[Helios: Real Real-Time Long Video Generation Model](https://huggingface.co/papers/2603.04379) from Peking University & ByteDance & etc, by Shenghai Yuan, Yuanyang Yin, Zongjian Li, Xinwei Huang, Xiao Yang, Li Yuan.
*<u>We introduce Helios, the first 14B video generation model that runs at 17 FPS on a single NVIDIA H100 GPU and supports minute-scale generation while matching a strong baseline in quality.</u> We make breakthroughs along three key dimensions: (1) robustness to long-video drifting without commonly used anti-drift heuristics such as self-forcing, error banks, or keyframe sampling; (2) real-time generation without standard acceleration techniques such as KV-cache, causal masking, or sparse attention; and (3) training without parallelism or sharding frameworks, enabling image-diffusion-scale batch sizes while fitting up to four 14B models within 80 GB of GPU memory. Specifically, Helios is a 14B autoregressive diffusion model with a unified input representation that natively supports T2V, I2V, and V2V tasks. To mitigate drifting in long-video generation, we characterize its typical failure modes and propose simple yet effective training strategies that explicitly simulate drifting during training, while eliminating repetitive motion at its source. For efficiency, we heavily compress the historical and noisy context and reduce the number of sampling steps, yielding computational costs comparable to—or lower than—those of 1.3B video generative models. Moreover, we introduce infrastructure-level optimizations that accelerate both inference and training while reducing memory consumption. Extensive experiments demonstrate that Helios consistently outperforms prior methods on both short- and long-video generation. All the code and models are available at [this https URL](https://pku-yuangroup.github.io/Helios-Page).
The following Helios models are supported in Diffusers:
-[Helios-Base](https://huggingface.co/BestWishYsh/Helios-Base): Best Quality, with v-prediction, standard CFG and custom HeliosScheduler.
-[Helios-Mid](https://huggingface.co/BestWishYsh/Helios-Mid): Intermediate Weight, with v-prediction, CFG-Zero* and custom HeliosScheduler.
-[Helios-Distilled](https://huggingface.co/BestWishYsh/Helios-Distilled): Best Efficiency, with x0-prediction and custom HeliosDMDScheduler.
> [!TIP]
> Click on the Helios models in the right sidebar for more examples of video generation.
### Optimizing Memory and Inference Speed
The example below demonstrates how to generate a video from text optimized for memory or inference speed.
<hfoptionsid="optimization">
<hfoptionid="memory">
Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
[Compilation](../../optimization/fp16#torchcompile) is slow the first time but subsequent calls to the pipeline are faster. [Attention Backends](../../optimization/attention_backends) such as FlashAttention and SageAttention can significantly increase speed by optimizing the computation of the attention mechanism. [Context Parallelism](../../training/distributed_inference#context-parallelism) splits the input sequence across multiple devices to enable processing of long contexts in parallel, reducing memory pressure and latency. [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
[HunyuanVideo](https://huggingface.co/papers/2412.03603) is a 13B parameter diffusion transformer model designed to be competitive with closed-source video foundation models and enable wider community access. This model uses a "dual-stream to single-stream" architecture to separately process the video and text tokens first, before concatenating and feeding them to the transformer to fuse the multimodal information. A pretrained multimodal large language model (MLLM) is used as the encoder because it has better image-text alignment, better image detail description and reasoning, and it can be used as a zero-shot learner if system instructions are added to user prompts. Finally, HunyuanVideo uses a 3D causal variational autoencoder to more efficiently process video data at the original resolution and frame rate.
You can find all the original HunyuanVideo checkpoints under the [Tencent](https://huggingface.co/tencent) organization.
> [!TIP]
> Click on the HunyuanVideo models in the right sidebar for more examples of video generation tasks.
>
> The examples below use a checkpoint from [hunyuanvideo-community](https://huggingface.co/hunyuanvideo-community) because the weights are stored in a layout compatible with Diffusers.
The example below demonstrates how to generate a video optimized for memory or inference speed.
<hfoptionsid="usage">
<hfoptionid="memory">
Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
The quantized HunyuanVideo model below requires ~14GB of VRAM.
<!-- 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. -->
# HunyuanVideo-1.5
HunyuanVideo-1.5 is a lightweight yet powerful video generation model that achieves state-of-the-art visual quality and motion coherence with only 8.3 billion parameters, enabling efficient inference on consumer-grade GPUs. This achievement is built upon several key components, including meticulous data curation, an advanced DiT architecture with selective and sliding tile attention (SSTA), enhanced bilingual understanding through glyph-aware text encoding, progressive pre-training and post-training, and an efficient video super-resolution network. Leveraging these designs, we developed a unified framework capable of high-quality text-to-video and image-to-video generation across multiple durations and resolutions. Extensive experiments demonstrate that this compact and proficient model establishes a new state-of-the-art among open-source models.
You can find all the original HunyuanVideo checkpoints under the [Tencent](https://huggingface.co/tencent) organization.
> [!TIP]
> Click on the HunyuanVideo models in the right sidebar for more examples of video generation tasks.
