and [nextframe] tokens, Lumina-T2X seamlessly unifies the representations of different modalities across various spatial-temporal resolutions. This unified approach enables training within a single framework for different modalities and allows for flexible generation of multimodal data at any resolution, aspect ratio, and length during inference. Advanced techniques like RoPE, RMSNorm, and flow matching enhance the stability, flexibility, and scalability of Flag-DiT, enabling models of Lumina-T2X to scale up to 7 billion parameters and extend the context window to 128K tokens. This is particularly beneficial for creating ultra-high-definition images with our Lumina-T2I model and long 720p videos with our Lumina-T2V model. Remarkably, Lumina-T2I, powered by a 5-billion-parameter Flag-DiT, requires only 35% of the training computational costs of a 600-million-parameter naive DiT. Our further comprehensive analysis underscores Lumina-T2X's preliminary capability in resolution extrapolation, high-resolution editing, generating consistent 3D views, and synthesizing videos with seamless transitions. We expect that the open-sourcing of Lumina-T2X will further foster creativity, transparency, and diversity in the generative AI community.*
You can find the original codebase at [Alpha-VLLM](https://github.com/Alpha-VLLM/Lumina-T2X) and all the available checkpoints at [Alpha-VLLM Lumina Family](https://huggingface.co/collections/Alpha-VLLM/lumina-family-66423205bedb81171fd0644b).
**Highlights**: Lumina-T2X supports Any Modality, Resolution, and Duration.
Lumina-T2X has the following components:
* It uses a Flow-based Large Diffusion Transformer as the backbone
* It supports different any modalities with one backbone and corresponding encoder, decoder.
This pipeline was contributed by [PommesPeter](https://github.com/PommesPeter). The original codebase can be found [here](https://github.com/Alpha-VLLM/Lumina-T2X). The original weights can be found under [hf.co/Alpha-VLLM](https://huggingface.co/Alpha-VLLM).
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
### Inference (Text-to-Image)
Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
image=pipeline(prompt="Upper body of a young woman in a Victorian-era outfit with brass goggles and leather straps. Background shows an industrial revolution cityscape with smoky skies and tall, metal structures").images[0]
```
## Quantization
Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`LuminaPipeline`] for inference with bitsandbytes.
[Lumina Image 2.0: A Unified and Efficient Image Generative Model](https://huggingface.co/Alpha-VLLM/Lumina-Image-2.0) is a 2 billion parameter flow-based diffusion transformer capable of generating diverse images from text descriptions.
The abstract from the paper is:
*We introduce Lumina-Image 2.0, an advanced text-to-image model that surpasses previous state-of-the-art methods across multiple benchmarks, while also shedding light on its potential to evolve into a generalist vision intelligence model. Lumina-Image 2.0 exhibits three key properties: (1) Unification – it adopts a unified architecture that treats text and image tokens as a joint sequence, enabling natural cross-modal interactions and facilitating task expansion. Besides, since high-quality captioners can provide semantically better-aligned text-image training pairs, we introduce a unified captioning system, UniCaptioner, which generates comprehensive and precise captions for the model. This not only accelerates model convergence but also enhances prompt adherence, variable-length prompt handling, and task generalization via prompt templates. (2) Efficiency – to improve the efficiency of the unified architecture, we develop a set of optimization techniques that improve semantic learning and fine-grained texture generation during training while incorporating inference-time acceleration strategies without compromising image quality. (3) Transparency – we open-source all training details, code, and models to ensure full reproducibility, aiming to bridge the gap between well-resourced closed-source research teams and independent developers.*
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
## Using Single File loading with Lumina Image 2.0
Single file loading for Lumina Image 2.0 is available for the `Lumina2Transformer2DModel`
| [prs-eth/marigold-depth-v1-1](https://huggingface.co/prs-eth/marigold-depth-v1-1) | Depth | Affine-invariant depth prediction assigns each pixel a value between 0 (near plane) and 1 (far plane), with both planes determined by the model during inference. |
| [prs-eth/marigold-normals-v0-1](https://huggingface.co/prs-eth/marigold-normals-v0-1) | Normals | The surface normals predictions are unit-length 3D vectors in the screen space camera, with values in the range from -1 to 1. |
| [prs-eth/marigold-iid-appearance-v1-1](https://huggingface.co/prs-eth/marigold-iid-appearance-v1-1) | Intrinsics | InteriorVerse decomposition is comprised of Albedo and two BRDF material properties: Roughness and Metallicity. |
| [prs-eth/marigold-iid-lighting-v1-1](https://huggingface.co/prs-eth/marigold-iid-lighting-v1-1) | Intrinsics | HyperSim decomposition of an image $I$ is comprised of Albedo $A$, Diffuse shading $S$, and Non-diffuse residual $R$: $I = A*S+R$. |
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff
> between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to
> efficiently load the same components into multiple pipelines.
> Also, to know more about reducing the memory usage of this pipeline, refer to the ["Reduce memory usage"] section
> Only a research preview of the model weights is available at the moment.
[Mochi 1](https://huggingface.co/genmo/mochi-1-preview) is a video generation model by Genmo with a strong focus on prompt adherence and motion quality. The model features a 10B parameter Asmmetric Diffusion Transformer (AsymmDiT) architecture, and uses non-square QKV and output projection layers to reduce inference memory requirements. A single T5-XXL model is used to encode prompts.
*Mochi 1 preview is an open state-of-the-art video generation model with high-fidelity motion and strong prompt adherence in preliminary evaluation. This model dramatically closes the gap between closed and open video generation systems. The model is released under a permissive Apache 2.0 license.*
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
## Quantization
Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`MochiPipeline`] for inference with bitsandbytes.
