Repository navigation
Description
Describe the bug
Passing pre-generated vision latents to Cosmos 3 Omni skips the step that initializes the masked positions from the reference image or video. With the same supplied noise, switching a red reference to blue leaves the standard pipeline output identical in the example below. Sampling the noise inside the pipeline correctly responds to the reference.
The same missing initialization also causes the modular image-to-video and video-to-video paths to diverge when given noise drawn from the same seed. I can send one fix for the standard pipeline and the corresponding modular block if this scope looks right.
Reproduction
The following uses the standard pipeline and compares its internal noise with the same noise supplied explicitly.
import torch
from PIL import Image
from tokenizers import Tokenizer, models, pre_tokenizers
from transformers import PreTrainedTokenizerFast
from diffusers import AutoencoderKLWan, Cosmos3OmniPipeline, Cosmos3OmniTransformer, UniPCMultistepScheduler
from diffusers.utils.torch_utils import randn_tensor
def build_pipeline():
tokenizer = Tokenizer(models.WordLevel({"<unk>": 0, "<eos>": 1, "<|vision_start|>": 2}, unk_token="<unk>"))
tokenizer.pre_tokenizer = pre_tokenizers.Whitespace()
tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer, unk_token="<unk>", eos_token="<eos>")
tokenizer.chat_template = "{% for message in messages %}{{ message['content'] }}{% endfor %}"
torch.manual_seed(0)
transformer = Cosmos3OmniTransformer(
head_dim=6, hidden_act="relu2", hidden_size=6, intermediate_size=12, latent_channel=16, latent_patch_size=1,
num_attention_heads=1, num_hidden_layers=1, num_key_value_heads=1, patch_latent_dim=16,
qk_norm_for_text=False, rms_norm_eps=1e-5, rope_axes_dim=[1, 1, 1], vocab_size=3,
)
vae = AutoencoderKLWan(
base_dim=3, z_dim=16, dim_mult=[1, 1, 1, 1], num_res_blocks=1, temperal_downsample=[False, True, True],
)
pipe = Cosmos3OmniPipeline(
transformer=transformer, text_tokenizer=tokenizer, vae=vae, scheduler=UniPCMultistepScheduler(),
enable_safety_checker=False,
)
pipe.set_progress_bar_config(disable=True)
return pipe
torch.set_num_threads(1)
pipe = build_pipeline()
common = dict(
prompt="a dog", negative_prompt="bad quality", height=16, width=16, num_inference_steps=2,
guidance_scale=1.0, output_type="latent", use_system_prompt=False, add_resolution_template=False,
add_duration_template=False,
)
def gen():
return torch.Generator().manual_seed(0)
red = Image.new("RGB", (16, 16), "red")
blue = Image.new("RGB", (16, 16), "blue")
for key, a, b, frames in (
("image", red, blue, 5),
("video", [red] * 5, [blue] * 5, 9),
):
expected = pipe(**common, **{key: a}, num_frames=frames, generator=gen()).video
other_reference = pipe(**common, **{key: b}, num_frames=frames, generator=gen()).video
assert (expected[:, :, :1] - other_reference[:, :, :1]).abs().max() > 0
noise = randn_tensor(expected.shape, generator=gen(), dtype=expected.dtype)
actual = pipe(**common, **{key: a}, num_frames=frames, latents=noise).video
other = pipe(**common, **{key: b}, num_frames=frames, latents=noise).video
same_inputs_max_error = (expected - actual).abs().max().item()
changed_condition_max_effect = (actual - other).abs().max().item()
print(
f"{key} conditioning with explicit noise: "
f"same inputs max error={same_inputs_max_error:.6f}; "
f"changing conditioning max effect={changed_condition_max_effect:.6f}"
)Logs
image conditioning with explicit noise: same inputs max error=663.352112; changing conditioning max effect=0.000000
video conditioning with explicit noise: same inputs max error=484.900635; changing conditioning max effect=0.000000
System Info
Diffusers 0.44.0.dev0 at 8edb6d8, Windows 11 x64, Python 3.12.14, PyTorch 2.11.0+cpu, Transformers 5.20.0. CPU with tiny randomly initialized components.
Activity