>
> The examples below use a checkpoint from [hunyuanvideo-community](https://huggingface.co/hunyuanvideo-community) because the weights are stored in a layout compatible with Diffusers.
The example below demonstrates how to generate a video optimized for memory or inference speed.
<hfoptionsid="usage">
<hfoptionid="memory">
Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
- HunyuanVideo1.5 use attention masks with variable-length sequences. For best performance, we recommend using an attention backend that handles padding efficiently.
- **H100/H800:** `_flash_3_hub` or `_flash_3_varlen_hub`
- **A100/A800/RTX 4090:** `flash_hub` or `flash_varlen_hub`
- **Other GPUs:** `sage_hub`
Refer to the [Attention backends](../../optimization/attention_backends) guide for more details about using a different backend.
```py
pipe.transformer.set_attention_backend("flash_hub")# or your preferred backend
```
- [`HunyuanVideo15Pipeline`] use guider and does not take `guidance_scale` parameter at runtime.
You can check the default guider configuration using `pipe.guider`:
```py
>>>pipe.guider
ClassifierFreeGuidance{
"_class_name":"ClassifierFreeGuidance",
"_diffusers_version":"0.36.0.dev0",
"enabled":true,
"guidance_rescale":0.0,
"guidance_scale":6.0,
"start":0.0,
"stop":1.0,
"use_original_formulation":false
}
State:
step:None
num_inference_steps:None
timestep:None
count_prepared:0
enabled:True
num_conditions:2
```
To update guider configuration, you can run `pipe.guider = pipe.guider.new(...)`
```py
pipe.guider=pipe.guider.new(guidance_scale=5.0)
```
Read more on Guider [here](../../using-diffusers/guiders).
<!--Copyright 2025 The HuggingFace Team and Tencent Hunyuan 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.
-->
# Hunyuan-DiT
[Hunyuan-DiT : A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding](https://huggingface.co/papers/2405.08748) from Tencent Hunyuan.
The abstract from the paper is:
*We present Hunyuan-DiT, a text-to-image diffusion transformer with fine-grained understanding of both English and Chinese. To construct Hunyuan-DiT, we carefully design the transformer structure, text encoder, and positional encoding. We also build from scratch a whole data pipeline to update and evaluate data for iterative model optimization. For fine-grained language understanding, we train a Multimodal Large Language Model to refine the captions of the images. Finally, Hunyuan-DiT can perform multi-turn multimodal dialogue with users, generating and refining images according to the context. Through our holistic human evaluation protocol with more than 50 professional human evaluators, Hunyuan-DiT sets a new state-of-the-art in Chinese-to-image generation compared with other open-source models.*
You can find the original codebase at [Tencent/HunyuanDiT](https://github.com/Tencent/HunyuanDiT) and all the available checkpoints at [Tencent-Hunyuan](https://huggingface.co/Tencent-Hunyuan/HunyuanDiT).
* It combines two text encoders, a bilingual CLIP and a multilingual T5 encoder
> [!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 [`HungyuanDiTPipeline`] to the [SDXL refiner](../../using-diffusers/sdxl#base-to-refiner-model) model.
## Optimization
You can optimize the pipeline's runtime and memory consumption with torch.compile and feed-forward chunking. To learn about other optimization methods, check out the [Speed up inference](../../optimization/fp16) and [Reduce memory usage](../../optimization/memory) guides.
### Inference
Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
The [benchmark](https://gist.github.com/sayakpaul/29d3a14905cfcbf611fe71ebd22e9b23) results on a 80GB A100 machine are:
```bash
With torch.compile(): Average inference time: 12.470 seconds.
Without torch.compile(): Average inference time: 20.570 seconds.
```
### Memory optimization
By loading the T5 text encoder in 8 bits, you can run the pipeline in just under 6 GBs of GPU VRAM. Refer to [this script](https://gist.github.com/sayakpaul/3154605f6af05b98a41081aaba5ca43e) for details.
Furthermore, you can use the [`~HunyuanDiT2DModel.enable_forward_chunking`] method to reduce memory usage. Feed-forward chunking runs the feed-forward layers in a transformer block in a loop instead of all at once. This gives you a trade-off between memory consumption and inference runtime.
> [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
## HunyuanImage-2.1
HunyuanImage-2.1 applies [Adaptive Projected Guidance (APG)](https://huggingface.co/papers/2410.02416) combined with Classifier-Free Guidance (CFG) in the denoising loop. `HunyuanImagePipeline` has a `guider` component (read more about [Guider](../../using-diffusers/guiders)) and does not take a `guidance_scale` parameter at runtime. To change guider-related parameters, e.g., `guidance_scale`, you can update the `guider` configuration instead.
To update the guider with a different configuration, use the `new()` method. For example, to generate an image with `guidance_scale=5.0` while keeping all other default guidance parameters:
<!--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.
