Introducing: MINIMAX H3 | EZ WF
UPDATE: The workflow has been completely redone on 08/08/2026. if you previously downloaded and it didn't work for you, try now. i have tested in a fresh install, it works, if you are having issues you most likely don't have the most up to date version on Comfy from the master branch here: https://github.com/Comfy-Org/ComfyUI/branches.
do NOT use the releases version 0.31.0, it will not work. Also be sure to update all of the nodes correctly, pull them from GitHub. Check your console for any missing dependencies like sage-attention etc and install them. i will no longer be responding to any issues regarding node pack issues, you just don't have the updated versions or you have something in your setup that conflicts. also make sure your models are compatible, users have been reporting it doesn't work, only to find the are using models that aren't compatible, this model was made with INT8 in mind.
If you are having black outputs you need to either remove the start-up flag --use-sage-attention, change to the normal VAE if you are using the INT8 one i linked, or again update comfy to latest version on master branch.
a custom workflow for Minimax's H3 video model. Its not too different to the base one Comfy provides, just has some little tweaks that i like to use, including Sage Attention/Turbo LoRAs and more!
This workflow is my personal workflow that I use for all of my Minimax H3 stuff, it may not be to everyone's taste but it works well for me and I'm sure others may find it useful.
Generated a 5 second clip with 0.2MP > 1x LTX Upscale at 16:9 in 7:32 on an RTX 3060 12GB/32GB Sys RAM using this WF. base workflow was more than double that time.
The workflow uses the most optimal boosters there are right now (all optional, you can just run without them)
Works on 12GB VRAM + 32 GB System RAM. not tested on anything below.
INCLUDES:
3 Ref Images | 2 Ref Audio | 1 Ref Video
Sage Attention Patch (For Lower VRAM Use)
Block Cache & Sol-Attn with Optimised Settings
Bob INT8 Loader (Faster INT8 Loading)
LTX Refiner
RTX Upscaler
ITV/T2V/RTV Modes
Nodes Needed:
[Bob INT8](https://github.com/BobJohnson24/ComfyUI-INT8-Fast)
[Sol-Att](https://github.com/kijai/ComfyUI-SolAttn_triton)
[KJNodes](https://github.com/kijai/ComfyUI-KJNodes)
[rgthree](https://github.com/rgthree/rgthree-comfy)
[EasyUse](https://github.com/yolain/ComfyUI-Easy-Use)
[VHS](https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite)
[RTX Upscaler](https://github.com/Comfy-Org/Nvidia_RTX_Nodes_ComfyUI)
[LTXVideo](https://github.com/Lightricks/ComfyUI-LTXVideo)
Models Needed:
-- H3 MODELS --
[TE]
-- LTX MODELS --
TURBO LORAS:
[LIGHTx]
Troubleshooting:
I'm missing the LTX Nodes and i get this error, what do i do?
ImportError: cannot import name 'pad' from 'kornia.geometry.transform.pyramid' (G:\StabilityMatrix-win-x64\Data\Packages\ComfyUIClean\venv\Lib\site-packages\kornia\geometry\transform\pyramid.py)-- downgrade Kornia in python packages to 0.8.2
The RTX Nodes are missing, even though i have them installed?Go to your python packages and install: nvidia-vfx
I'm missing MiniMaxH3MemoryEfficientSageAttentionPatchUpdate KJNodes through Github (go to your custom_nodes folder and open a git bash and type: git clone https://github.com/kijai/ComfyUI-KJNodes.git
MAKE SURE YOU ARE UPDATED LIKE SO:
Kofi: https://ko-fi.com/mrweaz
Disclaimer: I hold no liability for any images produced, including their legality, ethical implications, or potential misuse. By using this model, you agree that all responsibility for generated content rests solely with you. Use responsibly and within legal boundaries.
Description
Added Power Lora Loader
Added Chunking node for lower spec PCs
updated to the latest speed optimisations.
configured for Larrys 600 EMA LoRA by default.
added an Add Noise to images toggle (can help prompt adherence)
Easier to navigate.
included links to node packs and models, from the links i used.
