krea/krea-realtime-video

Krea Realtime 14B is distilled from the Wan 2.1 14B text-to-video model using Self-Forcing, a technique for converting regular video diffusion models into autoregressive models. It achieves a text-to-video inference speed of 11fps using 4 inference steps on a single NVIDIA B200 GPU. For more details on our training methodology and sampling innovations, refer to our technical blog post.

Inference code can be found here.

• Our model is over 10x larger than existing realtime video models
• We introduce novel techniques for mitigating error accumulation, including KV Cache Recomputation and KV Cache Attention Bias
• We develop memory optimizations specific to autoregressive video diffusion models that facilitate training large autoregressive models
• Our model enables realtime interactive capabilities: Users can modify prompts mid-generation, restyle videos on-the-fly, and see first frames within 1 second

Video To Video

Krea realtime allows users to stream real videos, webcam inputs, or canvas primitives into the model, unlocking controllable video synthesis and editing

Text To Video

Krea realtime allows users to generate videos in a streaming fashion with ~1s time to first frame.

Use it with our inference code

Set up

sudo apt install ffmpeg # install if you haven't already
git clone https://github.com/krea-ai/realtime-video
cd realtime-video
uv sync
uv pip install flash_attn --no-build-isolation
huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B --local-dir-use-symlinks False --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download krea/krea-realtime-video krea-realtime-video-14b.safetensors --local-dir-use-symlinks False --local-dir checkpoints/krea-realtime-video-14b.safetensors

Run

export MODEL_FOLDER=Wan-AI
export CUDA_VISIBLE_DEVICES=0 # pick the GPU you want to serve on
export DO_COMPILE=true

uvicorn release_server:app --host 0.0.0.0 --port 8000

And use the web app at http://localhost:8000/ in your browser (for more advanced use-cases and custom pipeline check out our GitHub repository: https://github.com/krea-ai/realtime-video)

Use it with 🧨 diffusers

Krea Realtime 14B can be used with the diffusers library utilizing the new Modular Diffusers structure

# Install diffusers from main
pip install git+github.com/huggingface/diffusers.git

Text to Video

import torch
from tqdm import tqdm
from diffusers.utils import export_to_video
from diffusers import ModularPipeline
from diffusers.modular_pipelines import PipelineState

repo_id = "krea/krea-realtime-video"
pipe = ModularPipeline.from_pretrained(repo_id, trust_remote_code=True)
pipe.load_components(
    trust_remote_code=True,
    device_map="cuda",
    torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
)
for block in pipe.transformer.blocks:
    block.self_attn.fuse_projections()

num_blocks = 9

frames = []
state = PipelineState()
prompt = ["a cat sitting on a boat"]

generator = torch.Generator(device=pipe.device).manual_seed(42)
for block_idx in tqdm(range(num_blocks)):
    state = pipe(
        state,
        prompt=prompt,
        num_inference_steps=6,
        num_blocks=num_blocks,
        block_idx=block_idx,
        generator=generator,
    )
    frames.extend(state.values["videos"][0])

export_to_video(frames, "output.mp4", fps=24)

Video to Video

import torch
from tqdm import tqdm
from diffusers.utils import load_video, export_to_video
from diffusers import ModularPipeline
from diffusers.modular_pipelines import PipelineState

repo_id = "krea/krea-realtime-video"
pipe = ModularPipeline.from_pretrained(repo_id, trust_remote_code=True)
pipe.load_components(
    trust_remote_code=True,
    device_map="cuda",
    torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
)
for block in pipe.transformer.blocks:
    block.self_attn.fuse_projections()

num_blocks = 9
video = load_video("https://app-uploads.krea.ai/public/a8218957-1a80-43dc-81b2-da970b5f2221-video.mp4")

frames = []
prompt = ["A car racing down a snowy mountain"]

state = PipelineState()
generator = torch.Generator("cuda").manual_seed(42)
for block_idx in tqdm(range(num_blocks)):
    state = pipe(
        state,
        video=video,
        prompt=prompt,
        num_inference_steps=6,
        strength=0.3,
        block_idx=block_idx,
        generator=generator,
    )
    frames.extend(state.values["videos"][0])

export_to_video(frames, "output-v2v.mp4", fps=24)

Streaming Video to Video

Using the video_stream input will process video frames in as they arrive, while maintaining temporal consistency across chunks.