-->
# Kandinsky 2.1
Kandinsky 2.1 is created by [Arseniy Shakhmatov](https://github.com/cene555), [Anton Razzhigaev](https://github.com/razzant), [Aleksandr Nikolich](https://github.com/AlexWortega), [Vladimir Arkhipkin](https://github.com/oriBetelgeuse), [Igor Pavlov](https://github.com/boomb0om), [Andrey Kuznetsov](https://github.com/kuznetsoffandrey), and [Denis Dimitrov](https://github.com/denndimitrov).
The description from it's GitHub page is:
*Kandinsky 2.1 inherits best practicies from Dall-E 2 and Latent diffusion, while introducing some new ideas. As text and image encoder it uses CLIP model and diffusion image prior (mapping) between latent spaces of CLIP modalities. This approach increases the visual performance of the model and unveils new horizons in blending images and text-guided image manipulation.*
The original codebase can be found at [ai-forever/Kandinsky-2](https://github.com/ai-forever/Kandinsky-2).
> [!TIP]
> Check out the [Kandinsky Community](https://huggingface.co/kandinsky-community) organization on the Hub for the official model checkpoints for tasks like text-to-image, image-to-image, and inpainting.
> [!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.
Kandinsky 3 is created by [Vladimir Arkhipkin](https://github.com/oriBetelgeuse),[Anastasia Maltseva](https://github.com/NastyaMittseva),[Igor Pavlov](https://github.com/boomb0om),[Andrei Filatov](https://github.com/anvilarth),[Arseniy Shakhmatov](https://github.com/cene555),[Andrey Kuznetsov](https://github.com/kuznetsoffandrey),[Denis Dimitrov](https://github.com/denndimitrov), [Zein Shaheen](https://github.com/zeinsh)
The description from it's GitHub page:
*Kandinsky 3.0 is an open-source text-to-image diffusion model built upon the Kandinsky2-x model family. In comparison to its predecessors, enhancements have been made to the text understanding and visual quality of the model, achieved by increasing the size of the text encoder and Diffusion U-Net models, respectively.*
Its architecture includes 3 main components:
1.[FLAN-UL2](https://huggingface.co/google/flan-ul2), which is an encoder decoder model based on the T5 architecture.
2. New U-Net architecture featuring BigGAN-deep blocks doubles depth while maintaining the same number of parameters.
3. Sber-MoVQGAN is a decoder proven to have superior results in image restoration.
The original codebase can be found at [ai-forever/Kandinsky-3](https://github.com/ai-forever/Kandinsky-3).
> [!TIP]
> Check out the [Kandinsky Community](https://huggingface.co/kandinsky-community) organization on the Hub for the official model checkpoints for tasks like text-to-image, image-to-image, and inpainting.
> [!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.
<!--Copyright 2025 The HuggingFace Team and Kandinsky Lab 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.
-->
# Kandinsky 5.0 Image
[Kandinsky 5.0](https://arxiv.org/abs/2511.14993) is a family of diffusion models for Video & Image generation.
Kandinsky 5.0 Image Lite is a lightweight image generation model (6B parameters).
The model introduces several key innovations:
-**Latent diffusion pipeline** with **Flow Matching** for improved training stability
-**Diffusion Transformer (DiT)** as the main generative backbone with cross-attention to text embeddings
- Dual text encoding using **Qwen2.5-VL** and **CLIP** for comprehensive text understanding
-**Flux VAE** for efficient image encoding and decoding
The original codebase can be found at [kandinskylab/Kandinsky-5](https://github.com/kandinskylab/Kandinsky-5).
> [!TIP]
> Check out the [Kandinsky Lab](https://huggingface.co/kandinskylab) organization on the Hub for the official model checkpoints for text-to-video generation, including pretrained, SFT, no-CFG, and distilled variants.
| [**kandinskylab/Kandinsky-5.0-T2I-Lite-pretrain-Diffusers**](https://huggingface.co/kandinskylab/Kandinsky-5.0-T2I-Lite-pretrain-Diffusers) | 6B image Base pretrained model | Research and fine-tuning |
| [**kandinskylab/Kandinsky-5.0-I2I-Lite-pretrain-Diffusers**](https://huggingface.co/kandinskylab/Kandinsky-5.0-I2I-Lite-pretrain-Diffusers) | 6B image editing Base pretrained model | Research and fine-tuning |
prompt="A fluffy, expressive cat wearing a bright red hat with a soft, slightly textured fabric. The hat should look cozy and well-fitted on the cat’s head. On the front of the hat, add clean, bold white text that reads “SWEET”, clearly visible and neatly centered. Ensure the overall lighting highlights the hat’s color and the cat’s fur details."
prompt="Change the background from a winter night scene to a bright summer day. Place the character on a sandy beach with clear blue sky, soft sunlight, and gentle waves in the distance. Replace the winter clothing with a light short-sleeved T-shirt (in soft pastel colors) and casual shorts. Ensure the character’s fur reflects warm daylight instead of cold winter tones. Add small beach details such as seashells, footprints in the sand, and a few scattered beach toys nearby. Keep the oranges in the scene, but place them naturally on the sand."