FAQ
Comments (85)
Something just went wrong for me. I tried the workflow the first time for a 10 second clip and it worked fine. Then I pushed it to 15 seconds and it crashed my PC. I had Patch Sage Attention KJ node enabled on auto setting, but bypassed the 'Sage H3 Patch' node disabled as it threw up an error for me. I'm running this on an Nvidia RTX 5090, and just troubleshooting with Gemini, it mentioned something about KJ Nodes Sage Attention Patch not working with the newest Nvidia Blackwell architecture on my card...? The workflow looks really great, so I'm wondering if there is a fix to get it to not crash my PC...
Yeah there may be something in the boosters that's not compatible with that card, I can't tell you though as I use a 3060 and this is built around optimising that card, try disabling the SolAttn and Cache nodes might help a bit
also it seems the model struggles past 10 seconds, youll need to use a continuous video wf for that it seems. or use the last frame of your prevous gen to continue another clip
@mrweaz Thanks for the heads up and the tip.
kijai/ComfyUI-SolAttn_triton cant find this node/having huge problems with it :/
otimo trabalho!
Looking forward to trying this today as it has the ltx upscaler and I am to lazy to wire it up in a another workflow but might have to do it if this workflow fails me.
I do wonder about a few things.... like the model we are using here is quite large. I do find w4a8 model to be excellent and its small and because it's small I can do much higher resolution clips on a 16gb card.
I also have no idea what the boosters option is doing.
Yeah the LTX refiner works nicely. Adds alot to the gen times but if you get a good seed then it's worth it. The workflow should work fine, as long as you are updated, most people who are having issues haven't updated comfy to the latest branch.
I'm not sure about models, INT8 works fine for me, but I have heard that the w4a8 model is fast but doesn't work on Windows, I could be wrong though, not tested myself.
So, it was a caching option, I have since removed since it doesn't play nice with the Loras, the booster options now are just Sage Attention patch until I find a better way to implement Sol Att or Caching.
@mrweaz I have gotten this error: [ERROR] !!! Exception during processing !!! The size of tensor a (9344) must match the size of tensor b (128) at non-singleton dimension 2
This is on the ltx refine node. I am using the same ltx model as you had in the workflow which is https://huggingface.co/CornLogic/10EROS_1.4_Int8_ConvRot/blob/main/10Eros_v1.4_DMD_int8_convrot.safetensors, everything else is the same as well. Not sure why this is happening.
@dragonite9000263 That mistake is very strange. You should copy and paste it directly into Gemini or Chatgpt with reasoning so that it looks for a solution.
@dragonite9000263 ill have a look, last i checked it was working but i may have forgot something on the remake
@mrweaz Problem appears to be fixed. Also ltx upscaler is godly compared to other options. Def worth using this workflow.
@dragonite9000263 just realised now the audio is swapped, I'm gonna quickly fix so LTX doesn't replace the Audio.
A serious question. With CUDA 11.8, is this impossible to run? I think Sage Attention doesn't run outdated CUDAs, can anyone confirm me?
as far as i am aware this only works for CUDA 13
@mrweaz Python, Pytorch and the comfyui Kitchen are also supposed to be up to date, right? You who run this model, what configuration do you use apart from the workflow? Do you have updated Python, Pytorch and Kitchen?
@chamo9009 i use Stability Matrix to run Comfy, latest branch of Comfy and that's it.
@mrweaz I thank you, your answer is not really of much use to me, but I think I will be able to run your Workflow very interesting, of course, I think so... Because OOM is possible.
But it doesn't matter, I can use CUDA 13.
So RTX 3060 can benefit with Torch2.10+Cuda13 instead of Torch2.8+Cuda12.8 ?
I'm getting error in the Basic Guider node. I'm using only the Turbo Lora with Sage Attention + VRAM Clean. With and without Sol Attn, same error. All reference nodes are turned off and no upscaling or LTX detail. All node packs updated via Comfy Manager and using Comfy 0.31. I'm using a GGUF H3 encoder instead of the INT8 one, not sure if that's the problem. It's taking too long to download.
