Workflow templates (for SeedVR2.5)

[naxci1]

1. Workflow that takes the files in the directory and processes them sequentially (output resolution is automatically adjusted).

HB2k-SeedVR2.5-Auto.json

[adrientoupet]

Thanks @naxci1 - might want to enable cache_model on the VAE and DiT if doing batch processing. Thanks for sharing

[naxci1]

2. Workflow that processes a single video file according to the specified frame count (output resolution is automatically adjusted).

HB2k-SeedVR2.5-Auto-File.json

[fortelos-boop]

Dear Naxci1, I used your provided workflow. my file size is roughly 6mg with 30 seconds duration on 5060 TI 16G VRAM
I encountered out of memory messsages at the end of the processs.
Should I use smaller sized video file to avoid this error message or should adjust some option values on the work flow ?

[naxci1]

Dear Naxci1, I used your provided workflow. my file size is roughly 6mg with 30 seconds duration on 5060 TI 16G VRAM I encountered out of memory messsages at the end of the processs. Should I use smaller sized video file to avoid this error message or should adjust some option values on the work flow ?

Hi @fortelos-boop

What are the dimensions of the input video? If it's large, you should adjust the settings or reduce the dimensions.

https://www.youtube.com/watch?v=MBtWYXq_r60&t=24s

He explains these settings towards the end of his video.

[fortelos-boop]

didnt know the dimesions, but i changed it into 640X480 now. I will try with the changed dimesion video file.
The youtube you refered is not usable(understandible) to me since i am not good at technical details at all.

[naxci1]

3. Workflow for image file (output resolution is set manually).

SeedVR2_simple_image_upscale.json

[naxci1]

4. Workflow for image file (output resolution is set by multiplying).

SeedVR2_simple_image_multiply.json

[Fenixdg3duy]

@naxci1 Thank you very much for sharing.

[skv89]

Here are two of my workflows modified from the workflow by LHKJacky. The main difference I see between my workflow vs naxci1's workflows is that there is the metabatch manager node so you process long videos in batches of "x" # of frames so you don't overload your memory. I spent days trying to get the SEEDVR2 Torch Compile node working to no avail. I got problems after problems using every single config and reinstalling different versions of pytorch+cuda, triton, xformers, flash_attn and I finally just gave up. I might lose out on some signifcant speed gains by using this node but it is too much of a headache getting it to work and I might wait a few months before I revisit this node so it is currently set to bypassed.

You can drag and drop the .png workflow images below into ComfyUI and it will automatically load up the workflow.

](url) This workflow has the ResizeAndPadImage node built in so you can downscale the resolution within the workflow. Normally for 720p videos, I downscale 0.5x to 360p and then upscale back to 720p or even 1080p for best results. For 1080p videos, I usually get better results downscaling to 0.33x instead of 0.5x and upscaling it back to 1080p with SEEDVR2. But you have to test a small section of the video first because for some videos, you lose too much original info by downscaling to 0.33x so the final video might have too much artificial details that don't match the original. I find that downscaling using an external app using high bitrate (high quality, slow encoding) settings yield slightly better results than using the built in downscale node but the difference is very slight. This workflow does not have ResizeAndPadImage node to downscale the video within the workflow. You also must upscale with integers so you cannot manually set the resolution.

[urixl]

You can drag and drop the .png workflow images below into ComfyUI and it will automatically load up the workflow.

Mind. Blown. ComfyUI embeds workflow into the PNG metadata!

One thing: ComfyUI throws the error: "Request entity too large" for a 114Mb input file.
That's strange.

[skv89]

Just change the default limit from 100mb to however big you need in the ComfyUI settings

[urixl]

Thank you. Haven't found it in Settings though, but command-line parameter works.

[fortelos-boop]

I have encountered "memory limitation error" when I try to use FalshVR on 5060 TI (16G VRAM) with Intel 14th generation CPU and 16G RAM. Is it related to VRAM or CPU RAM ? Could you suggest a modified FlashVR workflow that can be working on my system?

