If you run into any problems feel free to pm me on civitai/discord
Hunyuan 720p I2V
1316.72s 73F 688x800 22steps dpmpp_2m simple
Hunyuan720pI2V Q6_K gguf (adjust as needed)
https://huggingface.co/city96/HunyuanVideo-I2V-gguf/tree/main
llava_llama3_vision
https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/blob/main/split_files/clip_vision/llava_llama3_vision.safetensors
clip_l (renamed to clip_hunyuan)
https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/tree/main/split_files/text_encoders
hunyuan_video_vae_bf16
https://huggingface.co/Comfy-Org/HunyuanVideo_repackaged/tree/main/split_files/vae
Python version 3.12.7 Cuda 12.6 Torch 2.6.0+cu126
Triton windows: https://github.com/woct0rdho/triton-windows/releases
Once you’ve downloaded the appropriate wheel file for your Python version, proceed to open your command prompt and navigate to the directory where the downloaded file is located. Then, run the following command:
Through python_embeded
python.exe -m pip install triton-3.2.0-(filename)
python.exe -m pip install sageattention==1.0.6
------------------------------------------------------------------------------------------------------
Wan2.1
562.51s 512x512 uni_pc simple 33F
12step & 8step split works as intended
81F 1018.89s!
81F 573.99s!
8step Split 161F/10s (16fps) 512x512 uni_pc simple 6760.70seconds but it works! (metadata baked png posted)
I got buzz to tip, post your creations to the workflow gallery or add the resource to your posts, Have fun!
Wan2.1 I2V update published!
49F
512x512
12step(2stage 6+6)
Uni_pc
Simple
Seems like each lora I add +200-400s inference time
33F 700-900s
49F 1000-1500s
Wan2.1 480p I2V /unet (Adjust as needed)
https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/blob/main/wan2.1-i2v-14b-480p-Q6_K.gguf
Clip vision /clip_vision
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/clip_vision/clip_vision_h.safetensors
Vae /vae
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors
Text encoder /clip or /text_encoders
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors
(Optional) Upscale /upscale_models
https://huggingface.co/lokCX/4x-Ultrasharp/blob/main/4x-UltraSharp.pth
-----------------------------------------------------------------------------------------------------
Skyreels
Final barebones+ text weighted Hunyuan Lora compatibility update published
831.61 seconds (NO US)
932.07 seconds (NO US)
published vids in showcase
Could potentially work on 8GBVRAM or lower if you tinker with virtual_vram_gb on the UnetLoaderGGUFDisTorchMultiGPU custom node (if you have sufficient RAM GB)
Stage 1 415.369 Stage 2 315.937 VAE 70.838 total 837.93seconds. Q6+6stepLORA+SmoothLORA+DollyLORA
(I have defaulted to DPM++2M\Beta + Smooth LORA always (without for human-centric), AVG runtime: 700-900s 73F No US)
Comfyui_MultiGPU = UnetLoaderGGUFDisTorchMultiGPU (image latent batch 4 flux-finetune Q8, replace gguf loader in txt2img workflow)
Comfyui_KJNodes = TorchCompileModelHyVideo, Patch Sage Attention KJ, Patch Model Patcher Order (Add nodes>KJNodes>Experimental)
∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨∨
https://huggingface.co/spacepxl/skyreels-i2v-smooth-lora
∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧∧
Finetune the virtual_vram_gb to fit your requirements (I suggest looking at the Comfyui cmd for the distorch allocation values that show up after loading the model into SamplerCustom) or use normal Unet Loader (GGUF) with skyreels-hunyuan-I2V-Q?_
1st load
Prompt executed in 1662.22 seconds -587.365 seconds for upscale = 1075 seconds
640x864
73 frames (stable/generation time)
Steps: 6-12 (Stage 1 6 steps + Stage 2 6 steps)
cfg: 4.0
Sampler: Euler
Scheduler: Simple
(Original Kijai WF https://huggingface.co/Kijai/SkyReels-V1-Hunyuan_comfy/blob/main/skyreels_hunyuan_I2V_native_example_01.json)
Barebones I2V workflow with Upscaler, optimised on 306012GBVRAM + 32GBRAM
Make sure you update comfyui, torch & cuda
Run the update_comfyui.bat from the update folder
Go back to your python_embeded folder