negative_prompt=""
output=pipe(
image=image,
prompt=prompt,
negative_prompt=negative_prompt,
guidance_scale=3.5,
).image[0]
```
## Kandinsky5T2IPipeline
[[autodoc]] Kandinsky5T2IPipeline
- all
- __call__
## Kandinsky5I2IPipeline
[[autodoc]] Kandinsky5I2IPipeline
- all
- __call__
## Citation
```bibtex
@misc{kandinsky2025,
author={Alexander Belykh and Alexander Varlamov and Alexey Letunovskiy and Anastasia Aliaskina and Anastasia Maltseva and Anastasiia Kargapoltseva and Andrey Shutkin and Anna Averchenkova and Anna Dmitrienko and Bulat Akhmatov and Denis Dimitrov and Denis Koposov and Denis Parkhomenko and Dmitrii and Ilya Vasiliev and Ivan Kirillov and Julia Agafonova and Kirill Chernyshev and Kormilitsyn Semen and Lev Novitskiy and Maria Kovaleva and Mikhail Mamaev and Mikhailov and Nikita Kiselev and Nikita Osterov and Nikolai Gerasimenko and Nikolai Vaulin and Olga Kim and Olga Vdovchenko and Polina Gavrilova and Polina Mikhailova and Tatiana Nikulina and Viacheslav Vasilev and Vladimir Arkhipkin and Vladimir Korviakov and Vladimir Polovnikov and Yury Kolabushin},
title={Kandinsky 5.0: A family of diffusion models for Video & Image generation},
<!--Copyright 2025 The HuggingFace Team Kandinsky Lab 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.
-->
# Kandinsky 5.0 Video
[Kandinsky 5.0](https://arxiv.org/abs/2511.14993) is a family of diffusion models for Video & Image generation.
Kandinsky 5.0 Lite line-up of lightweight video generation models (2B parameters) that ranks #1 among open-source models in its class. It outperforms larger models and offers the best understanding of Russian concepts in the open-source ecosystem.
Kandinsky 5.0 Pro line-up of large high quality video generation models (19B parameters). It offers high qualty generation in HD and more generation formats like I2V.
The model introduces several key innovations:
-**Latent diffusion pipeline** with **Flow Matching** for improved training stability
-**Diffusion Transformer (DiT)** as the main generative backbone with cross-attention to text embeddings
- Dual text encoding using **Qwen2.5-VL** and **CLIP** for comprehensive text understanding
-**HunyuanVideo 3D VAE** for efficient video encoding and decoding
-**Sparse attention mechanisms** (NABLA) for efficient long-sequence processing
The original codebase can be found at [kandinskylab/Kandinsky-5](https://github.com/kandinskylab/Kandinsky-5).
> [!TIP]
> Check out the [Kandinsky Lab](https://huggingface.co/kandinskylab) organization on the Hub for the official model checkpoints for text-to-video generation, including pretrained, SFT, no-CFG, and distilled variants.
## Available Models
Kandinsky 5.0 T2V Pro:
| model_id | Description | Use Cases |
|------------|-------------|-----------|
| **kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers** | 5 second Text-to-Video Pro model | High-quality text-to-video generation |
| **kandinskylab/Kandinsky-5.0-I2V-Pro-sft-5s-Diffusers** | 5 second Image-to-Video Pro model | High-quality image-to-video generation |
Kandinsky 5.0 T2V Lite:
| model_id | Description | Use Cases |
|------------|-------------|-----------|
| **kandinskylab/Kandinsky-5.0-T2V-Lite-sft-5s-Diffusers** | 5 second Supervised Fine-Tuned model | Highest generation quality |
| **kandinskylab/Kandinsky-5.0-T2V-Lite-sft-10s-Diffusers** | 10 second Supervised Fine-Tuned model | Highest generation quality |
The evaluation is based on the expanded prompts from the [Movie Gen benchmark](https://github.com/facebookresearch/MovieGenBench), which are available in the expanded_prompt column of the benchmark/moviegen_bench.csv file.
author={Alexander Belykh and Alexander Varlamov and Alexey Letunovskiy and Anastasia Aliaskina and Anastasia Maltseva and Anastasiia Kargapoltseva and Andrey Shutkin and Anna Averchenkova and Anna Dmitrienko and Bulat Akhmatov and Denis Dimitrov and Denis Koposov and Denis Parkhomenko and Dmitrii and Ilya Vasiliev and Ivan Kirillov and Julia Agafonova and Kirill Chernyshev and Kormilitsyn Semen and Lev Novitskiy and Maria Kovaleva and Mikhail Mamaev and Mikhailov and Nikita Kiselev and Nikita Osterov and Nikolai Gerasimenko and Nikolai Vaulin and Olga Kim and Olga Vdovchenko and Polina Gavrilova and Polina Mikhailova and Tatiana Nikulina and Viacheslav Vasilev and Vladimir Arkhipkin and Vladimir Korviakov and Vladimir Polovnikov and Yury Kolabushin},
title={Kandinsky 5.0: A family of diffusion models for Video & Image generation},
<!--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.