[urixl]

It's related to your system RAM.
I found that even 32Gb RAM is just enough for upscale, but 64Gb seems reasonable.

[naxci1]

5. Workflow that takes the files in the directory and processes them sequentially (output resolution is automatically and manually adjusted).

Naxci1-SeedVR2.5-Auto-Manuel.json

[skv89]

I just tried your workflow again and for some reason the ComfyMath packs cannot be installed onto my comfyui anymore. I tested it on both my machines running the latest desktop ComfyUI. Do you get the same error?

[naxci1]

I don't have a problem; it's giving this error because the node isn't installed. Install the missing nodes in comfyui manager, or manually install the comfyui-easy-use node.

[skv89]

These are the node packs in your workflow. The only one that can't be installed is the ComfyMath node pack. It's strange because your workflow worked fine until I updated my comfyui to the latest version. Same issue on both pcs.

[naxci1]

Go to the missing part and install it from there. Actually, the ComfyUI code sometimes acts up, otherwise I use it with both the old and new versions without any problems. In fact, the old ComfyUI runs faster, so I'm only using the new one for testing.

[skv89]

Yea I tried installing the missing nodes many many times on both machines. It just doesn't work. Oh well that is fine. I already tweaked the Numz workflow to work well for my setups. Before I was trying to avoid enabling offload_devices and enabling cache_models. But now after I enabled them, it actually runs noticeably faster for my multi-batch runs (about 20%+ speedup). The only offload_device I didn't enable is in the Upscaler node. The only issue is that enabling too much of these takes up System Ram and that occassionally causes oom in system ram because VHS_LoadVideoPath tries to load the entire clip (2-minute 540p clip in my case) into one giant float32 numpy array (and failing). Despite I have 96gb of ram, the loader crashes when it can’t allocate a contiguous array to fit the block in and the more cache comfyui uses in system ram, the more fragmented the ram usage is resulting in less contiguous free memory blocks. So the more caching used, despite the speed increase, the more risk of random crashes due to more fragmentation and less large contiguous blocks of free memory. I hope ComfyUI can fix this in the future.

[naxci1]

My source videos are generally 240p; when they're larger, I downscale them to 240p. Because the more you upscale, the higher the quality, I recommend at least 3x upscaling, and 4x if possible. However, I've noticed that upscaling 4x+ doesn't significantly improve quality; for example, 4x and 5x upscaling resulted in almost the same quality. Therefore, 3x-4x is the best upscaling.

Since I don't have enough VRAM, I'm forced to download to 240p and then upscale it to 720p. So I adjust it according to my hardware, but if I had higher-end hardware, I would also upscale larger videos by 3x.

[brumskysb]

Interesting, I'll test this out and see how it works for my content. Thanks again.

[merhovon]

Hi, thanks a lot for the workflows! I just tried to use one of them, but I keep running into the following error message. Do you have any idea what might be causing this?

`
** ComfyUI startup time: 2026-02-19 09:30:42.049
** Platform: Windows
** Python version: 3.12.11 (main, Aug 18 2025, 19:17:54) [MSC v.1944 64 bit (AMD64)]
** Python executable: C:\Users\Shadow\Documents\ComfyUI.venv\Scripts\python.exe
** ComfyUI Path: C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI
** ComfyUI Base Folder Path: C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI
** User directory: C:\Users\Shadow\Documents\ComfyUI\user
** ComfyUI-Manager config path: C:\Users\Shadow\Documents\ComfyUI\user__manager\config.ini
** Log path: C:\Users\Shadow\Documents\ComfyUI\user\comfyui.log
[ComfyUI-Manager] Skipped fixing the 'comfyui-frontend-package' dependency because the ComfyUI is outdated.
[PRE] ComfyUI-Manager
Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Found comfy_kitchen backend triton: {'available': False, 'disabled': True, 'unavailable_reason': "ImportError: No module named 'triton'", 'capabilities': []}
Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['apply_rope', 'apply_rope1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'scaled_mm_nvfp4']}
Checkpoint files will always be loaded safely.
Total VRAM 15356 MB, total RAM 16364 MB
pytorch version: 2.10.0+cu130
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA RTX 2000 Ada Generation : cudaMallocAsync
Reserving 1024.0MB vram for other applications.
Using async weight offloading with 2 streams
Enabled pinned memory 7363.0
working around nvidia conv3d memory bug.
Using pytorch attention
Python version: 3.12.11 (main, Aug 18 2025, 19:17:54) [MSC v.1944 64 bit (AMD64)]
ComfyUI version: 0.14.1