2026-08-08T18:47:30.284873 - [32m[INFO][0m [32mPrompt executed in 0.02 seconds[0m 2026-08-08T18:49:55.792712 - [32m[INFO][0m got prompt 2026-08-08T18:49:55.817051 - [1m[31m[ERROR][0m !!! Exception during processing !!! 'NoneType' object is not iterable 2026-08-08T18:49:55.819319 - [1m[31m[ERROR][0m Traceback (most recent call last): File "c:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, 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:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, 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:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in _async_map_node_over_list await process_inputs(input_dict, i) File "c:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs result = f(**inputs) File "c:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func return method(locked_class, **inputs) File "c:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\comfy_api\latest\_io.py", line 1990, in EXECUTE_NORMALIZED to_return = cls.execute(*args, **kwargs) File "C:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_custom_sampler.py", line 812, in execute guider.set_conds(conditioning) ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^ File "C:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\comfy_extras\nodes_custom_sampler.py", line 793, in set_conds self.inner_set_conds({"positive": positive}) ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^ File "c:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\comfy\samplers.py", line 1203, in inner_set_conds if self.model_patcher.is_dynamic() and comfy.sampler_helpers.cond_has_hooks(conds[k]): ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^ File "c:\tools\comfy_270\ComfyUI_windows_portable\ComfyUI\comfy\sampler_helpers.py", line 72, in cond_has_hooks for c in cond: ^^^^ TypeError: 'NoneType' object is not iterable 2026-08-08T18:49:55.823542 - [32m[INFO][0m [32mPrompt executed in 0.02 seconds[0m
just curious, are you using the most up to date version of the WF, there was a bug in one of them that caused this, its fixed now
@mrweaz I believe so, the file name has V32 REMADE in it. I even tried re-downloading the WF and it's throwing the same error on my VM. I used a modified "yetAnotherWorkFlowEasy v5" by adding SolAttn and Patch Sage Attn and that worked without the Basic Guider node throwing an error, so I don't think the error is related to the installed packages.
@Cowgary hm how odd, everything seems to work on my end.
"All node packs updated via Comfy Manager and using Comfy 0.31" ~ currently this is a bad way to approach Minimax H3, i had a ton of issues using the Manager to update nodes and the Release version of Comfy. its advisable to use the branch version of comfy, and clone the nodes directly from github. H3 is brand new and you need to be using the latest updates and nodes. my WF follows that principle for the latest optimisations. so I'm guessing its a node issue, or related to GGUF. if not i have no idea.
@mrweaz Hey buddy, I managed to get it running, but my video outputs are coming out black (in img2vid mode)—even though the first frame is active (node). Did that happen to you? I'm using roughly the same settings as the guy in this comment, but I can't figure out why the outputs are black.
@dupe_900 yes i had that exact issue, you need to make sure your Cross Attention Method is set to --use-pytorch-cross-attention instead of --use-sage-attention. and also make sure comfy is up to date from branches, not releases. im currently on version "master@cbbc9da" and i had the issue when using v0.30.0
@mrweaz From the branches and cross atenttion, I don't understand very well, I'm not using sage attention, that is, it doesn't work for me no matter how hard I try to use it, but I don't know if that means anything. Could you explain it to me with apples and not in professional terms? Please
@dupe_900 if you are using Stability Matrix you select Change Version > change version type to branches from Release and update. and for the --sage-attention you go into your launch arguments and remove --use-sage-attention and replace with --use-pytorch-cross-attention.
if you are using a downloaded version of comfy without stability matrix, in your git bash console do: git pull https://github.com/Comfy-Org/ComfyUI.git
@mrweaz I finally managed to download the Qwen 3 VL 32B text encoder safetensor and use that instead of the GGUF text encoder from Civitai. It's still throwing the same error with WF 3.3. So it must be something else and I don't have time at the moment to re-install all the packages via Git tonight. I might try it sometime next week.
@Cowgary wait, are you using all GGUF models, or have you got a mix of bf/INT8/GGUF? I feel like you probably have an incompatible model structure
This workflow looks promising, but unfortunately does not work for me even after ensuring comfy-ui and all the custom nodes are up to date this workflow still crashes my comfy-ui for some reason.
excellent wf. The video preview node is a huge timesaver. Much appreciated.