-->
# Kandinsky 2.2
Kandinsky 2.2 is created by [Arseniy Shakhmatov](https://github.com/cene555), [Anton Razzhigaev](https://github.com/razzant), [Aleksandr Nikolich](https://github.com/AlexWortega), [Vladimir Arkhipkin](https://github.com/oriBetelgeuse), [Igor Pavlov](https://github.com/boomb0om), [Andrey Kuznetsov](https://github.com/kuznetsoffandrey), and [Denis Dimitrov](https://github.com/denndimitrov).
The description from it's GitHub page is:
*Kandinsky 2.2 brings substantial improvements upon its predecessor, Kandinsky 2.1, by introducing a new, more powerful image encoder - CLIP-ViT-G and the ControlNet support. The switch to CLIP-ViT-G as the image encoder significantly increases the model's capability to generate more aesthetic pictures and better understand text, thus enhancing the model's overall performance. The addition of the ControlNet mechanism allows the model to effectively control the process of generating images. This leads to more accurate and visually appealing outputs and opens new possibilities for text-guided image manipulation.*
The original codebase can be found at [ai-forever/Kandinsky-2](https://github.com/ai-forever/Kandinsky-2).
> [!TIP]
> Check out the [Kandinsky Community](https://huggingface.co/kandinsky-community) organization on the Hub for the official model checkpoints for tasks like text-to-image, image-to-image, and inpainting.
> [!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.
Kolors is a large-scale text-to-image generation model based on latent diffusion, developed by [the Kuaishou Kolors team](https://github.com/Kwai-Kolors/Kolors). Trained on billions of text-image pairs, Kolors exhibits significant advantages over both open-source and closed-source models in visual quality, complex semantic accuracy, and text rendering for both Chinese and English characters. Furthermore, Kolors supports both Chinese and English inputs, demonstrating strong performance in understanding and generating Chinese-specific content. For more details, please refer to this [technical report](https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf).
The abstract from the technical report is:
*We present Kolors, a latent diffusion model for text-to-image synthesis, characterized by its profound understanding of both English and Chinese, as well as an impressive degree of photorealism. There are three key insights contributing to the development of Kolors. Firstly, unlike large language model T5 used in Imagen and Stable Diffusion 3, Kolors is built upon the General Language Model (GLM), which enhances its comprehension capabilities in both English and Chinese. Moreover, we employ a multimodal large language model to recaption the extensive training dataset for fine-grained text understanding. These strategies significantly improve Kolors’ ability to comprehend intricate semantics, particularly those involving multiple entities, and enable its advanced text rendering capabilities. Secondly, we divide the training of Kolors into two phases: the concept learning phase with broad knowledge and the quality improvement phase with specifically curated high-aesthetic data. Furthermore, we investigate the critical role of the noise schedule and introduce a novel schedule to optimize high-resolution image generation. These strategies collectively enhance the visual appeal of the generated high-resolution images. Lastly, we propose a category-balanced benchmark KolorsPrompts, which serves as a guide for the training and evaluation of Kolors. Consequently, even when employing the commonly used U-Net backbone, Kolors has demonstrated remarkable performance in human evaluations, surpassing the existing open-source models and achieving Midjourney-v6 level performance, especially in terms of visual appeal. We will release the code and weights of Kolors at <https://github.com/Kwai-Kolors/Kolors>, and hope that it will benefit future research and applications in the visual generation community.*
Kolors needs a different IP Adapter to work, and it uses [Openai-CLIP-336](https://huggingface.co/openai/clip-vit-large-patch14-336) as an image encoder.
> [!TIP]
> Using an IP Adapter with Kolors requires more than 24GB of VRAM. To use it, we recommend using [`~DiffusionPipeline.enable_model_cpu_offload`] on consumer GPUs.
> [!TIP]
> While Kolors is integrated in Diffusers, you need to load the image encoder from a revision to use the safetensor files. You can still use the main branch of the original repository if you're comfortable loading pickle checkpoints.
Latent Consistency Models (LCMs) were proposed in [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://huggingface.co/papers/2310.04378) by Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao.
The abstract of the paper is as follows:
*Latent Diffusion models (LDMs) have achieved remarkable results in synthesizing high-resolution images. However, the iterative sampling process is computationally intensive and leads to slow generation. Inspired by Consistency Models (song et al.), we propose Latent Consistency Models (LCMs), enabling swift inference with minimal steps on any pre-trained LDMs, including Stable Diffusion (rombach et al). Viewing the guided reverse diffusion process as solving an augmented probability flow ODE (PF-ODE), LCMs are designed to directly predict the solution of such ODE in latent space, mitigating the need for numerous iterations and allowing rapid, high-fidelity sampling. Efficiently distilled from pre-trained classifier-free guided diffusion models, a high-quality 768 x 768 2~4-step LCM takes only 32 A100 GPU hours for training. Furthermore, we introduce Latent Consistency Fine-tuning (LCF), a novel method that is tailored for fine-tuning LCMs on customized image datasets. Evaluation on the LAION-5B-Aesthetics dataset demonstrates that LCMs achieve state-of-the-art text-to-image generation performance with few-step inference. Project Page: [this https URL](https://latent-consistency-models.github.io/).*
A demo for the [SimianLuo/LCM_Dreamshaper_v7](https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7) checkpoint can be found [here](https://huggingface.co/spaces/SimianLuo/Latent_Consistency_Model).