"Close-up of a cats eye, with the galaxy reflected in the cats eye. Ultra high resolution 4k.",
num_inference_steps=28,
guidance_scale=3.5
).frames[0]
export_to_video(video,"cat.mp4")
```
## Generating videos with Mochi-1 Preview
The following example will download the full precision `mochi-1-preview` weights and produce the highest quality results but will require at least 42GB VRAM to run.
The following example will use the `bfloat16` variant of the model and requires 22GB VRAM to run. There is a slight drop in the quality of the generated video as a result.
prompt="Close-up of a chameleon's eye, with its scaly skin changing color. Ultra high resolution 4k."
frames=pipe(prompt,num_frames=85).frames[0]
export_to_video(frames,"mochi.mp4",fps=30)
```
## Reproducing the results from the Genmo Mochi repo
The [Genmo Mochi implementation](https://github.com/genmoai/mochi/tree/main) uses different precision values for each stage in the inference process. The text encoder and VAE use `torch.float32`, while the DiT uses `torch.bfloat16` with the [attention kernel](https://pytorch.org/docs/stable/generated/torch.nn.attention.sdpa_kernel.html#torch.nn.attention.sdpa_kernel) set to `EFFICIENT_ATTENTION`. Diffusers pipelines currently do not support setting different `dtypes` for different stages of the pipeline. In order to run inference in the same way as the original implementation, please refer to the following example.
> [!TIP]
> The original Mochi implementation zeros out empty prompts. However, enabling this option and placing the entire pipeline under autocast can lead to numerical overflows with the T5 text encoder.
>
> When enabling `force_zeros_for_empty_prompt`, it is recommended to run the text encoding step outside the autocast context in full precision.
> [!TIP]
> Decoding the latents in full precision is very memory intensive. You will need at least 70GB VRAM to generate the 163 frames in this example. To reduce memory, either reduce the number of frames or run the decoding step in `torch.bfloat16`.
It is possible to split the large Mochi transformer across multiple GPUs using the `device_map` and `max_memory` options in `from_pretrained`. In the following example we split the model across two GPUs, each with 24GB of VRAM.
*The emergence of Large Language Models (LLMs) has unified language generation tasks and revolutionized human-machine interaction. However, in the realm of image generation, a unified model capable of handling various tasks within a single framework remains largely unexplored. In this work, we introduce OmniGen, a new diffusion model for unified image generation. OmniGen is characterized by the following features: 1) Unification: OmniGen not only demonstrates text-to-image generation capabilities but also inherently supports various downstream tasks, such as image editing, subject-driven generation, and visual conditional generation. 2) Simplicity: The architecture of OmniGen is highly simplified, eliminating the need for additional plugins. Moreover, compared to existing diffusion models, it is more user-friendly and can complete complex tasks end-to-end through instructions without the need for extra intermediate steps, greatly simplifying the image generation workflow. 3) Knowledge Transfer: Benefit from learning in a unified format, OmniGen effectively transfers knowledge across different tasks, manages unseen tasks and domains, and exhibits novel capabilities. We also explore the model’s reasoning capabilities and potential applications of the chain-of-thought mechanism. This work represents the first attempt at a general-purpose image generation model, and we will release our resources at https://github.com/VectorSpaceLab/OmniGen to foster future advancements.*
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.md) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading.md#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
This pipeline was contributed by [staoxiao](https://github.com/staoxiao). The original codebase can be found [here](https://github.com/VectorSpaceLab/OmniGen). The original weights can be found under [hf.co/shitao](https://huggingface.co/Shitao/OmniGen-v1).
For text-to-image, pass a text prompt. By default, OmniGen generates a 1024x1024 image.
You can try setting the `height` and `width` parameters to generate images with different size.
```python
prompt="Realistic photo. A young woman sits on a sofa, holding a book and facing the camera. She wears delicate silver hoop earrings adorned with tiny, sparkling diamonds that catch the light, with her long chestnut hair cascading over her shoulders. Her eyes are focused and gentle, framed by long, dark lashes. She is dressed in a cozy cream sweater, which complements her warm, inviting smile. Behind her, there is a table with a cup of water in a sleek, minimalist blue mug. The background is a serene indoor setting with soft natural light filtering through a window, adorned with tasteful art and flowers, creating a cozy and peaceful ambiance. 4K, HD."
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Pipelines
Pipelines provide a simple way to run state-of-the-art diffusion models in inference by bundling all of the necessary components (multiple independently-trained models, schedulers, and processors) into a single end-to-end class. Pipelines are flexible and they can be adapted to use different schedulers or even model components.
All pipelines are built from the base [`DiffusionPipeline`] class which provides basic functionality for loading, downloading, and saving all the components. Specific pipeline types (for example [`StableDiffusionPipeline`]) loaded with [`~DiffusionPipeline.from_pretrained`] are automatically detected and the pipeline components are loaded and passed to the `__init__` function of the pipeline.
> [!WARNING]
> You shouldn't use the [`DiffusionPipeline`] class for training. Individual components (for example, [`UNet2DModel`] and [`UNet2DConditionModel`]) of diffusion pipelines are usually trained individually, so we suggest directly working with them instead.
>
> <br>
>
> Pipelines do not offer any training functionality. You'll notice PyTorch's autograd is disabled by decorating the [`~DiffusionPipeline.__call__`] method with a [`torch.no_grad`](https://pytorch.org/docs/stable/generated/torch.no_grad.html) decorator because pipelines should not be used for training. If you're interested in training, please take a look at the [Training](../../training/overview) guides instead!
The table below lists all the pipelines currently available in 🤗 Diffusers and the tasks they support. Click on a pipeline to view its abstract and published paper.
Ovis-Image is a 7B text-to-image model specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints.
