Face Restore / Face Boost running on CPU instead of GPU

[paradisebuster44]

First, confirm

• I have read the instruction carefully
• I have searched the existing issues
• I have updated the extension to the latest version

What happened?

I recently upgraded to a 5090 GPU and noticed that when I select GFPGANv.14.pth under face_restore_model, it runs on CPU instead of GPU.

The initial face swap is super fast but when it starts running the 'Restoring with GFPGANv1.4.pth' step, its very slow. These were my tests ::

Video Length - 5 second

-- First Swap --
Node - Fast Face swap
Enabled - ON
swap_model - inswapper_128.onnx
facedetection - retinaface_resnet50
face_restore_model - None
Time to complete - 74 seconds
Note: During the face swap I noticed little bit of GPU activity upto 20-25% but no CPU spike/activity. CPU activity stayed around 10-11%

-- Second Swap -
Node - Fast Face swap
Enabled - ON
swap_model - inswapper_128.onnx
facedetection - retinaface_resnet50
face_restore_model - GFPGANv1.4.pth
Time to complete - 244 seconds
Note: This time, during the face swap, similar GPU/CPU activity as before, but during the face restore step, GPU activity stayed around 10% while CPU activity was at 50-60% throughout

-- Third Swap --
Node - Fast Face swap
Enabled - ON
swap_model - inswapper_128.onnx
facedetection - retinaface_resnet50
face_restore_model - GFPGANv1.4.pth
Face Boost - ON - boost_model: GFPGANv1.4.pth - interpolation: Lancosz - restore_with_main_after: ON
Time to complete - 377 seconds
Note: Similar to the second swap, during the face boost step, only CPU showed activity.

This feels slower than what speeds I was getting on my older 3080Ti card. I alo tried the hyperswap, but it completely changes the area around the lips so its unusable in my case.

Any help with be appreciated. If there is a way to run face_restore and face boost on the GPU that would be great!

Steps to reproduce the problem

Your workflow
Run a simple face swap with a face image as source and video frames as input. Turn on face_restore_model and enable face boost

Sysinfo

Windows 11 + 5090 GPU

Relevant console log

Python version: 3.13.9 (tags/v3.13.9:8183fa5, Oct 14 2025, 14:09:13) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.3.75
ComfyUI frontend version: 1.32.9
[ReActor] - STATUS - Running v0.6.2-b1 in ComfyUI
Torch version: 2.8.0+cu128

Additional information

No response

[Buumcode]

1/ I'm not sure about 'Face Restore' CPU/GPU, but you can try this: Check your Reactor uses GPU -- not CPU!

2/ Best Face Restorer paired with Inswapper-128 is GPEN-BFR-512 (with face-restore_visibility 0.7-0.8). In my taste. (Best similarity face result). (Other versions of 'GPEN' is bad - only '512' works cool).

3/ Just INFO: 'Face Booster' is absolutely the same as 'Face Restore' - just performed on Target Face. So, in your test №3 you perform Face Restoring twice - before face swap and after. And get this long time as result. 'Face Booster' need ONLY on very poor target image (video) quality - to help Face Swapper RECOGNIZE where the hell face is (which over do face swap).

[paradisebuster44]

Tried Step 1 and it seems like I got the correct installation of packages :

onnx                                   1.20.0
onnxruntime-gpu                        1.23.2
opencv-python                          4.10.0.84

2. Tried inswapper with GPEN-BFR-512 and got these results for a 5 second video (160 frames) - No face booster:

[ReActor] 12:03:04 - STATUS - Working: source face index [0], target face index [0]
[ReActor] 12:03:06 - STATUS - Using Hashed Source Face(s) Model...
[ReActor] 12:03:06 - STATUS - Using Hashed Target Face(s) Model...
[ReActor] 12:03:08 - STATUS - Swapping...
[ReActor] 12:03:55 - STATUS - --Done!--
[ReActor] 12:04:01 - STATUS - Restoring with GPEN-BFR-512.onnx | Face Size is set to 512
Prompt executed in **227.14 seconds**

3. You're right. I have turned OFF the face booster.

Tried again without face boost and face restore, the swap took 74.86 seconds.

