Cosmos Predict2

[a-r-r-o-w]

The cosmos is within us. We are made of star-stuff. We are a way for the universe to know itself.

cc @pjannaty @chenhsuanlin @fitsumreda @asfiyab-nvidia @amolfasale

[a-r-r-o-w]

@yiyixuxu This PR contains the version that works with FlowMatchEulerDiscreteScheduler. For the version that works with EDMEulerDiscrete, it is in this branch.

We will probably be sticking with this PR IIUC from our discussion, so I'll update the scheduler configs once you confirm (currently only the text2image pipeline has been updated, so will update video2world soon too)

PRs:

• https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image/discussions/2
• https://huggingface.co/nvidia/Cosmos-Predict2-14B-Text2Image/discussions/1
• https://huggingface.co/nvidia/Cosmos-Predict2-2B-Video2World/discussions/1
• https://huggingface.co/nvidia/Cosmos-Predict2-14B-Video2World/discussions/1

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[pjannaty]

LGTM! Thank you, Aryan and the HuggingFace team!

[yiyixuxu]

looks good!

[a-r-r-o-w]

We have new weight PRs now since the original weights were updated at the time of release:

• https://huggingface.co/nvidia/Cosmos-Predict2-2B-Text2Image/discussions/4
• https://huggingface.co/nvidia/Cosmos-Predict2-14B-Text2Image/discussions/2
• https://huggingface.co/nvidia/Cosmos-Predict2-2B-Video2World/discussions/2
• https://huggingface.co/nvidia/Cosmos-Predict2-14B-Video2World/discussions/2

[Arslan-Mehmood1]

python3.10

code

import time
import torch
from diffusers import Cosmos2TextToImagePipeline

# Available checkpoints: nvidia/Cosmos-Predict2-2B-Text2Image, nvidia/Cosmos-Predict2-14B-Text2Image
model_id = "nvidia/Cosmos-Predict2-14B-Text2Image"
pipe = Cosmos2TextToImagePipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16)
pipe.to("cuda")

prompt = "A close-up shot captures a vibrant yellow scrubber vigorously working on a grimy plate, its bristles moving in circular motions to lift stubborn grease and food residue. The dish, once covered in remnants of a hearty meal, gradually reveals its original glossy surface. Suds form and bubble around the scrubber, creating a satisfying visual of cleanliness in progress. The sound of scrubbing fills the air, accompanied by the gentle clinking of the dish against the sink. As the scrubber continues its task, the dish transforms, gleaming under the bright kitchen lights, symbolizing the triumph of cleanliness over mess."
negative_prompt = "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. Overall, the video is of poor quality."

start_time = time.time()
print("Generating image...")

output = pipe(
    prompt=prompt, negative_prompt=negative_prompt, generator=torch.Generator().manual_seed(1)
).images[0]

output.save("output_cosmos14b.png")

end_time = time.time()
time_taken = end_time - start_time
print(f"Image generated and saved as 'output.png'. Time taken: {time_taken:.2f} seconds")

better-profanity==0.7.0
boto3==1.38.31
decord==0.6.0
diffusers==0.33.1
ftfy==6.3.1
fvcore==0.1.5.post20221221
h11==0.16.0
huggingface-hub==0.32.4
hydra-core==1.3.2
imageio[pyav,ffmpeg]==2.37.0
iopath==0.1.10
ipdb==0.13.13
loguru==0.7.3
mediapy==1.2.4
megatron-core==0.12.1
modelscope==1.26.0
nltk==3.9.1
omegaconf==2.3.0
opencv-python==4.11.0.86
peft==0.15.2
qwen-vl-utils[decord]==0.0.11
retinaface-py==0.0.2
scikit-image==0.25.2
sentencepiece==0.2.0
termcolor==3.1.0
transformers==4.51.3
webdataset==0.2.111

ERROR:

Traceback (most recent call last):
  File "/home/paperspace/Ahmer/cosmos/test.py", line 3, in <module>
    from diffusers import Cosmos2TextToImagePipeline
ImportError: cannot import name 'Cosmos2TextToImagePipeline' from 'diffusers' (/home/paperspace/.virtualenvs/cosmos/lib/python3.10/site-packages/diffusers/__init__.py)

[a-r-r-o-w]

Since there hasn't been a diffusers release yet, you need to install from the main branch to use Cosmos. A release will be happening soon, but for the time being, please try: pip install git+https://github.com/huggingface/diffusers

[matabear-wyx]

why num_channels_latents = self.transformer.config.in_channels - 1 ?

[a-r-r-o-w]

Video2World models have an additional channel for concatenated conditioning mask, which indicates what frames to use for video extending condition signal. The actual latent channels is one less than the transformer in_channels