[Ovis-Image Technical Report](https://arxiv.org/abs/2511.22982) from Alibaba Group, by Guo-Hua Wang, Liangfu Cao, Tianyu Cui, Minghao Fu, Xiaohao Chen, Pengxin Zhan, Jianshan Zhao, Lan Li, Bowen Fu, Jiaqi Liu, Qing-Guo Chen.
The abstract from the paper is:
*We introduce Ovis-Image, a 7B text-to-image model specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints. Built upon our previous Ovis-U1 framework, Ovis-Image integrates a diffusion-based visual decoder with the stronger Ovis 2.5 multimodal backbone, leveraging a text-centric training pipeline that combines large-scale pre-training with carefully tailored post-training refinements. Despite its compact architecture, Ovis-Image achieves text rendering performance on par with significantly larger open models such as Qwen-Image and approaches closed-source systems like Seedream and GPT4o. Crucially, the model remains deployable on a single high-end GPU with moderate memory, narrowing the gap between frontier-level text rendering and practical deployment. Our results indicate that combining a strong multimodal backbone with a carefully designed, text-focused training recipe is sufficient to achieve reliable bilingual text rendering without resorting to oversized or proprietary models.*
**Highlights**:
***Strong text rendering at a compact 7B scale**: Ovis-Image is a 7B text-to-image model that delivers text rendering quality comparable to much larger 20B-class systems such as Qwen-Image and competitive with leading closed-source models like GPT4o in text-centric scenarios, while remaining small enough to run on widely accessible hardware.
***High fidelity on text-heavy, layout-sensitive prompts**: The model excels on prompts that demand tight alignment between linguistic content and rendered typography (e.g., posters, banners, logos, UI mockups, infographics), producing legible, correctly spelled, and semantically consistent text across diverse fonts, sizes, and aspect ratios without compromising overall visual quality.
***Efficiency and deployability**: With its 7B parameter budget and streamlined architecture, Ovis-Image fits on a single high-end GPU with moderate memory, supports low-latency interactive use, and scales to batch production serving, bringing near–frontier text rendering to applications where tens-of-billions–parameter models are impractical.
This pipeline was contributed by Ovis-Image Team. The original codebase can be found [here](https://github.com/AIDC-AI/Ovis-Image).
[Perturbed-Attention Guidance (PAG)](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) is a new diffusion sampling guidance that improves sample quality across both unconditional and conditional settings, achieving this without requiring further training or the integration of external modules.
PAG was introduced in [Self-Rectifying Diffusion Sampling with Perturbed-Attention Guidance](https://huggingface.co/papers/2403.17377) by Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin and Seungryong Kim.
The abstract from the paper is:
*Recent studies have demonstrated that diffusion models are capable of generating high-quality samples, but their quality heavily depends on sampling guidance techniques, such as classifier guidance (CG) and classifier-free guidance (CFG). These techniques are often not applicable in unconditional generation or in various downstream tasks such as image restoration. In this paper, we propose a novel sampling guidance, called Perturbed-Attention Guidance (PAG), which improves diffusion sample quality across both unconditional and conditional settings, achieving this without requiring additional training or the integration of external modules. PAG is designed to progressively enhance the structure of samples throughout the denoising process. It involves generating intermediate samples with degraded structure by substituting selected self-attention maps in diffusion U-Net with an identity matrix, by considering the self-attention mechanisms' ability to capture structural information, and guiding the denoising process away from these degraded samples. In both ADM and Stable Diffusion, PAG surprisingly improves sample quality in conditional and even unconditional scenarios. Moreover, PAG significantly improves the baseline performance in various downstream tasks where existing guidances such as CG or CFG cannot be fully utilized, including ControlNet with empty prompts and image restoration such as inpainting and deblurring.*
PAG can be used by specifying the `pag_applied_layers` as a parameter when instantiating a PAG pipeline. It can be a single string or a list of strings. Each string can be a unique layer identifier or a regular expression to identify one or more layers.
- Full identifier as a normal string: `down_blocks.2.attentions.0.transformer_blocks.0.attn1.processor`
- Full identifier as a RegEx: `down_blocks.2.(attentions|motion_modules).0.transformer_blocks.0.attn1.processor`
- Partial identifier as a RegEx: `down_blocks.2`, or `attn1`
- List of identifiers (can be combo of strings and ReGex): `["blocks.1", "blocks.(14|20)", r"down_blocks\.(2,3)"]`
> [!WARNING]
> Since RegEx is supported as a way for matching layer identifiers, it is crucial to use it correctly otherwise there might be unexpected behaviour. The recommended way to use PAG is by specifying layers as `blocks.{layer_index}` and `blocks.({layer_index_1|layer_index_2|...})`. Using it in any other way, while doable, may bypass our basic validation checks and give you unexpected results.
[InstructPix2Pix: Learning to Follow Image Editing Instructions](https://huggingface.co/papers/2211.09800) is by Tim Brooks, Aleksander Holynski and Alexei A. Efros.
The abstract from the paper is:
*We propose a method for editing images from human instructions: given an input image and a written instruction that tells the model what to do, our model follows these instructions to edit the image. To obtain training data for this problem, we combine the knowledge of two large pretrained models -- a language model (GPT-3) and a text-to-image model (Stable Diffusion) -- to generate a large dataset of image editing examples. Our conditional diffusion model, InstructPix2Pix, is trained on our generated data, and generalizes to real images and user-written instructions at inference time. Since it performs edits in the forward pass and does not require per example fine-tuning or inversion, our model edits images quickly, in a matter of seconds. We show compelling editing results for a diverse collection of input images and written instructions.*
You can find additional information about InstructPix2Pix on the [project page](https://www.timothybrooks.com/instruct-pix2pix), [original codebase](https://github.com/timothybrooks/instruct-pix2pix), and try it out in a [demo](https://huggingface.co/spaces/timbrooks/instruct-pix2pix).