No problem! Glad you like it 😊
I like this Workflow, but thre is one thing i don't understand.
I tried a Ref2Vid with the Workflow, it kinda worked but in the End i got 2 files, one video and one audio file instead of 1 finished one, also the Video looks choppy when opening with VLC, in the H3 Base Video Node the Video is smooth and with sound
If the video is choppy you can try change the video encoder, I know it works cause I did that monkey video in the gallery right before going to bed, and that's straight from the H3 pass. In terms of two files, that's just how the saving video works, I may swap it out for a combine video node instead next update though, I'll have a mess round and see what I like better.
Thanks I try it out, maybe that was the case because i used nvenv hevc what vlc has troubles with, i tried another video player and there it worked.
Really love that Workflow even it looks complicated at first, but the options to choose to use upscaling and so on without having a workflow for everything is amazing
@trying_out_ai glad it's sorted out, yeah no problem. Glad people like it, it does what I need it to 😅
v3.4 WF is so Good
using Chkpt600 Lora
8 Steps 1MP
from Native 15FPS to 30FPS interpolation+RTX UPScale took around 30minutes for 10sec clip
bcoz of interpolation fast movement's are not good
but slow motion face closeup shots are awesome, bcoz of 1MP
Config-48GB System RAM RTX 3060 12GB
AI Model Files are stored in 7200RPM HDD
OS and Paging File set to Nvme SSD
Yeah I added that interpolation node exactly for close up slow mo shots, glad you liked the WF! Just remember you can toggle everything on and off 👍 and if you want faster gen times, do a 0.4MP base and use the LTX upscaler, shaves off a bunch of time that a 1MP gen would get you from base H3 and removes that ghosting H3 gives.
@mrweaz yeah
i tried LTX Refiner but showing OOM error
also tried as a separate WF via copy Paste All of the LTX Nodes from ur WF
still same
@Yathukku strange, I'm running a 3060 12GB with 32GB RAM and I don't get any OOM errors, so with your extra RAM you shouldn't be either
@Yathukku going to update soon with a Chunking node that should help with lower VRAM 👍
@mrweaz tried just now the custom WF
No OOM Error
but not even 1 step complete out of 3
even after 15 minutes
waiting for the Chunking node
@Yathukku just updated, if it doesnt work, i dont know what to tell you, i have a lower spec PC and its working for me just fine.
@mrweaz v3.5 Worked like a Magic
everything is so Good Dude
and also Sage Attn and Sol Attn can be used simultaneously?
@Yathukku nice, glad you got it working, and as far as I'm aware yes, i haven't come across any issue using both. so you should be good to go.
I have updated everything and I get an error when I run the sol attention node because the dense block is set on false. when I deactivate the node or leave the value empty, I get a totally grainy video with white noise audio. I don't know if it's because of sol attention but other workflows work for me.
edit: I think it might be because I dont't have a reference image? does this workflow not have a T2V mode? when I turn off RFTV and ITV, then it doesn't work
edit 2: even with images, it just produces grain
i dont know what to tell you, it works for me and many others. for T2V you just disable any image input on the I2V node. have you changed any of the settings? like the Lora, lora strength, steps etc?
@mrweaz found the problem. it seems to be happening when using the unpruned int8 model, the 34gb one. any explanation for that?
@hakinn90 oh yeah so it may be incompatible with the Lora maybe? i only use the pruned one so i'm not 100% sure.
@mrweaz even with everything optional turned off, it's still the same grainy video. I think it has nothing to do with your workflow and more with the model lol
@hakinn90 yeah the workflow works, i've had multiple people on discord try it. usually issues stem from the models, updates or wrong node versions. hence why i put links to the exact models i use for it in the description and workflow.