[ComfyUI-Easy-Use] server: v1.3.6 Loaded
[ComfyUI-Easy-Use] web root: C:\Users\Shadow\Documents\ComfyUI\custom_nodes\comfyui-easy-use\web_version/v2 Loaded
Trying to load custom node C:\Users\Shadow\Documents\ComfyUI\custom_nodes\comfyui-videohelpersuite
Trying to load custom node C:\Users\Shadow\Documents\ComfyUI\custom_nodes\seedvr2_videoupscaler
⚠️ SeedVR2 optimizations check: SageAttention ❌ | Flash Attention ❌ | Triton ❌
💡 For best performance: pip install sageattention flash-attn triton
🔧 Conv3d workaround active: PyTorch 2.10.0, cuDNN 91200 (fixing VAE 3x memory bug)
📊 Initial CUDA memory: 13.89GB free / 15.00GB total

[ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
FETCH DATA from: C:\Users\Shadow\Documents\ComfyUI\user__manager\cache\1514988643_custom-node-list.json [DONE]
[ComfyUI-Manager] All startup tasks have been completed.
got prompt
Failed to validate prompt for output 3:

• SeedVR2LoadDiTModel 54:
  • Value not in list: attention_mode: 'flash_attn' not in ['sdpa', 'flash_attn_2', 'flash_attn_3', 'sageattn_2', 'sageattn_3']
    Output will be ignored
    Using proactor: IocpProactor
    !!! Exception during processing !!! cannot allocate memory for array
    Traceback (most recent call last):
    File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 530, in execute
    output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 334, in get_output_data
    return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 308, in _async_map_node_over_list
    await process_inputs(input_dict, i)
    File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 296, in process_inputs
    result = f(**inputs)
    ^^^^^^^^^^^
    File "C:\Users\Shadow\Documents\ComfyUI\custom_nodes\comfyui-videohelpersuite\videohelpersuite\load_video_nodes.py", line 462, in load_video
    return load_video(**kwargs)
    ^^^^^^^^^^^^^^^^^^^^
    File "C:\Users\Shadow\Documents\ComfyUI\custom_nodes\comfyui-videohelpersuite\videohelpersuite\load_video_nodes.py", line 378, in load_video
    images = torch.from_numpy(np.fromiter(gen, np.dtype((np.float32, (new_height, new_width, 4 if alpha else 3)))))
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    MemoryError: cannot allocate memory for array

`

[naxci1]

Hi @merhovon

Change Node Setting: Go to the SeedVR2LoadDiTModel node and change attention_mode from flash_attn to sdpa. This is the most immediate fix for the crash.

Memory Management: When using Load Video (VHS), do not try to load long or high-resolution videos all at once. Set a frame_load_cap (e.g., 24 or 48) to avoid the MemoryError.

Install Dependencies: Open your terminal in the ComfyUI virtual environment and run:
pip install sageattention flash-attn triton

[naxci1]

I don't perform these operations in ComfyUI; I've written separate applications for this and shared them previously. You can use the Shutter Encoder application to reduce the size of your lossless video.