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
[PixArt-α: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis](https://huggingface.co/papers/2310.00426) is Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li.
The abstract from the paper is:
*The most advanced text-to-image (T2I) models require significant training costs (e.g., millions of GPU hours), seriously hindering the fundamental innovation for the AIGC community while increasing CO2 emissions. This paper introduces PIXART-α, a Transformer-based T2I diffusion model whose image generation quality is competitive with state-of-the-art image generators (e.g., Imagen, SDXL, and even Midjourney), reaching near-commercial application standards. Additionally, it supports high-resolution image synthesis up to 1024px resolution with low training cost, as shown in Figure 1 and 2. To achieve this goal, three core designs are proposed: (1) Training strategy decomposition: We devise three distinct training steps that separately optimize pixel dependency, text-image alignment, and image aesthetic quality; (2) Efficient T2I Transformer: We incorporate cross-attention modules into Diffusion Transformer (DiT) to inject text conditions and streamline the computation-intensive class-condition branch; (3) High-informative data: We emphasize the significance of concept density in text-image pairs and leverage a large Vision-Language model to auto-label dense pseudo-captions to assist text-image alignment learning. As a result, PIXART-α's training speed markedly surpasses existing large-scale T2I models, e.g., PIXART-α only takes 10.8% of Stable Diffusion v1.5's training time (675 vs. 6,250 A100 GPU days), saving nearly $300,000 ($26,000 vs. $320,000) and reducing 90% CO2 emissions. Moreover, compared with a larger SOTA model, RAPHAEL, our training cost is merely 1%. Extensive experiments demonstrate that PIXART-α excels in image quality, artistry, and semantic control. We hope PIXART-α will provide new insights to the AIGC community and startups to accelerate building their own high-quality yet low-cost generative models from scratch.*
You can find the original codebase at [PixArt-alpha/PixArt-alpha](https://github.com/PixArt-alpha/PixArt-alpha) and all the available checkpoints at [PixArt-alpha](https://huggingface.co/PixArt-alpha).
Some notes about this pipeline:
* It uses a Transformer backbone (instead of a UNet) for denoising. As such it has a similar architecture as [DiT](./dit).
* It was trained using text conditions computed from T5. This aspect makes the pipeline better at following complex text prompts with intricate details.
* It is good at producing high-resolution images at different aspect ratios. To get the best results, the authors recommend some size brackets which can be found [here](https://github.com/PixArt-alpha/PixArt-alpha/blob/08fbbd281ec96866109bdd2cdb75f2f58fb17610/diffusion/data/datasets/utils.py).
* It rivals the quality of state-of-the-art text-to-image generation systems (as of this writing) such as Stable Diffusion XL, Imagen, and DALL-E 2, while being more efficient than them.
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
## Inference with under 8GB GPU VRAM
Run the [`PixArtAlphaPipeline`] with under 8GB GPU VRAM by loading the text encoder in 8-bit precision. Let's walk through a full-fledged example.
First, install the [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) library:
If you want a report of your memory-usage, run this [script](https://gist.github.com/sayakpaul/3ae0f847001d342af27018a96f467e4e).
> [!WARNING]
> Text embeddings computed in 8-bit can impact the quality of the generated images because of the information loss in the representation space caused by the reduced precision. It's recommended to compare the outputs with and without 8-bit.
While loading the `text_encoder`, you set `load_in_8bit` to `True`. You could also specify `load_in_4bit` to bring your memory requirements down even further to under 7GB.
[PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation](https://huggingface.co/papers/2403.04692) is Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li.
The abstract from the paper is:
*In this paper, we introduce PixArt-Σ, a Diffusion Transformer model (DiT) capable of directly generating images at 4K resolution. PixArt-Σ represents a significant advancement over its predecessor, PixArt-α, offering images of markedly higher fidelity and improved alignment with text prompts. A key feature of PixArt-Σ is its training efficiency. Leveraging the foundational pre-training of PixArt-α, it evolves from the ‘weaker’ baseline to a ‘stronger’ model via incorporating higher quality data, a process we term “weak-to-strong training”. The advancements in PixArt-Σ are twofold: (1) High-Quality Training Data: PixArt-Σ incorporates superior-quality image data, paired with more precise and detailed image captions. (2) Efficient Token Compression: we propose a novel attention module within the DiT framework that compresses both keys and values, significantly improving efficiency and facilitating ultra-high-resolution image generation. Thanks to these improvements, PixArt-Σ achieves superior image quality and user prompt adherence capabilities with significantly smaller model size (0.6B parameters) than existing text-to-image diffusion models, such as SDXL (2.6B parameters) and SD Cascade (5.1B parameters). Moreover, PixArt-Σ’s capability to generate 4K images supports the creation of high-resolution posters and wallpapers, efficiently bolstering the production of highquality visual content in industries such as film and gaming.*
You can find the original codebase at [PixArt-alpha/PixArt-sigma](https://github.com/PixArt-alpha/PixArt-sigma) and all the available checkpoints at [PixArt-alpha](https://huggingface.co/PixArt-alpha).
Some notes about this pipeline:
* It uses a Transformer backbone (instead of a UNet) for denoising. As such it has a similar architecture as [DiT](https://hf.co/docs/transformers/model_doc/dit).
* It was trained using text conditions computed from T5. This aspect makes the pipeline better at following complex text prompts with intricate details.
* It is good at producing high-resolution images at different aspect ratios. To get the best results, the authors recommend some size brackets which can be found [here](https://github.com/PixArt-alpha/PixArt-sigma/blob/master/diffusion/data/datasets/utils.py).