@mrweaz it works with the pruned model, but now using the ltx refiner, my 5090 goes oom and the upscaling takes 175s/it. am I doing something wrong? I downloaded all the models and am using 10eros 1.4 int8 convrot, which has 27gb
@hakinn90 I may try a different model for LTX, I haven't gotten an oom issues, but it is a large model, let me okay around with some different ones to see how quality is impacted and I'll get back to you 👍
@mrweaz Thanks, I appreciate it. I just tried the RTX upscaler and it just doubles the seconds of the video, so slowing down everything. Idk whats wrong, I left everything at default from the beginning
@hakinn90 that would be the interpolation, if you enable the interpolation it will double the frames.
Brother, version 3.2 doesn't work anymore? I've been struggling for 3 whole days because my generations are black, it just gives me black generations and I even connect node in the Vae decoder to save image and all the frames are black, I'm using cross attention and Comfyui 0.31.0, but I don't understand anymore what the fuck is happening that the outputs are black. Besides, itinerant as it should be, something very strange is happening.
try V3.5, and update comfy to the latest branch not release, try the normal VAE, try not using INT8 models, try updating the nodes from github. its all in the description. the workflow works.
You really need to be more clear on your comfyui instructions, because you can't just saying "latest" when referring to github. "Latest" is literally a tag on github, pinned to the latest release (so 0.31.0 as of this comment), so saying "update to latest branch and not release" is nonsensical, since they're the same thing. Do you mean to update to the most recent commit on the master branch?
@Undecked yes the latest commit in master branch
if you are using stabilitymatrix, change version and toggle the Releases Button.
@mrweaz Ok dude, I'm going to try again Hahaha, I'm a obsessive, I need it to work at least 1 time to finish it, I won't use sage or use Triton.
With SageAttn, VRAM Cleaner, RTX and LTX enabled, a 0.2MP 15s generation on a 16GB card (and 32GB RAM) the H3 & RTX both succeed and generate videos, but the LTX step always fails with an OOM error. Am I misusing the workflow?
My LTX amplification process is very slow; what could be the problem?
I already have the environment, the recent update of branches, the most updated comfy, cuda, Pytorch and Python well updated, everything necessary for the workflow and still my generations are black, although it does a roaming process, everything is completely black, both frames and video LOL. I don't know where the mistake is, but, I don't understand anything anymore HAHA
Great workflow, though I am currently unable to get it consistently working. I've gotten a few good generations from it but it has been erroring out for some reason.
It seems to be something with this particular workflow, the basic right-out-of-the-box H3 template works just fine consistently.
Not 100% sure what the deal is. I imagine there is some conflict but I don't know what. The fact that I CAN get it to work only for it to fault out on the next generation is confusing.
ComfyUI Error Report
Error Details
Node ID: 735
Node Type: SamplerCustomAdvanced
Exception Type: torch.AcceleratorError
Exception Message: torch.AcceleratorError: CUDA error: an illegal memory access was encountered
Search for cudaErrorIllegalAddress' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with TORCH_USE_CUDA_DSA` to enable device-side assertions.
Stack Trace
File "H:\ComfyUI_windows_portable\ComfyUI\execution.py", line 550, in execute comfy.model_management.reset_cast_buffers() ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^ File "H:\ComfyUI_windows_portable\ComfyUI\comfy\model_management.py", line 1413, in reset_cast_buffers offload_stream.synchronize() ~~~~~~~~~~~~~~~~~~~~~~~~~~^^ File "H:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\cuda\streams.py", line 102, in synchronize super().synchronize() ~~~~~~~~~~~~~~~~~~~^^
System Information
ComfyUI Version: 0.31.1
Arguments: ComfyUI\main.py --windows-standalone-build
OS: win32
Python Version: 3.13.9 (tags/v3.13.9:8183fa5, Oct 14 2025, 14:09:13) [MSC v.1944 64 bit (AMD64)]
Embedded Python: true
PyTorch Version: 2.9.1+cu130
Devices
Name: cuda:0 NVIDIA GeForce RTX 3090 : cudaMallocAsync
Type: cuda
VRAM Total: 25769279488
VRAM Free: 24425398272
Torch VRAM Total: 33554432
Torch VRAM Free: 25034752
That's weird. I recommend you install Pytorch 2.13, Pytorch 2.9.1 is already a bit late. But I don't blame you, I've been trying to generate something for 4 days but I just can't do it, that is, the Ksampler does generate and itiner, but, when it decodes my videos and whole frames are black XD
i just dont understand how so many people are having issues, ive tried it on a clean comfy install and its working fine, others in Discord who have tried it say its working fine too. im kinda lost on how these errors keep coming up.