The pipelines were contributed by [luosiallen](https://luosiallen.github.io/), [nagolinc](https://github.com/nagolinc), and [dg845](https://github.com/dg845).
<!--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.
-->
# Latent Diffusion
Latent Diffusion was proposed in [High-Resolution Image Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer.
The abstract from the paper is:
*By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond. Additionally, their formulation allows for a guiding mechanism to control the image generation process without retraining. However, since these models typically operate directly in pixel space, optimization of powerful DMs often consumes hundreds of GPU days and inference is expensive due to sequential evaluations. To enable DM training on limited computational resources while retaining their quality and flexibility, we apply them in the latent space of powerful pretrained autoencoders. In contrast to previous work, training diffusion models on such a representation allows for the first time to reach a near-optimal point between complexity reduction and detail preservation, greatly boosting visual fidelity. By introducing cross-attention layers into the model architecture, we turn diffusion models into powerful and flexible generators for general conditioning inputs such as text or bounding boxes and high-resolution synthesis becomes possible in a convolutional manner. Our latent diffusion models (LDMs) achieve a new state of the art for image inpainting and highly competitive performance on various tasks, including unconditional image generation, semantic scene synthesis, and super-resolution, while significantly reducing computational requirements compared to pixel-based DMs.*
The original codebase can be found at [CompVis/latent-diffusion](https://github.com/CompVis/latent-diffusion).
> [!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.
[Latte: Latent Diffusion Transformer for Video Generation](https://huggingface.co/papers/2401.03048) from Monash University, Shanghai AI Lab, Nanjing University, and Nanyang Technological University.
The abstract from the paper is:
*We propose a novel Latent Diffusion Transformer, namely Latte, for video generation. Latte first extracts spatio-temporal tokens from input videos and then adopts a series of Transformer blocks to model video distribution in the latent space. In order to model a substantial number of tokens extracted from videos, four efficient variants are introduced from the perspective of decomposing the spatial and temporal dimensions of input videos. To improve the quality of generated videos, we determine the best practices of Latte through rigorous experimental analysis, including video clip patch embedding, model variants, timestep-class information injection, temporal positional embedding, and learning strategies. Our comprehensive evaluation demonstrates that Latte achieves state-of-the-art performance across four standard video generation datasets, i.e., FaceForensics, SkyTimelapse, UCF101, and Taichi-HD. In addition, we extend Latte to text-to-video generation (T2V) task, where Latte achieves comparable results compared to recent T2V models. We strongly believe that Latte provides valuable insights for future research on incorporating Transformers into diffusion models for video generation.*
**Highlights**: Latte is a latent diffusion transformer proposed as a backbone for modeling different modalities (trained for text-to-video generation here). It achieves state-of-the-art performance across four standard video benchmarks - [FaceForensics](https://huggingface.co/papers/1803.09179), [SkyTimelapse](https://huggingface.co/papers/1709.07592), [UCF101](https://huggingface.co/papers/1212.0402) and [Taichi-HD](https://huggingface.co/papers/2003.00196). To prepare and download the datasets for evaluation, please refer to [this https URL](https://github.com/Vchitect/Latte/blob/main/docs/datasets_evaluation.md).
This pipeline was contributed by [maxin-cn](https://github.com/maxin-cn). The original codebase can be found [here](https://github.com/Vchitect/Latte). The original weights can be found under [hf.co/maxin-cn](https://huggingface.co/maxin-cn).
> [!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
Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
First, load the pipeline:
```python
importtorch
fromdiffusersimportLattePipeline
pipeline=LattePipeline.from_pretrained(
"maxin-cn/Latte-1",torch_dtype=torch.float16
).to("cuda")
```
Then change the memory layout of the pipelines `transformer` and `vae` components to `torch.channels-last`:
video=pipeline(prompt="A dog wearing sunglasses floating in space, surreal, nebulae in background").frames[0]
```
The [benchmark](https://gist.github.com/a-r-r-o-w/4e1694ca46374793c0361d740a99ff19) results on an 80GB A100 machine are:
```
Without torch.compile(): Average inference time: 16.246 seconds.
With torch.compile(): Average inference time: 14.573 seconds.
```
## 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 [`LattePipeline`] for inference with bitsandbytes.
LEDITS++ was proposed in [LEDITS++: Limitless Image Editing using Text-to-Image Models](https://huggingface.co/papers/2311.16711) by Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, Apolinário Passos.
The abstract from the paper is:
*Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDITS++, an efficient yet versatile and precise textual image manipulation technique. LEDITS++'s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDITS++ and its improvements over previous methods. The project page is available at https://leditsplusplus-project.static.hf.space .*
> [!TIP]
> You can find additional information about LEDITS++ on the [project page](https://leditsplusplus-project.static.hf.space/index.html) and try it out in a [demo](https://huggingface.co/spaces/editing-images/leditsplusplus).
