Wan 2.2 AI Video Upscaler Detailer Workflow: ComfyUI Guide

AI Overview

What is a Wan 2.2 AI video upscaler detailer workflow?

It is a ComfyUI video-to-video pipeline that enlarges a clip and runs a low-denoise Wan 2.2 pass to rebuild texture, edges, and faces while preserving the original motion.

Which Wan 2.2 upscaling method should I use?

Use a Wan 2.2 5B latent or utility workflow when you want generative detail. Use a tiled or conservative pixel-space pass when identity, text, or product geometry must remain nearly unchanged.

What denoise strength works best for Wan 2.2 upscaling?

Start around 0.20–0.30 for faithful enhancement. Raise it only when the source is genuinely soft, because stronger denoise can invent texture, alter faces, and create frame-to-frame flicker.

Can Wan 2.2 upscale video to 1080p or 4K?

Yes. A stable route is source cleanup, then 1080p, then 4K if needed. Test a short section first and keep target width and height divisible by 64.

What Is the Wan 2.2 AI Video Upscaler Detailer Workflow?

The workflow separates two jobs that are often confused. Upscaling increases pixel dimensions. Detailing asks a generative model to reinterpret missing information. A traditional resize may produce a larger but still soft image; an aggressive detailer may create sharp hair or skin while changing the person. The useful middle ground is a low-noise second pass constrained by the original video.

The core graph is straightforward:

Load video → decode frames → choose target size → latent or pixel upscale → low-noise Wan pass → optional face detailer → rebuild video → restore audio

When this workflow is the right choice

Use it for 480p or 720p AI clips with good motion but weak micro-detail, compressed backgrounds, soft fabrics, or faces that collapse in wider shots. It is less suitable when the source has broken anatomy, severe scene changes, incorrect lip movement, or baked-in subtitles. Fix those defects first; upscaling makes visible errors larger.

Before a full render, test three to five representative seconds containing motion, fine texture, and the darkest part of the scene.

A useful review frame shows the detected region plus matched before-and-after crops. Judge structure and identity before judging sharpness.

What You Need: Models, Custom Nodes and Workflow Files

Update ComfyUI before importing the graph. Old core nodes or mismatched custom-node versions are the most common reason a shared workflow opens with red boxes. Restart after installing nodes so every model loader and video input is registered.

Required model stack

For a face branch, use a Wan video wrapper, video load/combine nodes, segmentation, and basic utilities. Keep the first version minimal; extra routing and style LoRAs make debugging harder.

Place each file in the directory expected by its loader, then verify the exact filename inside the graph. Do not assume a model with “high noise” and one with “low noise” are interchangeable: the low-noise stage is the polishing pass. If you regularly run long local workflows on limited hardware, the low-VRAM long-video guide explains the same practical ideas—short tests, tiled decoding, block movement, and segment handoffs.

How to Run the Wan 2.2 Video Upscale Workflow Step by Step

1. Load the approved source. Record its FPS, dimensions, duration, and audio state. Avoid variable-frame-rate inputs.
2. Trim a test window. Start with 49–81 frames and include movement plus a face when relevant.
3. Set the target dimensions. Choose a 1.5× or 2× pass and round width and height to values divisible by 64. Do not jump from a very small source directly to 4K.
4. Choose the upscale path. Latent upscale rebuilds detail; tiled processing is safer for long clips, product shapes, signs, and limited VRAM.
5. Run the low-noise sampler. Begin with conservative denoise and a neutral enhancement prompt such as “natural fine detail, stable texture, consistent face.” Avoid adding new story content.
6. Enable the face branch only if needed. Detect, segment, crop, refine, feather the mask, and composite the crop back into the same frame.
7. Rebuild and review. Restore the original FPS and audio, then watch at normal speed and frame by frame.

Before you queue the full clip

Compare the same frame numbers, not random thumbnails. Check eyes, fingers, logos, fabric direction, background texture, and color drift. If the short test fails, do not queue the full clip.

For inputs with heavy stylization, transitions, or effect layers, finish those creative decisions in your video-effects workflow before the upscale pass. The detailer should polish an approved edit rather than become another unpredictable generation stage.

Best Wan 2.2 Upscale Settings for Detail, Fidelity and VRAM

There is no universal best preset. The right value depends on how much valid information remains in the source and how much change the shot can tolerate.

How to read the result

More sharpness is not automatically better. Reduce denoise if hair, lettering, or jewelry changes between frames. Reduce the scale or tile size if memory spikes. If the result is stable but flat, add a little denoise before increasing steps.

Resolution, frame count, model precision, VAE mode, and concurrent nodes all affect VRAM. Queue one graph at a time and clear cached models between stages when supported. For a delivery checklist, see the Seedance 4K video tutorial.

How to Use Wan 2.2 Face Detailer Without Flicker

A face detailer gives one important region more pixels and focused sampling: detect, segment, expand the crop, refine, feather the mask, and composite it back.

Keep the detailer subordinate to the source

Start with one face and low denoise. Include the hairline, jaw, and some context; a tight crop can look pasted on. Keep the face index fixed in multi-person shots and watch for mask jumps.

Wan 2.2 face-detailer sample · rainy two-person shot

Watch the woman’s face through head movement and rain. A successful pass improves readability without changing expression, age, or identity.

For changing eyes, plastic skin, or a mask halo, reduce denoise, increase feathering, and simplify the prompt. During fast turns or occlusion, avoid forcing a full-strength replacement.

Wan 2.2 face-detailer sample · turning subject

A turning subject is a tougher temporal test than a static close-up: review the transition into profile and back again.

Troubleshooting and Choosing the Right Upscaler

Common failures and fixes

Choose generative Wan upscaling when the source genuinely lacks detail and some creative reconstruction is acceptable. Choose a conservative or pixel-space upscaler when text, products, architecture, or documentary identity must stay exact. A hybrid workflow often works best: fix obvious AI artifacts, perform a restrained Wan pass to 1080p, then use a conservative final resize.

Compare motion stability, error rate, render time, and shippable output—not one attractive frame. For a broader production comparison, read Wan 3.0 vs Seedance 2.5.

Conclusion

The most reliable Wan 2.2 AI video upscaler detailer workflow keeps changes controlled: test a short segment, upscale by 1.5×–2×, use low denoise, and apply face detailing only where needed. Review motion, identity, and texture at normal speed before processing the full clip, then create your next AI video with Seedance.