Compare this to FaceFusion, same video and same input image with these settings :

face_swapper - ON
face_enhancer - ON
Face enhancer Model - gpen_bfr_512
Face Enhancer Blend - 63
Face Swapper Model - inswapper_128
Face Swapper Pixel Boost - 128x128
Face Swapper weight - .75
Execution Providers - CUDA (ON) / tensorrt (OFF) / CPU (OFF)

1408x2560 and 32 frames per second
Extracting: 100% (161/161)
Processing: 100% (161/161)
[FACEFUSION.CORE] Merging video with a resolution of 1408x2560 and 32 frames per second
Merging: 100% (161/161)
[FACEFUSION.CORE] Skipping audio
[FACEFUSION.CORE] Processing to video succeeded in **32.63 seconds**

But with CUDA and tensorrt OFF and only cpu set to ON, I get these results :

[FACEFUSION.CORE] Processing step 1 of 1
[FACEFUSION.CORE] Extracting frames with a resolution of 1408x2560 and 32.0 frames per second
Extracting: 100% (161/161)
Processing: 100% (161/161)
[FACEFUSION.CORE] Merging video with a resolution of 1408x2560 and 32.0 frames per second
Merging: 100% (161/161)
[FACEFUSION.CORE] Skipping audio
[FACEFUSION.CORE] Processing to video succeeded in **279.71 seconds**

So definitely, Reactor is using CPU when face_boost is ON.

I tried YanWenKun's Portable ComfyUI with pre-installed ReActor but it has the same issue. Also, I saw this script to check if ReActor uses CPU or GPU. Not sure how to use it though.

[Buumcode]

Also, I saw script to check if ReActor uses CPU or GPU. Not sure how to use it though.

@Gourieff

Here is ZIP for testing on CPU-GPU.
How to use:
unpack to custom_nodes, as ~:\ComfyUI\custom_nodes\comfyui-reactor-node
make your test and look in console
***In every node you have line 'Device' -- type in it (small letters): 'auto', or 'cpu', or 'gpu'

--- Console Results (example, in short) ---
-Using CPU session for Hyperswap <--- you use CPU
-Using GPU session for Hyperswap <--- you use GPU
-inswapper_128.onnx with providers: ['CPUExecutionProvider'] <--- you use CPU
-inswapper_128.onnx with providers: ['CUDAExecutionProvider', 'CPUExecutionProvider'] <--- you use GPU
-Face Restore using device: cuda <------- GPU
-Face Restore using device: cpu <----- CPU

TEST:
1/ make face-to-video swap (without Face Restorer -- choose None);
2/ add after node Fast-Face-Swap -- node 'Restore Face' --> and make Face Restore over the same video (on GPU).
3/ Now choose CPU - again make only Restoring - and compare times.

RESULT: for me, I test GPEN-BFR-512 as restorer - and get VERY SIMILAR results on GPU and CPU:
cuda -- Prompt executed in 67.14 seconds
cpu -- Prompt executed in 67.03 seconds

What the hell??? CPU is even better.
Investigation needed.
2/ And need to compare with FaceFusion results on same files.

Reactor_v0.6.1-b2_CPU-GPU_face-restore_DEBUG.zip <---- 👀

[Buumcode]

***Just INFO:
for Face Boost users:

If you use Face Boost -- it makes Face Restore on Target Face (before Face Swap) and AUTOMATICALLY makes Face Restore on Swapped Face (after Face Swap).

[Buumcode]

Make some experiments to compare with FaceFusion.
RESULTS:

---Reactor:---
GPEN-BFR-512 (restore only) 21 frames, 720p:
cpu -- 10.93 sec
cuda -- 10.66 sec

GFPGAN_v1.4 -- (restore only) 21 frames, 720p:
cpu -- 144.42 sec
cuda -- 11.52 sec

---FaceFusion v3.3.2:---
GPEN-BFR-512 (restore only) 21 frames, 720p:
cpu -- 80.64 sec
cuda -- 05.74 sec

GFPGAN_v1.4 -- (restore only) 21 frames, 720p:
cpu -- 91.90 sec
cuda -- 03.22 sec

1/ Looks like GPEN-BFR-512 in Reactor ALWAYS use only GPU.
2/ GFPGAN_v1.4 show normal time slowing on forced CPU. In 'auto' Reactor uses only GPU.
3/ FaceFusion look faster (in times!) -- but MAY BE this difference from his 'Preview' function? Investigation needed.