* It rivals the quality of state-of-the-art text-to-image generation systems (as of this writing) such as PixArt-α, Stable Diffusion XL, Playground V2.0 and DALL-E 3, while being more efficient than them.
* It shows the ability of generating super high resolution images, such as 2048px or even 4K.
* It shows that text-to-image models can grow from a weak model to a stronger one through several improvements (VAEs, datasets, and so on.)
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
> [!TIP]
> You can further improve generation quality by passing the generated image from [`PixArtSigmaPipeline`] to the [SDXL refiner](../../using-diffusers/sdxl#base-to-refiner-model) model.
## Inference with under 8GB GPU VRAM
Run the [`PixArtSigmaPipeline`] with under 8GB GPU VRAM by loading the text encoder in 8-bit precision. Let's walk through a full-fledged example.
First, install the [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) library:
If you want a report of your memory-usage, run this [script](https://gist.github.com/sayakpaul/3ae0f847001d342af27018a96f467e4e).
> [!WARNING]
> Text embeddings computed in 8-bit can impact the quality of the generated images because of the information loss in the representation space caused by the reduced precision. It's recommended to compare the outputs with and without 8-bit.
While loading the `text_encoder`, you set `load_in_8bit` to `True`. You could also specify `load_in_4bit` to bring your memory requirements down even further to under 7GB.
<!-- Copyright 2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License. -->
# PRX
PRX generates high-quality images from text using a simplified MMDIT architecture where text tokens don't update through transformer blocks. It employs flow matching with discrete scheduling for efficient sampling and uses Google's T5Gemma-2B-2B-UL2 model for multi-language text encoding. The ~1.3B parameter transformer delivers fast inference without sacrificing quality. You can choose between Flux VAE (8x compression, 16 latent channels) for balanced quality and speed or DC-AE (32x compression, 32 latent channels) for latent compression and faster processing.
## Available models
PRX offers multiple variants with different VAE configurations, each optimized for specific resolutions. Base models excel with detailed prompts, capturing complex compositions and subtle details. Fine-tuned models trained on the [Alchemist dataset](https://huggingface.co/datasets/yandex/alchemist) improve aesthetic quality, especially with simpler prompts.
| [`Photoroom/prx-256-t2i`](https://huggingface.co/Photoroom/prx-256-t2i)| 256 | No | No | Base model pre-trained at 256 with Flux VAE|Works best with detailed prompts in natural language|28 steps, cfg=5.0| `torch.bfloat16` |
| [`Photoroom/prx-256-t2i-sft`](https://huggingface.co/Photoroom/prx-256-t2i-sft)| 512 | Yes | No | Fine-tuned on the [Alchemist dataset](https://huggingface.co/datasets/yandex/alchemist) dataset with Flux VAE | Can handle less detailed prompts|28 steps, cfg=5.0| `torch.bfloat16` |
| [`Photoroom/prx-512-t2i`](https://huggingface.co/Photoroom/prx-512-t2i)| 512 | No | No | Base model pre-trained at 512 with Flux VAE |Works best with detailed prompts in natural language|28 steps, cfg=5.0| `torch.bfloat16` |
| [`Photoroom/prx-512-t2i-sft`](https://huggingface.co/Photoroom/prx-512-t2i-sft)| 512 | Yes | No | Fine-tuned on the [Alchemist dataset](https://huggingface.co/datasets/yandex/alchemist) dataset with Flux VAE | Can handle less detailed prompts in natural language|28 steps, cfg=5.0| `torch.bfloat16` |
| [`Photoroom/prx-512-t2i-sft-distilled`](https://huggingface.co/Photoroom/prx-512-t2i-sft-distilled)| 512 | Yes | Yes | 8-step distilled model from [`Photoroom/prx-512-t2i-sft`](https://huggingface.co/Photoroom/prx-512-t2i-sft) | Can handle less detailed prompts in natural language|8 steps, cfg=1.0| `torch.bfloat16` |
| [`Photoroom/prx-512-t2i-dc-ae`](https://huggingface.co/Photoroom/prx-512-t2i-dc-ae)| 512 | No | No | Base model pre-trained at 512 with [Deep Compression Autoencoder (DC-AE)](https://hanlab.mit.edu/projects/dc-ae)|Works best with detailed prompts in natural language|28 steps, cfg=5.0| `torch.bfloat16` |
| [`Photoroom/prx-512-t2i-dc-ae-sft`](https://huggingface.co/Photoroom/prx-512-t2i-dc-ae-sft)| 512 | Yes | No | Fine-tuned on the [Alchemist dataset](https://huggingface.co/datasets/yandex/alchemist) dataset with [Deep Compression Autoencoder (DC-AE)](https://hanlab.mit.edu/projects/dc-ae) | Can handle less detailed prompts in natural language|28 steps, cfg=5.0| `torch.bfloat16` |
| [`Photoroom/prx-512-t2i-dc-ae-sft-distilled`](https://huggingface.co/Photoroom/prx-512-t2i-dc-ae-sft-distilled)| 512 | Yes | Yes | 8-step distilled model from [`Photoroom/prx-512-t2i-dc-ae-sft-distilled`](https://huggingface.co/Photoroom/prx-512-t2i-dc-ae-sft-distilled) | Can handle less detailed prompts in natural language|8 steps, cfg=1.0| `torch.bfloat16` |s
Refer to [this](https://huggingface.co/collections/Photoroom/prx-models-68e66254c202ebfab99ad38e) collection for more information.
## Loading the pipeline
Load the pipeline with [`~DiffusionPipeline.from_pretrained`].