> [!WARNING]
> Due to some backward compatibility issues with the current diffusers implementation of [`~schedulers.DPMSolverMultistepScheduler`] this implementation of LEdits++ can no longer guarantee perfect inversion.
> This issue is unlikely to have any noticeable effects on applied use-cases. However, we provide an alternative implementation that guarantees perfect inversion in a dedicated [GitHub repo](https://github.com/ml-research/ledits_pp).
We provide two distinct pipelines based on different pre-trained models.
We introduce LongCat-Image, a pioneering open-source and bilingual (Chinese-English) foundation model for image generation, designed to address core challenges in multilingual text rendering, photorealism, deployment efficiency, and developer accessibility prevalent in current leading models.
### Key Features
- 🌟 **Exceptional Efficiency and Performance**: With only **6B parameters**, LongCat-Image surpasses numerous open-source models that are several times larger across multiple benchmarks, demonstrating the immense potential of efficient model design.
- 🌟 **Superior Editing Performance**: LongCat-Image-Edit model achieves state-of-the-art performance among open-source models, delivering leading instruction-following and image quality with superior visual consistency.
- 🌟 **Powerful Chinese Text Rendering**: LongCat-Image demonstrates superior accuracy and stability in rendering common Chinese characters compared to existing SOTA open-source models and achieves industry-leading coverage of the Chinese dictionary.
- 🌟 **Remarkable Photorealism**: Through an innovative data strategy and training framework, LongCat-Image achieves remarkable photorealism in generated images.
- 🌟 **Comprehensive Open-Source Ecosystem**: We provide a complete toolchain, from intermediate checkpoints to full training code, significantly lowering the barrier for further research and development.
For more details, please refer to the comprehensive [***LongCat-Image Technical Report***](https://arxiv.org/abs/2412.11963)
[LTX-2](https://hf.co/papers/2601.03233) is a DiT-based foundation model designed to generate synchronized video and audio within a single model. It brings together the core building blocks of modern video generation, with open weights and a focus on practical, local execution.
You can find all the original LTX-Video checkpoints under the [Lightricks](https://huggingface.co/Lightricks) organization.
The original codebase for LTX-2 can be found [here](https://github.com/Lightricks/LTX-2).
## Two-stages Generation
Recommended pipeline to achieve production quality generation, this pipeline is composed of two stages:
- Stage 1: Generate a video at the target resolution using diffusion sampling with classifier-free guidance (CFG). This stage produces a coherent low-noise video sequence that respects the text/image conditioning.
- Stage 2: Upsample the Stage 1 output by 2 and refine details using a distilled LoRA model to improve fidelity and visual quality. Stage 2 may apply lighter CFG to preserve the structure from Stage 1 while enhancing texture and sharpness.
# Stage 2 inference with distilled LoRA and sigmas
video,audio=pipe(
latents=upscaled_video_latent,
audio_latents=audio_latent,
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=3,
noise_scale=STAGE_2_DISTILLED_SIGMA_VALUES[0],# renoise with first sigma value https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/src/ltx_pipelines/ti2vid_two_stages.py#L218
noise_scale=STAGE_2_DISTILLED_SIGMA_VALUES[0],# renoise with first sigma value https://github.com/Lightricks/LTX-2/blob/main/packages/ltx-pipelines/src/ltx_pipelines/distilled.py#L178
You can use `LTX2ConditionPipeline` to specify image and/or video conditions at arbitrary latent indices. For example, we can specify both a first-frame and last-frame condition to perform first-last-frame-to-video (FLF2V) generation:
Because the conditioning is done via latent frames, the 8 data space frames corresponding to the specified latent frame for an image condition will tend to be static.
## Multimodal Guidance
LTX-2.X pipelines support multimodal guidance. It is composed of three terms, all using a CFG-style update rule:
1. Classifier-Free Guidance (CFG): standard [CFG](https://huggingface.co/papers/2207.12598) where the perturbed ("weaker") output is generated using the negative prompt.
2. Spatio-Temporal Guidance (STG): [STG](https://huggingface.co/papers/2411.18664) moves away from a perturbed output created from short-cutting self-attention operations and substitutes in the attention values instead. The idea is that this creates sharper videos and better spatiotemporal consistency.
3. Modality Isolation Guidance: moves away from a perturbed output created from disabling cross-modality (audio-to-video and video-to-audio) cross attention. This guidance is more specific to [LTX-2.X](https://huggingface.co/papers/2601.03233) models, with the idea that this produces better consistency between the generated audio and video.
These are controlled by the `guidance_scale`, `stg_scale`, and `modality_scale` arguments and can be set separately for video and audio. Additionally, for STG the transformer block indices where self-attention is skipped needs to be specified via the `spatio_temporal_guidance_blocks` argument. The LTX-2.X pipelines also support [guidance rescaling](https://huggingface.co/papers/2305.08891) to help reduce over-exposure, which can be a problem when the guidance scales are set to high values.