RESOLUTION:
In FaceRestore/FaceBoost -- Reactor uses only GPU. All works well.

ZIP -- FaceFusion_v3.3.2 - 'freedom' version. Might be useful to someone for testing.

FaceFusion_v3.3.2.zip

[paradisebuster44]

Benchmark: ReActor Node vs FaceFusion 3.4.1 – RTX 5090

Test conditions

• Hardware: RTX 5090 + Ryzen 9 3900X + 64 GB RAM
• Video: 1280×720, 21 frames (~0.7 s at 30 fps)
• Source face image: 1200×1200
• All other settings default (no upscaling, no extra blending)

1. ReActor Node (ComfyUI) results

Test	Swap Device	Face Restore	Restore Device	Total Time	Time per frame	Notes
1	CPU (forced)	Off	–	4.10 s	195 ms	Log shows CUDA was still used
2	GPU	Off	–	4.06 s	193 ms
3	GPU	GPEN-BFR-512	GPU	12.01 s	572 ms
4	GPU	GPEN-BFR-512	CPU	11.59 s	552 ms	Almost no difference vs GPU restore

2. FaceFusion 3.4.1 (standalone) results

Test	Execution Provider	Face Restore	Total Time	Time per frame	Speedup vs ReActor (same config)
1	CUDA	Off	1.80 s	86 ms	2.26× faster
2	CPU	Off	12.90 s	614 ms	–
3	CUDA	GPEN-BFR-512	2.54 s	121 ms	4.72× faster
4	CPU	GPEN-BFR-512	36.41 s	1734 ms	–

3. Head-to-head comparison (RTX 5090)

Configuration	ReActor time	FaceFusion time	FaceFusion is faster by
Pure swap only (no restore)	~4.1 s	1.80 s	2.26× (takes only 44% of ReActor time)
Swap + GPEN-BFR-512 restore	~12.0 s	2.54 s	4.72× (takes only 21% of ReActor time)

Here is the combined, coherent version. It blends your punchy, real-world impact statements with the specific technical explanations into a solid summary.

Key Findings

• FaceFusion dominates on speed: It is 2.3× faster for raw swapping and ~4.7× faster with GPEN restoration on an RTX 5090. This is largely due to native frame batching that saturates the CUDA cores, whereas ReActor processes frames sequentially (one at a time).
• The "Restore" penalty is severe in ReActor: Adding GPEN-BFR-512 costs FaceFusion only +0.74s, but costs ReActor +8.00s. ReActor appears to treat Swap and Restore as disconnected events potentially reloading models or failing to persist tensors in VRAM.
• Device Selection Anomalies:
  • Phantom CPU Mode: ReActor ALWAYS uses the GPU for the swap phase, ignoring the 'CPU' selection in the node (logs confirm CUDA providers are loaded).
  • Identical Restore Times: For the restore phase, ReActor respects the CPU/GPU setting, yet execution times are nearly identical (~12s). This confirms that there is a bottleneck during data transfer.

---

Technical Analysis: Sequential Phases and Lack of Batching

The debug logs confirm that ReActor suffers from two major architectural inefficiencies: it separates the workflow into distinct, disconnected phases (Swap then Restore) and processes the video linearly (Frame-by-Frame) rather than using GPU batching.

1. Proof of Disconnected Phases (Swap $\to$ Memory Flush $\to$ Restore)

ReActor does not pipeline the restoration step. Instead, it iterates through the entire video to perform the swap, closes that task, and only then initializes the restoration model. This forces a full read/write cycle to System RAM between steps.

Evidence from Log:
The log shows the Swap phase completing (--Done!--) before the Restore phase even initializes.