```py
fromdiffusers.pipelines.prximportPRXPipeline
# Load pipeline - VAE and text encoder will be loaded from HuggingFace
Qwen-Image from the Qwen team is an image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. Experiments show strong general capabilities in both image generation and editing, with exceptional performance in text rendering, especially for Chinese.
prompt="a tiny astronaut hatching from an egg on the moon, Ultra HD, 4K, cinematic composition."
negative_prompt=" "
image=pipe(
prompt=prompt,
negative_prompt=negative_prompt,
width=1024,
height=1024,
num_inference_steps=8,
true_cfg_scale=1.0,
generator=torch.manual_seed(0),
).images[0]
image.save("qwen_fewsteps.png")
```
</details>
> [!TIP]
> The `guidance_scale` parameter in the pipeline is there to support future guidance-distilled models when they come up. Note that passing `guidance_scale` to the pipeline is ineffective. To enable classifier-free guidance, please pass `true_cfg_scale` and `negative_prompt` (even an empty negative prompt like " ") should enable classifier-free guidance computations.
## Multi-image reference with QwenImageEditPlusPipeline
With [`QwenImageEditPlusPipeline`], one can provide multiple images as input reference.
### Batched Inference with Variable-Length Prompts
When using classifier-free guidance (CFG) with prompts of different lengths, the pipeline properly handles padding through attention masking. This ensures padding tokens do not influence the generated output.
```python
# CFG with different prompt lengths works correctly
image=pipe(
prompt="A cat",
negative_prompt="blurry, low quality, distorted",
true_cfg_scale=3.5,
num_inference_steps=50,
).images[0]
```
For detailed benchmark scripts and results, see [this gist](https://gist.github.com/cdutr/bea337e4680268168550292d7819dc2f).
[SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
The abstract from the paper is:
*We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096×4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8×, we trained an AE that can compress images 32×, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024×1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released.*
> [!TIP]
> Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj) and [chenjy2003](https://github.com/chenjy2003). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model).
Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-673efba2a57ed99843f11f9e) collection for more information.
Note: The recommended dtype mentioned is for the transformer weights. The text encoder and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype.
> [!TIP]
> Make sure to pass the `variant` argument for downloaded checkpoints to use lower disk space. Set it to `"fp16"` for models with recommended dtype as `torch.float16`, and `"bf16"` for models with recommended dtype as `torch.bfloat16`. By default, `torch.float32` weights are downloaded, which use twice the amount of disk storage. Additionally, `torch.float32` weights can be downcasted on-the-fly by specifying the `torch_dtype` argument. Read about it in the [docs](https://huggingface.co/docs/diffusers/v0.31.0/en/api/pipelines/overview#diffusers.DiffusionPipeline.from_pretrained).
## Quantization
Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaPipeline`] for inference with bitsandbytes.
[SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation](https://huggingface.co/papers/2503.09641) from NVIDIA, MIT HAN Lab, and Hugging Face by Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Enze Xie, Song Han
The abstract from the paper is:
*This paper presents SANA-Sprint, an efficient diffusion model for ultra-fast text-to-image (T2I) generation. SANA-Sprint is built on a pre-trained foundation model and augmented with hybrid distillation, dramatically reducing inference steps from 20 to 1-4. We introduce three key innovations: (1) We propose a training-free approach that transforms a pre-trained flow-matching model for continuous-time consistency distillation (sCM), eliminating costly training from scratch and achieving high training efficiency. Our hybrid distillation strategy combines sCM with latent adversarial distillation (LADD): sCM ensures alignment with the teacher model, while LADD enhances single-step generation fidelity. (2) SANA-Sprint is a unified step-adaptive model that achieves high-quality generation in 1-4 steps, eliminating step-specific training and improving efficiency. (3) We integrate ControlNet with SANA-Sprint for real-time interactive image generation, enabling instant visual feedback for user interaction. SANA-Sprint establishes a new Pareto frontier in speed-quality tradeoffs, achieving state-of-the-art performance with 7.59 FID and 0.74 GenEval in only 1 step — outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10× faster (0.1s vs 1.1s on H100). It also achieves 0.1s (T2I) and 0.25s (ControlNet) latency for 1024×1024 images on H100, and 0.31s (T2I) on an RTX 4090, showcasing its exceptional efficiency and potential for AI-powered consumer applications (AIPC). Code and pre-trained models will be open-sourced.*
This pipeline was contributed by [lawrence-cj](https://github.com/lawrence-cj), [shuchen Xue](https://github.com/scxue) and [Enze Xie](https://github.com/xieenze). The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://huggingface.co/Efficient-Large-Model/).
Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-sprint-67d6810d65235085b3b17c76) collection for more information.
Note: The recommended dtype mentioned is for the transformer weights. The text encoder must stay in `torch.bfloat16` and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype.
## Quantization
Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaSprintPipeline`] for inference with bitsandbytes.
prompt="a tiny astronaut hatching from an egg on the moon"
image=pipeline(prompt).images[0]
image.save("sana.png")
```
## Setting `max_timesteps`
Users can tweak the `max_timesteps` value for experimenting with the visual quality of the generated outputs. The default `max_timesteps` value was obtained with an inference-time search process. For more details about it, check out the paper.
## Image to Image
The [`SanaSprintImg2ImgPipeline`] is a pipeline for image-to-image generation. It takes an input image and a prompt, and generates a new image based on the input image and the prompt.
[SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer](https://huggingface.co/papers/2509.24695) from NVIDIA and MIT HAN Lab, by Junsong Chen, Yuyang Zhao, Jincheng Yu, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie.
The abstract from the paper is:
*We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation. [this https URL](https://github.com/NVlabs/SANA).*
This pipeline was contributed by SANA Team. The original codebase can be found [here](https://github.com/NVlabs/Sana). The original weights can be found under [hf.co/Efficient-Large-Model](https://hf.co/collections/Efficient-Large-Model/sana-video).