Click on the file directory bar at the top, type cmd then hit enter
In cmd type "python.exe -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu126"
∨∨ May ruin older workflows ∨∨
Run the other update.bat if it still aint working: update_comfyui_and_python_dependencies.bat
∧∧ May ruin older workflows ∧∧
Workflow Resources:
Fast_Hunyuan Lora (models/lora): https://huggingface.co/Kijai/HunyuanVideo_comfy/blob/main/hyvideo_FastVideo_LoRA-fp8.safetensors
GGUF Model (Switch the models to fit your requirements) (models/unet):
https://huggingface.co/Kijai/SkyReels-V1-Hunyuan_comfy/blob/main/skyreels-hunyuan-I2V-Q6_K.gguf
VAE model (models/vae): https://huggingface.co/Kijai/HunyuanVideo_comfy/blob/main/hunyuan_video_vae_bf16.safetensors
Clip_l model (I renamed it to clip_hunyuan) (models/clip):
llava_llama3 model (models/clip):
https://huggingface.co/calcuis/hunyuan-gguf/blob/main/llava_llama3_fp8_scaled.safetensors
Upscale Model (models/upscale_models):
https://huggingface.co/uwg/upscaler/blob/main/ESRGAN/4x-UltraSharp.pth
Personal Generation Times
after 1st load base gen runtimes(2Stage+Vae Decode):
758.173 seconds
704.589 seconds
with suggested lora after 1st:
779.494
169F tests after 1st (No Load Test):
OOM
121F test after 1st+6stepLORA+smoothLORA (No Load Test):
1st stage
525.14s 1st iteration
729.66s 2nd
736.19s 3rd
645.15s 4th
665.55s 5th
764.12s 6th/Average
2nd stage
81.90s 1st+2nd iteration
OOM
Instant requeue after oom runs from 2nd stage
6.17s 1st Iteration
113.74s 2nd+3rd
222.92s 4th
327.62s 5th
282.29s 6th/Average
VAE 128.309s
97F tests I2V+6stepLora (posted in gallery) (no oom yet)
1123s
1013s
Description
(Original Kijai WF https://huggingface.co/Kijai/SkyReels-V1-Hunyuan_comfy/blob/main/skyreels_hunyuan_I2V_native_example_01.json)
Barebones I2V workflow with Upscaler, optimised for 306012GBVRAM + 32GBRAM (potentially on lower vram)
Make sure you update comfyui, torch & cuda
FAQ
Comments (44)
The achieve is RAR, but downloaded as ZIP. That's why some ppl cant open it under windows. To open it just renamed extension.
Updated, thank you
Really appreciate your putting together a workflow for us 12GBers. I've updated everything but I still can't find nodes:
- UnetLoaderGGUFDisTorchMultiGPU
- ImageNoiseAugmentation
- TorchCompileModelHyVideo
I see ImageNoiseAugmentation comes from KJNodes which I have installed and updated. Any help welcome!
For the ImageNoiseAugmentation side quest, you get the reward from here https://github.com/kijai/ComfyUI-KJNodes
same here missing torch!
@ToxicBot Except I have that installed and updated (as far as I can tell) and that node is not available. Uninstall and reinstall?
@sarashinai Comfyui_MultiGPU = UnetLoaderGGUFDisTorchMultiGPU
Comfyui_KJNodes = TorchCompileModelHyVideo (Add nodes>KJNodes>Experimental), ImageNoiseAugmentation
@tsolful Can confirm, Comfyui_MultiGPU package does contain the needed Unet loader node. For whatever reason, the KJNodes wasn't updating using the Comfyui Manager so I uninstalled it and did a manual git clone of in the custom nodes folder, that brought in the missing nodes. THANK YOU!
"TorchCompileModelHyVideo" error, Uninstall kjnodes in Manager and reinstall kjnodes nightly after restart.
1662.22 seconds !!!!
-587.365 seconds for upscale = 1075seconds for 73 frame
@tsolful What GPU
I have raised this topic for a long time. There is a memory overflow and the render increases significantly. It is not being cleared. Only restarting comfyui helps.
InstructPixToPixConditioning: Input type (CUDABFloat16Type) and weight type (CPUBFloat16Type) should be the same
Any ideas? Can't get past this error.
Didn't run into this error my self, but searching for this error returned me with this :
https://github.com/smthemex/ComfyUI_FramePainter/issues/3
https://stackoverflow.com/questions/59013109/runtimeerror-input-type-torch-floattensor-and-weight-type-torch-cuda-floatte
LLM Why Reasoning:
Model Loading: Your GGUF file (e.g., for HunyuanVideo) might not be properly moved to the GPU during loading.
