Wan Video Upscale in ComfyUI: Full Workflow Guide

🎬 Before / After — Same Clip, Full Pipeline

Split-screen: source clip on the left, output after the Wan enhance pass, RealESRGAN upscale, and RIFE frame interpolation on the right.

🎬 Before / After — Same Clip, Full Pipeline

Split-screen: source clip on the left, output after the Wan enhance pass, RealESRGAN upscale, and RIFE frame interpolation on the right.

Load your source video

Add a VHS_LoadVideo node (from Video Helper Suite). Click choose video to upload and select your file. Leave frame_load_cap at 0 to process the whole clip, or set a number to test on a short clip first.

🔍 Click to zoom

Resize the video down for the enhance pass

Connect the loaded video to an ImageResizeKJv2 node (from KJNodes). Set width and height to a size your GPU can handle — this workflow uses 512x896. Set the resize method to nearest-exact and crop mode to center.

This step exists because the Wan diffusion pass is expensive. Running it at full resolution would need far more VRAM than most GPUs have — you upscale to full size after the enhance pass, not before.

Load the Wan model, VAE, and text encoder

Add three nodes: WanVideoModelLoader (select Wan2_1-T2V-14B_fp8_e4m3fn.safetensors, base_precision fp16_fast, attention_mode sdpa), WanVideoVAELoader (Wan2_1_VAE_bf16.safetensors), and LoadWanVideoT5TextEncoder (umt5-xxl-enc-bf16.safetensors).

Build the LoRA stack

Add six WanVideoLoraSelect nodes and chain them together (each one's prev_lora output feeds the next one's prev_lora input). Set each LoRA and strength as listed in the table above. Connect the final chained output into the WanVideoModelLoader's LoRA input.

🔍 Click to zoom

Set up VRAM management

Add a WanVideoBlockSwap node and connect it to the model loader. Start with it bypassed. If you hit an out-of-memory error in step 7, enable it and raise the swap block count gradually (up to 40) until generation completes without crashing.

Configure the sampler

Add a WanVideoSampler node. Set steps to 4, cfg to 1.0, denoise to 0.5 (for cleanup without changing the video's content), and scheduler to flowmatch_causvid.

Leave the prompt empty in WanVideoTextEncode if you just want cleanup with no style changes. If your video is longer than 81 frames, add a WanVideoContextOptions node set to uniform_standard with context frames at 81 — this splits long videos into chunks so you don't run out of memory.

Run the enhance pass and check the output

Click Queue Prompt. Generation time depends on your video length and GPU — expect several minutes for a short clip on a 24 GB card. Once done, a WanVideoDecode node converts the result back to viewable frames, and a VHS_VideoCombine node saves it as an MP4.

🔍 Click to zoom

Upscale and interpolate

Feed the enhanced frames into UpscaleModelLoader (set to RealESRGAN_x2.pth) → ImageUpscaleWithModel → ImageScale (set to your target resolution, e.g. 2880x2160, method lanczos) → RIFE VFI (rife49.pth, multiplier 2 to double your frame rate).

Export the final video

Connect the RIFE output to a final VHS_VideoCombine node. Set frame_rate to double your source (e.g. 24 → 48 if you used a multiplier of 2 in RIFE). Click Queue Prompt to render the finished, upscaled video.

Yes — swap the checkpoint in WanVideoModelLoader for a Wan 2.2 model and check that your LoRAs are built for the same version, since Wan 2.1 and 2.2 LoRAs are not always cross-compatible. The node structure and settings stay the same.

The Wan diffusion pass is what fixes blur and compression artifacts, and running diffusion at full resolution needs far more VRAM than most consumer GPUs have. Processing small, then upscaling with RealESRGAN afterward, gets you the quality benefit without the memory cost.

On an RTX 4090, a short clip at 512x896 with 4 sampler steps takes a few minutes for the enhance pass, plus additional time for the RealESRGAN upscale and RIFE interpolation. Longer videos or lower-end GPUs will take considerably longer — use WanVideoContextOptions to chunk anything over 81 frames.

No. Realism Boost and Detail Enhancer have the biggest visible impact on upscale quality. MPS Rewards, AccVid, MoviiGen, and LightX2V mainly affect motion quality and render speed — drop them first if you're short on VRAM.

Yes — this pipeline works the same way on Wan-, LTX-, or other AI-generated clips. It's especially useful for cleaning up low-resolution AI video output before final export.

A plain ESRGAN upscale only enlarges existing pixels — it can't fix blur or motion artifacts already baked into the video. This workflow's Wan diffusion pass redraws the frames first, so the ESRGAN upscale that follows has cleaner source material to work with.