Refer to [this](https://huggingface.co/collections/Efficient-Large-Model/sana-video) collection for more information.
Note: The recommended dtype mentioned is for the transformer weights. The text encoder and VAE weights must stay in `torch.bfloat16` or `torch.float32` for the model to work correctly. Please refer to the inference example below to see how to load the model with the recommended dtype.
## Generation Pipelines
<hfoptionsid="generation pipelines">`
<hfoption id="Text-to-Video">
The example below demonstrates how to use the text-to-video pipeline to generate a video using a text description.
prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
prompt = "A woman stands against a stunning sunset backdrop, her long, wavy brown hair gently blowing in the breeze. She wears a sleeveless, light-colored blouse with a deep V-neckline, which accentuates her graceful posture. The warm hues of the setting sun cast a golden glow across her face and hair, creating a serene and ethereal atmosphere. The background features a blurred landscape with soft, rolling hills and scattered clouds, adding depth to the scene. The camera remains steady, capturing the tranquil moment from a medium close-up angle."
negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`SanaVideoPipeline`] for inference with bitsandbytes.
```py
import torch
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, SanaVideoTransformer3DModel, SanaVideoPipeline
from transformers import BitsAndBytesConfig as BitsAndBytesConfig, AutoModel
prompt = "Evening, backlight, side lighting, soft light, high contrast, mid-shot, centered composition, clean solo shot, warm color. A young Caucasian man stands in a forest, golden light glimmers on his hair as sunlight filters through the leaves. He wears a light shirt, wind gently blowing his hair and collar, light dances across his face with his movements. The background is blurred, with dappled light and soft tree shadows in the distance. The camera focuses on his lifted gaze, clear and emotional."
negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience."
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Shap-E
The Shap-E model was proposed in [Shap-E: Generating Conditional 3D Implicit Functions](https://huggingface.co/papers/2305.02463) by Alex Nichol and Heewoo Jun from [OpenAI](https://github.com/openai).
The abstract from the paper is:
*We present Shap-E, a conditional generative model for 3D assets. Unlike recent work on 3D generative models which produce a single output representation, Shap-E directly generates the parameters of implicit functions that can be rendered as both textured meshes and neural radiance fields. We train Shap-E in two stages: first, we train an encoder that deterministically maps 3D assets into the parameters of an implicit function; second, we train a conditional diffusion model on outputs of the encoder. When trained on a large dataset of paired 3D and text data, our resulting models are capable of generating complex and diverse 3D assets in a matter of seconds. When compared to Point-E, an explicit generative model over point clouds, Shap-E converges faster and reaches comparable or better sample quality despite modeling a higher-dimensional, multi-representation output space.*
The original codebase can be found at [openai/shap-e](https://github.com/openai/shap-e).
> [!TIP]
> See the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
# SkyReels-V2: Infinite-length Film Generative model
[SkyReels-V2](https://huggingface.co/papers/2504.13074) by the SkyReels Team from Skywork AI.
*Recent advances in video generation have been driven by diffusion models and autoregressive frameworks, yet critical challenges persist in harmonizing prompt adherence, visual quality, motion dynamics, and duration: compromises in motion dynamics to enhance temporal visual quality, constrained video duration (5-10 seconds) to prioritize resolution, and inadequate shot-aware generation stemming from general-purpose MLLMs' inability to interpret cinematic grammar, such as shot composition, actor expressions, and camera motions. These intertwined limitations hinder realistic long-form synthesis and professional film-style generation. To address these limitations, we propose SkyReels-V2, an Infinite-length Film Generative Model, that synergizes Multi-modal Large Language Model (MLLM), Multi-stage Pretraining, Reinforcement Learning, and Diffusion Forcing Framework. Firstly, we design a comprehensive structural representation of video that combines the general descriptions by the Multi-modal LLM and the detailed shot language by sub-expert models. Aided with human annotation, we then train a unified Video Captioner, named SkyCaptioner-V1, to efficiently label the video data. Secondly, we establish progressive-resolution pretraining for the fundamental video generation, followed by a four-stage post-training enhancement: Initial concept-balanced Supervised Fine-Tuning (SFT) improves baseline quality; Motion-specific Reinforcement Learning (RL) training with human-annotated and synthetic distortion data addresses dynamic artifacts; Our diffusion forcing framework with non-decreasing noise schedules enables long-video synthesis in an efficient search space; Final high-quality SFT refines visual fidelity. All the code and models are available at [this https URL](https://github.com/SkyworkAI/SkyReels-V2).*
You can find all the original SkyReels-V2 checkpoints under the [Skywork](https://huggingface.co/collections/Skywork/skyreels-v2-6801b1b93df627d441d0d0d9) organization.
The following SkyReels-V2 models are supported in Diffusers:
Key Pattern: Block `i` lags behind Block `i-1` by exactly `ar_step=5` timesteps, creating the
staggered "diffusion forcing" effect where later blocks condition on cleaner earlier blocks.
### Text-to-Video Generation
The example below demonstrates how to generate a video from text.