ComfyUI Config: A misconfiguration in ComfyUI or its nodes might default weights to CPU instead of CUDA.
Mixed Precision: Using FP16 (common with GGUF quants like Q5_K_M or Q6_K) can expose device bugs if not handled consistently.
@tsolful That's a handful. I'll have to dig in and see if I can figure those out. Thank you!
@sarashinai or its as simple as your PyTorch might be outdated or mismatched with your CUDA toolkit
BRO, i read somewhere that triton must match with cuda version. pytorch 2.6 + cuda 12.6, and so on. also u need to ad 2 cuda path Vars to windows sistem. I just erase my oldest comfyui and reinstall all step by step, its aworking amazing in my 306012gb . so kep trying!!!!
I'm getting this error during the Sampler step:
SamplerCustom
shape '[1, 32, 12, 80, 80]' is invalid for input of size 1228800
I haven't changed the default image size in the workflow. Using the Q6 GGUF and SageAttention
Could be something to do with sageattention, but i ain't to knowledgeable. I downloaded/use triton with my WF
@tsolful Same I use Triton as well, I'm not sure what's going on. Workflow doesn't work for me.
@Melty1989
LLM Possible causes:
Mismatched Input Dimensions
Incorrect Model or Checkpoint Configuration
Sampler-Specific Issue
Batch Size Mismatch
VRAM Issues or Corrupt Installation
Try 640x640 batch 1 Sampler Euler or Dpm++ simple/beta, clear cache models/restart comfyui
could possibly be pytorch and cuda mismatch
@tsolful Already tried 640x640. Cuda toolkit is also at 12.4 as is the pytorch. Not really sure what's going on.
@Melty1989 Could be Cuda and pytorch, im on 12.6
@tsolful Hmm, so I'd have to upgrade my current CUDA toolkit to 12.6 then. Goddamnit.
@tsolful I tried it with 12.6 as well, upgraded pytorch - no dice at 640x640
Fixed it (thanks tsolful!). The issue was I had an outdated TeaCache version
sageattention-2.1.1 must be compiled manually -> https://www.reddit.com/r/StableDiffusion/comments/1h7hunp/how_to_run_hunyuanvideo_on_a_single_24gb_vram_card/
it will take a long time to complete, so dont close cmd window
Seeds in the 000's seem to have good character movement
Encountered error: Exception: An error occurred in the ffmpeg subprocess:
which made my GPU crash temporarily making my screen black for 1 second.
Trying to queue a new batch i ran into this error: RuntimeError: CUDA error: unknown error
I rebooted Comfyui, works as expected
Where do we get UnetLoaderGGUFDisTorchMultiGPU
Its the only one im missing, and im not finding it.
ComfyUI-MultiGPU through comfyui manager
@tsolful Thank you! :) Found it using Comfyui MultiGPU
Im getting another error:
PathchSageAttentionKJ - No module named 'sageattention'
No problem, Install triton or sageattention
@tsolful For some reason it's still saying I don't have the nodes: unetloaderggufadvancedtorchmultgpu, wanimagetovideo, saveWEBM ???? I've restarted the container and installed multigpu through comfyui, but it still says I dont have th em.
@staythepath I am currently on Python 3.12.2, Cuda 12.6, Torch 2.6.0+cu126. if that helps let me know
AMAZING AMAZING AMAZING !!!!! THX SO MUCH !!!!!! 546 s. for people with bad results or error -> U NEED CRAZY INSTALATIONS BROS, I WAS LIKE 8 HRS INSTALLING ALL, WHEELS, PYTORCH, CUDA, ETC. AND TESTING
GPU?
works fine on my 3060 12go just disable "Patch Model Patcher Order"
Why?
"SamplerCustom
expected str, bytes or os.PathLike object, not NoneType"
I'm not using the portable version and installed triton and sage attention
bro. u need sageattention-2.1.1 to be installed manualy. i have days fixing problems for this , and all worked for me , very fast and good quality. 2 options -> install linux as a sub system or follow this -> https://www.reddit.com/r/StableDiffusion/comments/1h7hunp/how_to_run_hunyuanvideo_on_a_single_24gb_vram_card/
this is the crazyest install i have ever done. i have delete all and reinstall again, but finally its completely working
sageattention-2.1.1 -> compiled manually i mean, u need ti create ur own
@blablabla666234 GREAT lmao. I might skip this for awhile and come back later. I just reinstalled the portable version of comfy and it took hours to figure out how to use my backup of ReActor.