import torch
from collections import deque
from tqdm import tqdm
from diffusers.utils import load_video, export_to_video
from diffusers import ModularPipeline
from diffusers.modular_pipelines import PipelineState

repo_id = "krea/krea-realtime-video"
pipe = ModularPipeline.from_pretrained(repo_id, trust_remote_code=True)
pipe.load_components(
    trust_remote_code=True,
    device_map="cuda",
    torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
)
for block in pipe.transformer.blocks:
    block.self_attn.fuse_projections()

n_samples = 9
frame_sample_len = 12
video = load_video(
    "https://app-uploads.krea.ai/public/a8218957-1a80-43dc-81b2-da970b5f2221-video.mp4"
)

# Simulate streaming video input
frame_samples = [
    video[sample_start : sample_start + frame_sample_len]
    for sample_start in range(0, n_samples * frame_sample_len, frame_sample_len)
]

frames = []
state = PipelineState()
prompt = ["A car racing down a snowny mountain road"]

block_idx = 0
generator = torch.Generator("cpu").manual_seed(42)
for frame_sample in tqdm(frame_samples):
    state = pipe(
        state,
        video_stream=frame_sample,
        prompt=prompt,
        num_inference_steps=6,
        strength=0.3,
        block_idx=block_idx,
        generator=generator,
    )
    frames.extend(state.values["videos"][0])

    block_idx += 1

export_to_video(frames, "output-v2v-streaming.mp4", fps=24)

Using LoRAs

import torch
from collections import deque
from tqdm import tqdm
from diffusers.utils import export_to_video
from diffusers import ModularPipeline
from diffusers.modular_pipelines import PipelineState

repo_id = "krea/krea-realtime-video"
pipe = ModularPipeline.from_pretrained(repo_id, trust_remote_code=True)
pipe.load_components(
    trust_remote_code=True,
    device_map="cuda",
    torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
)
pipe.transformer.load_lora_adapter(
    "shauray/Origami_WanLora",
    prefix="diffusion_model",
    weight_name="origami_000000500.safetensors",
    adapter_name="origami",
)
for block in pipe.transformer.blocks:
    block.self_attn.fuse_projections()

num_blocks = 9

frames = []
state = PipelineState()
prompt = ["[origami] a cat sitting on a boat"]

generator = torch.Generator("cuda").manual_seed(42)
for block_idx in tqdm(range(num_blocks)):
    state = pipe(
        state,
        prompt=prompt,
        num_inference_steps=6,
        num_blocks=num_blocks,
        block_idx=block_idx,
        generator=generator,
    )
    frames.extend(state.values["videos"][0])

export_to_video(frames, "output.mp4", fps=24)

Optimized Inference

To optimize inference speed and memory usage on Hopper level GPUs (H100s), we recommend using torch.compile, Sageattention and FP8 quantization with torchao.

First let's set up our depedencies by enabling Sageattention via Hub kernels and installing the torchao and kernels packages.

export DIFFUSERS_ENABLE_HUB_KERNELS=true
pip install -U kernels torchao

Alternatively, you can use Flash Attention 3 via kernels by disabling Sageattention:

export DISABLE_SAGEATTENTION=1

Then we will iterate over the blocks of the transformer and apply quantization and torch.compile.

import torch
from collections import deque
from tqdm import tqdm
from diffusers.utils import export_to_video
from diffusers import ModularPipeline
from diffusers.modular_pipelines import PipelineState
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, quantize_

repo_id = "krea/krea-realtime-video"
pipe = ModularPipeline.from_pretrained(repo_id, trust_remote_code=True)
pipe.load_components(
    trust_remote_code=True,
    device_map="cuda",
    torch_dtype={"default": torch.bfloat16, "vae": torch.float16},
)

for block in pipe.transformer.blocks:
    block.self_attn.fuse_projections()

# Quantize just the transformer blocks
for block in pipe.transformer.blocks:
    quantize_(block, Float8DynamicActivationFloat8WeightConfig())

# Compile just the attention modules
for submod in pipe.transformer.modules():
    if submod.__class__.__name__ in ["CausalWanAttentionBlock"]:
        submod.compile(fullgraph=False)

num_blocks = 9

state = PipelineState()
prompt = ["a cat sitting on a boat"]

# Compile warmup
for block_idx in range(num_blocks):
    state = pipe(
        state,
        prompt=prompt,
        num_inference_steps=2,
        num_blocks=num_blocks,
        num_frames_per_block=num_frames_per_block,
        block_idx=block_idx,
        generator=torch.Generator("cuda").manual_seed(42),
    )

# Reset state
state = PipelineState()
generator = torch.Generator("cuda").manual_seed(42)
for block_idx in tqdm(range(num_blocks)):
    state = pipe(
        state,
        prompt=prompt,
        num_inference_steps=6,
        num_blocks=num_blocks,
        block_idx=block_idx,
        generator=generator,
    )
    frames.extend(state.values["videos"][0])

export_to_video(frames, "output.mp4", fps=24)