<hfoptionsid="T2V usage">
<hfoptionid="T2V memory">
Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
From the original repo:
>You can use --ar_step 5 to enable asynchronous inference. When asynchronous inference, --causal_block_size 5 is recommended while it is not supposed to be set for synchronous generation... Asynchronous inference will take more steps to diffuse the whole sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous inference may improve the instruction following and visual consistent performance.
prompt="A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."
output=pipeline(
prompt=prompt,
num_inference_steps=30,
height=544,# 720 for 720P
width=960,# 1280 for 720P
num_frames=97,
base_num_frames=97,# 121 for 720P
ar_step=5,# Controls asynchronous inference (0 for synchronous mode)
causal_block_size=5,# Number of frames in each block for asynchronous processing
overlap_history=None,# Number of frames to overlap for smooth transitions in long videos; 17 for long video generations
addnoise_condition=20,# Improves consistency in long video generation
The example below demonstrates how to use the image-to-video pipeline to generate a video using a text description, a starting frame, and an ending frame.
prompt="CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
prompt="CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Stable Audio
Stable Audio was proposed in [Stable Audio Open](https://huggingface.co/papers/2407.14358) by Zach Evans et al. . it takes a text prompt as input and predicts the corresponding sound or music sample.
Stable Audio Open generates variable-length (up to 47s) stereo audio at 44.1kHz from text prompts. It comprises three components: an autoencoder that compresses waveforms into a manageable sequence length, a T5-based text embedding for text conditioning, and a transformer-based diffusion (DiT) model that operates in the latent space of the autoencoder.
Stable Audio is trained on a corpus of around 48k audio recordings, where around 47k are from Freesound and the rest are from the Free Music Archive (FMA). All audio files are licensed under CC0, CC BY, or CC Sampling+. This data is used to train the autoencoder and the DiT.
The abstract of the paper is the following:
*Open generative models are vitally important for the community, allowing for fine-tunes and serving as baselines when presenting new models. However, most current text-to-audio models are private and not accessible for artists and researchers to build upon. Here we describe the architecture and training process of a new open-weights text-to-audio model trained with Creative Commons data. Our evaluation shows that the model's performance is competitive with the state-of-the-art across various metrics. Notably, the reported FDopenl3 results (measuring the realism of the generations) showcase its potential for high-quality stereo sound synthesis at 44.1kHz.*
This pipeline was contributed by [Yoach Lacombe](https://huggingface.co/ylacombe). The original codebase can be found at [Stability-AI/stable-audio-tools](https://github.com/Stability-AI/stable-audio-tools).
## Tips
When constructing a prompt, keep in mind:
* Descriptive prompt inputs work best; use adjectives to describe the sound (for example, "high quality" or "clear") and make the prompt context specific where possible (e.g. "melodic techno with a fast beat and synths" works better than "techno").
* Using a *negative prompt* can significantly improve the quality of the generated audio. Try using a negative prompt of "low quality, average quality".
During inference:
* The _quality_ of the generated audio sample can be controlled by the `num_inference_steps` argument; higher steps give higher quality audio at the expense of slower inference.
* Multiple waveforms can be generated in one go: set `num_waveforms_per_prompt` to a value greater than 1 to enable. Automatic scoring will be performed between the generated waveforms and prompt text, and the audios ranked from best to worst accordingly.
## Quantization
Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`StableAudioPipeline`] for inference with bitsandbytes.
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Stable Cascade
This model is built upon the [Würstchen](https://openreview.net/forum?id=gU58d5QeGv) architecture and its main
difference to other models like Stable Diffusion is that it is working at a much smaller latent space. Why is this
important? The smaller the latent space, the **faster** you can run inference and the **cheaper** the training becomes.
How small is the latent space? Stable Diffusion uses a compression factor of 8, resulting in a 1024x1024 image being
encoded to 128x128. Stable Cascade achieves a compression factor of 42, meaning that it is possible to encode a
1024x1024 image to 24x24, while maintaining crisp reconstructions. The text-conditional model is then trained in the
highly compressed latent space. Previous versions of this architecture, achieved a 16x cost reduction over Stable
Diffusion 1.5.
Therefore, this kind of model is well suited for usages where efficiency is important. Furthermore, all known extensions
like finetuning, LoRA, ControlNet, IP-Adapter, LCM etc. are possible with this method as well.
The original codebase can be found at [Stability-AI/StableCascade](https://github.com/Stability-AI/StableCascade).
## Model Overview
Stable Cascade consists of three models: Stage A, Stage B and Stage C, representing a cascade to generate images,
hence the name "Stable Cascade".
Stage A & B are used to compress images, similar to what the job of the VAE is in Stable Diffusion.
However, with this setup, a much higher compression of images can be achieved. While the Stable Diffusion models use a
spatial compression factor of 8, encoding an image with resolution of 1024 x 1024 to 128 x 128, Stable Cascade achieves
a compression factor of 42. This encodes a 1024 x 1024 image to 24 x 24, while being able to accurately decode the
image. This comes with the great benefit of cheaper training and inference. Furthermore, Stage C is responsible
for generating the small 24 x 24 latents given a text prompt.
The Stage C model operates on the small 24 x 24 latents and denoises the latents conditioned on text prompts. The model is also the largest component in the Cascade pipeline and is meant to be used with the `StableCascadePriorPipeline`
The Stage B and Stage A models are used with the `StableCascadeDecoderPipeline` and are responsible for generating the final image given the small 24 x 24 latents.
> [!WARNING]
> There are some restrictions on data types that can be used with the Stable Cascade models. The official checkpoints for the `StableCascadePriorPipeline` do not support the `torch.float16` data type. Please use `torch.bfloat16` instead.
>
> In order to use the `torch.bfloat16` data type with the `StableCascadeDecoderPipeline` you need to have PyTorch 2.2.0 or higher installed. This also means that using the `StableCascadeCombinedPipeline` with `torch.bfloat16` requires PyTorch 2.2.0 or higher, since it calls the `StableCascadeDecoderPipeline` internally.
>
> If it is not possible to install PyTorch 2.2.0 or higher in your environment, the `StableCascadeDecoderPipeline` can be used on its own with the `torch.float16` data type. You can download the full precision or `bf16` variant weights for the pipeline and cast the weights to `torch.float16`.
