[Cosmos 3] Preserve image and video conditioning with supplied latents

未關閉
#15,342 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
4/5
預估耗時
3-5 天
新手友好度
48/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
python, pytorch

研究方向

Start by reading the Cosmos 3 Omni standard pipeline’s latent preparation and masked-position initialization, then trace the corresponding modular image-to-video and video-to-video paths. Use the reproduction to compare internally sampled noise with the same noise passed through latents, for both image and video conditioning. Done means explicit latents preserve the reference conditioning and match the internally sampled path for the same inputs.

由索引模型根據 Issue 內容生成。

描述

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.

主要語言
Python
星號
34.9k
分支
7.5k
平均合併
3 天 16 小時
30 天內合併 PR
65

環境準備

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

huggingface/diffusers 的其他 Issue

查看 huggingface/diffusers 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。