Afterwards, you should use a small application to split it into lossless segments and then merge them together.

#530

[merhovon]

I used the application to split my video into 3-minute segments and started the workflow (HB2k-SeedVR2.5-Auto.json), but at the end I received the following error message:

`
[11:50:57.188] ⚡ └─ Upscaled batch 156: 6.90s
[11:50:57.188] ⚡ └─ DiT inference 156: 6.90s
[11:50:57.188] ⚡ └─ (other operations): 0.76s
[11:50:57.195] 📊 After phase 2 (DiT upscaling):
[11:50:57.196] 📊 [VRAM] 2.40GB allocated / 10.56GB reserved / Peak: 8.86GB / 3.04GB free / 15.00GB total
[11:50:57.196] 📊 [RAM] 4.84GB process / 2.81GB others / 8.33GB free / 15.98GB total
[11:50:57.197] 📊 Memory changes: RAM -4.70GB
[11:50:57.197] 📊 Resetting VRAM peak memory statistics
[11:50:57.200]
[11:50:57.201] ━━━━━━━━ Phase 3: VAE decoding ━━━━━━━━
[11:50:57.205] 🔧 Pre-allocating output tensor: 4500 frames, 1280x960px, RGB (30.90GB)
[11:50:58.116] 🧹 Starting full cleanup
[11:50:58.117] 🧹 Cleaning up VAE components
[11:50:58.284] 🧹 VAE model deleted
[11:50:58.286] 🧹 Clearing memory caches (deep)...
[11:51:01.049] ✅ Completed full cleanup
!!! Exception during processing !!! [enforce fail at alloc_cpu.cpp:117] data. DefaultCPUAllocator: not enough memory: you tried to allocate 33177600000 bytes.
Traceback (most recent call last):
File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 530, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 334, in get_output_data
return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 308, in async_map_node_over_list
await process_inputs(input_dict, i)
File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\execution.py", line 296, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\comfy_api\internal_init.py", line 149, in wrapped_func
return method(locked_class, **inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Shadow\AppData\Local\Programs\ComfyUI\resources\ComfyUI\comfy_api\latest_io.py", line 1716, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Shadow\Documents\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\interfaces\video_upscaler.py", line 580, in execute
raise e
File "C:\Users\Shadow\Documents\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\interfaces\video_upscaler.py", line 497, in execute
ctx = decode_all_batches(
^^^^^^^^^^^^^^^^^^^
File "C:\Users\Shadow\Documents\ComfyUI\custom_nodes\seedvr2_videoupscaler\src\core\generation_phases.py", line 879, in decode_all_batches
ctx['final_video'] = torch.empty((total_frames, true_h, true_w, C), dtype=ctx['compute_dtype'], device=target_device)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: [enforce fail at alloc_cpu.cpp:117] data. DefaultCPUAllocator: not enough memory: you tried to allocate 33177600000 bytes.

Prompt executed in 02:08:28
`

Should I adjust the following setting if necessary?

[naxci1]

This application consistently causes problems on AMD GPUs, so Nvidia is more suitable for these tasks. Try dividing the video into 1-minute segments, because according to the error log, the RAM isn't sufficient. Keep the VAE decode tile low, for example 512 or less. Increase the RAM page file size or set it to automatic. You can solve the problems with the help of artificial intelligence. Since you have both AMD and Linux systems, I can't fully help you because I have Nvidia and a Windows system.

use model: 3BQ8
blocks_to_swap: 32

[merhovon]

I have NVIDIA RTX 2000 16GB VRAM with Windows 11 no AMD an Linux.

[naxci1]

Sorry, I got confused with the other comment; that one had an AMD and Linux system. The RTX 2000 is a bit old, so its performance is weak. How much RAM does it have? Set the pagefile to automatic, because it uses a lot of RAM for this process, and allocate it to 1-minute videos for better accuracy.

use model: 3BQ8
blocks_to_swap: 32