Add WanImage2ImagePipeline #12368

Activity

1. DN6

   added

   contributions-welcome

   on Sep 29, 2025
2. DN6

   added the

   Feature

   issue type on Sep 29, 2025
3. github-actions commented on Jan 9, 2026

   

   github-actionsbot

   on Jan 9, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

   Please note that issues that do not follow the contributing guidelines are likely to be ignored.
4. github-actions

   added

   staleIssues that haven't received updatesIssues that haven't received updates

   on Jan 9, 2026
5. yiyixuxu

   removed

   staleIssues that haven't received updatesIssues that haven't received updates

   on Jan 9, 2026
6. devanshi00 commented on Jan 30, 2026

   

   devanshi00

   on Jan 30, 2026

   Last edited by devanshi00

   More actions

   Hello I m interested in contributing to this issue.
   To do this, I plan to add a WanImage2ImagePipeline that closely mirrors standard diffusion image-to-image workflows while making only minimal, targeted changes:

   Reuse the existing WanPipeline architecture and weights

   The new pipeline will share the same tokenizer, text encoder, VAE, transformer(s), and scheduler as WanPipeline. This ensures users can switch seamlessly between text-to-image and image-to-image workflows without reloading or swapping model checkpoints.

   Introduce denoising-strength–based timestep control

   Image-to-image behavior will be enabled by computing the starting timestep from a user-provided strength parameter and running the denoising process only over the remaining timesteps. This follows the behavior of established img2img pipelines in diffusers and allows controlled deviation from the input image.

   Extend prepare_latents to support image initialization

   When an input image is provided:

   • The image will be encoded into latent space using the existing VAE
   • Noise will be added to the encoded latents at the selected starting timestep
   • These noisy latents will be used as the initial state for denoising

   If no image is provided, the pipeline will fall back to the existing random latent initialization, preserving current text-to-image behavior.

   Single-frame output

   Image-to-image generation will operate with num_frames = 1, leveraging Wan’s strong photo-realistic capabilities while remaining consistent with current single-frame text-to-image usage patterns.

   If this plan is okay then I will proceed further with the pr.
7. DN6

   assigned

   devanshi00

   on Feb 16, 2026
8. devanshi00 commented on Feb 18, 2026

   

   devanshi00

   on Feb 18, 2026

   Last edited by devanshi00

   More actions

   Hello I m not sure but I m not getting the correct output for Wan-AI/Wan2.2-T2V-A14B-Diffusers
   My Script

   # pip install ftfy
   import torch
   import numpy as np
   from diffusers import  WanPipeline
   from diffusers.utils import export_to_video, load_image
   pipeline = WanPipeline.from_pretrained(
       "Wan-AI/Wan2.1-T2V-14B-Diffusers",
       torch_dtype=torch.bfloat16
   )
   
   pipeline.to("cuda")
   
   prompt = """
   The camera rushes from far to near in a low-angle shot, 
   revealing a white ferret on a log. It plays, leaps into the water, and emerges, as the camera zooms in 
   for a close-up. Water splashes berry bushes nearby, while moss, snow, and leaves blanket the ground. 
   Birch trees and a light blue sky frame the scene, with ferns in the foreground. Side lighting casts dynamic 
   shadows and warm highlights. Medium composition, front view, low angle, with depth of field.
   """
   negative_prompt = """
   Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, 
   low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, 
   misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards
   """
   
   output = pipeline(
       prompt=prompt,
       negative_prompt=negative_prompt,
       num_frames=81,
       guidance_scale=5.0,
   ).frames[0]
   
   export_to_video(output, "output.mp4", fps=16)

   I get the log:

   Fetching 3 files: 100%|███████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 52428.80it/s]
   Download complete: : 0.00B [00:00, ?B/s]                                                                     | 0/3 [00:00<?, ?it/s]
   Loading weights: 100%|██████████████████████████████████████| 242/242 [00:00<00:00, 4765.67it/s, Materializing param=shared.weight]
   UMT5EncoderModel LOAD REPORT from: Wan-AI/Wan2.2-T2V-A14B-Diffusers
   Key                         | Status  |
   ----------------------------+---------+-
   encoder.embed_tokens.weight | MISSING |
   
   Notes:
   - MISSING     :those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
   /home/zeus/miniconda3/envs/cloudspace/lib/python3.12/site-packages/huggingface_hub/utils/_validators.py:202: UserWarning: The `local_dir_use_symlinks` argument is deprecated and ignored in `hf_hub_download`. Downloading to a local directory does not use symlinks anymore.
     warnings.warn(
   Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
   Download complete: : 0.00B [00:00, ?B/s]                                                                     | 0/3 [00:00<?, ?it/s]
   Loading weights: 100%|██████████████████████████████████████| 242/242 [00:00<00:00, 4765.67it/s, Materializing param=shared.weight]
   UMT5EncoderModel LOAD REPORT from: Wan-AI/Wan2.2-T2V-A14B-Diffusers
   Key                         | Status  |
   ----------------------------+---------+-
   encoder.embed_tokens.weight | MISSING |
   
   Notes:
   - MISSING     :those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
   
   100%|██████████████████████████████████████████████████████████████████████████████████████████████| 50/50 [08:08<00:00,  9.78s/it]
   It is recommended to use `export_to_video` with `imageio` and `imageio-ffmpeg` as a backend.
   These libraries are not present in your environment. Attempting to use legacy OpenCV backend to export video.
   Support for the OpenCV backend will be deprecated in a future Diffusers version
   

   The output is:

   output.mp4

   Can you please look into this? I m also attaching a community issue with same problem here. Without the Text to Video Model Running correctly I cannot run my Image2ImagePipeline code also
   @DN6 @yiyixuxu @vladmandic
9. devanshi00 commented on Feb 20, 2026

   

   devanshi00

   on Feb 20, 2026

   More actions

   Hello any updates on this?
   @DN6 @yiyixuxu