[ReActor] 14:02:32 - DEBUG - Selected face: bbox=[...]   <-- Last frame of Swap
[ReActor] 14:02:32 - STATUS - --Done!--                  <-- SWAP PHASE ENDS
[ReActor] 14:02:33 - STATUS - Restoring with GPEN...     <-- RESTORE PHASE STARTS
[ReActor] 14:02:33 - DEBUG - Face Restore using device: cuda

• Implication: The GPU calculates the swap for 21 frames, moves them to CPU RAM, and then immediately has to fetch them back from CPU RAM to the GPU for restoration. This "Ping-Pong" of data creates the latency bottleneck.

2. Proof of Sequential Processing (No Batching)

ReActor's logs show a linear Python loop where the CPU interrupts the process to log detection data for every single frame individually.

Evidence from Log:
The log prints a Selected face debug entry for every single frame in the 21-frame sequence, with timestamps ticking forward linearly.

[ReActor] 14:02:31 - DEBUG - Selected face: bbox=[...]  <-- Frame 1
[ReActor] 14:02:31 - DEBUG - Selected face: bbox=[...]  <-- Frame 2
[ReActor] 14:02:31 - DEBUG - Selected face: bbox=[...]  <-- Frame 3
... (Repeats 21 times) ...
[ReActor] 14:02:32 - DEBUG - Selected face: bbox=[...]  <-- Frame 21

Summary

ReActor is running a Linear, Multi-Pass workflow:

1. Loop [Load Frame -> CPU Detect -> GPU Swap -> CPU Save]
2. Loop [Load Frame -> GPU Restore -> CPU Save]

FaceFusion runs a Batched, Single-Pass workflow:

1. Batch Load -> GPU [Swap + Restore] -> Save

[Gourieff]

Phantom CPU Mode: ReActor ALWAYS uses the GPU for the swap phase, ignoring the 'CPU' selection in the node (logs confirm CUDA providers are loaded).

Original ReActor node has no selection of 'GPU' or 'CPU' in the node, it always uses GPU where it can be used. In the version of Buumcode 'CPU' mode is not properly set, ONNX models are still in the GPU mode.
FF works faster because it utilizes parallel threads while processing.
ReActor was originally created as an image tool, not a video one. ComfyUI doesn't have chunking from the box when batching the images, and when you load a large amount of frames as a batch your RAM is likely to be OOMed. So that if you need to swap in the video you'd better use FF.
But I have these things in the road map, someday ReActor will be optimized for fast video processing as well.

But anyway guys thanks for your tests 👍

[paradisebuster44]

Phantom CPU Mode: ReActor ALWAYS uses the GPU for the swap phase, ignoring the 'CPU' selection in the node (logs confirm CUDA providers are loaded).

Original ReActor node has no selection of 'GPU' or 'CPU' in the node, it always uses GPU where it can be used. In the version of Buumcode 'CPU' mode is not properly set, ONNX models are still in the GPU mode. FF works faster because it utilizes parallel threads while processing. ReActor was originally created as an image tool, not a video one. ComfyUI doesn't have chunking from the box when batching the images, and when you load a large amount of frames as a batch your RAM is likely to be OOMed. So that if you need to swap in the video you'd better use FF. But I have these things in the road map, someday ReActor will be optimized for fast video processing as well.

But anyway guys thanks for your tests 👍

Thank you for the explanation and all your hard work and the ReActor node. I will continue to use it and will look forward to the updates. :)

[Buumcode]

But anyway guys thanks for your tests

Also:

1. I did test on speed of one face swap - image2image. Hot launch, on hashed source and hashed target - hyperswap_1b. FaceFusion make twice faster than Reactor - 0.76 sec vs 1.60 sec. But in Reactor quality of swapped face is better - FaceFusion did swapped face little blurry.

2. Also, for img2video face swap we need to test FaceFusion wrapper for ComfyUI - https://github.com/HJH-AILab/ComfyUI_Facefusion

***And also on market we have program VisoMaster, which maked for face swap too - https://github.com/visomaster/VisoMaster He has his own models for face swap.