The LTX-2.X models are sensitive to prompting style. Refer to the [official prompting guide](https://ltx.io/model/model-blog/prompting-guide-for-ltx-2) for recommendations on how to write a good prompt. Using prompt enhancement, where the supplied prompts are enhanced using the pipeline's text encoder (by default a [Gemma 3](https://huggingface.co/google/gemma-3-12b-it-qat-q4_0-unquantized) model) given a system prompt, can also improve sample quality. The optional `processor` pipeline component needs to be present to use prompt enhancement. Enable prompt enhancement by supplying a `system_prompt` argument:
[LTX-Video](https://huggingface.co/Lightricks/LTX-Video) is a diffusion transformer designed for fast and real-time generation of high-resolution videos from text and images. The main feature of LTX-Video is the Video-VAE. The Video-VAE has a higher pixel to latent compression ratio (1:192) which enables more efficient video data processing and faster generation speed. To support and prevent finer details from being lost during generation, the Video-VAE decoder performs the latent to pixel conversion *and* the last denoising step.
You can find all the original LTX-Video checkpoints under the [Lightricks](https://huggingface.co/Lightricks) organization.
> [!TIP]
> Click on the LTX-Video models in the right sidebar for more examples of other video generation tasks.
The example below demonstrates how to generate a video optimized for memory or inference speed.
<hfoptionsid="usage">
<hfoptionid="memory">
Refer to the [Reduce memory usage](../../optimization/memory) guide for more details about the various memory saving techniques.
[Compilation](../../optimization/fp16#torchcompile) is slow the first time but subsequent calls to the pipeline are faster. [Caching](../../optimization/cache) may also speed up inference by storing and reusing intermediate outputs.
- Refer to the following recommended settings for generation from the [LTX-Video](https://github.com/Lightricks/LTX-Video) repository.
- The recommended dtype for the transformer, VAE, and text encoder is `torch.bfloat16`. The VAE and text encoder can also be `torch.float32` or `torch.float16`.
- For guidance-distilled variants of LTX-Video, set `guidance_scale` to `1.0`. The `guidance_scale` for any other model should be set higher, like `5.0`, for good generation quality.
- For timestep-aware VAE variants (LTX-Video 0.9.1 and above), set `decode_timestep` to `0.05` and `image_cond_noise_scale` to `0.025`.
- For variants that support interpolation between multiple conditioning images and videos (LTX-Video 0.9.5 and above), use similar images and videos for the best results. Divergence from the conditioning inputs may lead to abrupt transitions in the generated video.
- LTX-Video 0.9.7 includes a spatial latent upscaler and a 13B parameter transformer. During inference, a low resolution video is quickly generated first and then upscaled and refined.
- LTX-Video 0.9.7 distilled model is guidance and timestep-distilled to speedup generation. It requires `guidance_scale` to be set to `1.0` and `num_inference_steps` should be set between `4` and `10` for good generation quality. You should also use the following custom timesteps for the best results.
- Base model inference to prepare for upscaling: `[1000, 993, 987, 981, 975, 909, 725, 0.03]`.
- LTX-Video 0.9.8 distilled model is similar to the 0.9.7 variant. It is guidance and timestep-distilled, and similar inference code can be used as above. An improvement of this version is that it supports generating very long videos. Additionally, it supports using tone mapping to improve the quality of the generated video using the `tone_map_compression_ratio` parameter. The default value of `0.6` is recommended.
prompt="""The camera pans over a snow-covered mountain range, revealing a vast expanse of snow-capped peaks and valleys.The mountains are covered in a thick layer of snow, with some areas appearing almost white while others have a slightly darker, almost grayish hue. The peaks are jagged and irregular, with some rising sharply into the sky while others are more rounded. The valleys are deep and narrow, with steep slopes that are also covered in snow. The trees in the foreground are mostly bare, with only a few leaves remaining on their branches. The sky is overcast, with thick clouds obscuring the sun. The overall impression is one of peace and tranquility, with the snow-covered mountains standing as a testament to the power and beauty of nature."""
# prompt = """A woman walks away from a white Jeep parked on a city street at night, then ascends a staircase and knocks on a door. The woman, wearing a dark jacket and jeans, walks away from the Jeep parked on the left side of the street, her back to the camera; she walks at a steady pace, her arms swinging slightly by her sides; the street is dimly lit, with streetlights casting pools of light on the wet pavement; a man in a dark jacket and jeans walks past the Jeep in the opposite direction; the camera follows the woman from behind as she walks up a set of stairs towards a building with a green door; she reaches the top of the stairs and turns left, continuing to walk towards the building; she reaches the door and knocks on it with her right hand; the camera remains stationary, focused on the doorway; the scene is captured in real-life footage."""
- LTX-Video supports loading from single files, such as [GGUF checkpoints](../../quantization/gguf), with [`loaders.FromOriginalModelMixin.from_single_file`] or [`loaders.FromSingleFileMixin.from_single_file`].