@mrweaz no clue. I'm going to try a fresh install. Perhaps there's something I'm missing on mine that I'm not considering. The inconsistency of the error is what's throwing me for a loop. Like I said, I can get it to run and it runs well enough to give me some good results, but it inevitably faults at some point and requires a restart. Doesn't always work, but sometimes it does.
@RoadRunner440 my guess is that maybe something is being kept in memory along the flow, ill double check to make sure LTX is unloading the models at the end etc.
@mrweaz Potentially. The error SEEMS, for the most part, to happen after one or more successful generations, so it does make some sense that something could be sitting in the pipe that needs to get unloaded and just isn't. As to what that could be, no idea.
When the interpolation node is on, the frame rate for rtx and ltx upscaled video stays 24 but it should be 48. is there any way to make it 48 automatically?
Where can we put other LoRAs in this workflow? I may be blind but I don't see anything except the turbo loader.
EDIT: Just updated the wf to include a power lora loader.
yeah i didn't add any lora loader for other loras, you can replace the current one for a power loader and add whatever you want.
In version 3.5, i2v doesn't work—only fl2v does. If you don't enable "last frame," an error occurs.
The same error for me too:
File "D:\GAMES\ComfyUI\python_embeded\Lib\site-packages\torch\cuda\__init__.py", line 1181, in synchronize
return torch._C._cuda_synchronize()
~~~~~~~~~~~~~~~~~~~~~~~~~~^^
torch.AcceleratorError: CUDA error: an illegal memory access was encountered
Search for `cudaErrorIllegalAddress' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "D:\GAMES\ComfyUI\ComfyUI\execution.py", line 550, in execute
comfy.model_management.reset_cast_buffers()
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "D:\GAMES\ComfyUI\ComfyUI\comfy\model_management.py", line 1413, in reset_cast_buffers
offload_stream.synchronize()
~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "D:\GAMES\ComfyUI\python_embeded\Lib\site-packages\torch\cuda\streams.py", line 108, in synchronize
super().synchronize()
~~~~~~~~~~~~~~~~~~~^^
torch.AcceleratorError: CUDA error: an illegal memory access was encountered
Search for `cudaErrorIllegalAddress' in https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html for more information.
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.
[INFO] Prompt executed in 6.96 seconds
Fatal Python error: Aborted
this means the gpu you selected does not support up to date cuda go into comyui and go to manager and click update all than install the custom nodes filter than click on install custom mission nodes after clicking the manager blue button select the top corner check box to install all of them click install than update after that and it should work is best to use a cloud gpu with cuda and rtx 5090
Since you're already using the Model Preview Override node in this workflow, you can add the H3 Spectrum node right before or after MPO for an additional speedup. It's compatible withLightX2V, too, if you're comfortable with a quality drop or compensating with additional steps.
Quick comparison with the same seed and prompt:
0.5 MP, 5s, res_multistep/simple:
- Base H3 with no acceleration, 20 Steps: 272.68s
- Spectrum only, 20 Steps: 169.07s
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 only, 8 Steps: 109.09s
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 and Spectrum, 8 Steps: 91.86s
1.0 MP, 5s, res_multistep/simple:
- Base H3 with no acceleration, 20 Steps: 642s
- Spectrum only, 20 Steps: 455.52s
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 only, 8 Steps: 280.10s
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 and Spectrum, 8 Steps: 236.51s
Notably, raising the steps to 12-14 with Spectrum and LX2V enabled takes about half the time to complete as H3 with no acceleration, and only incurs a marginal drop in quality and prompt adherence over base or spectrum only runs.
0.5 MP, 5s:
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 and Spectrum, 12 Steps: 129.15s
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 and Spectrum, 14 Steps: 156.00s
1.0 MP, 5s:
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 and Spectrum, 12 Steps: 313.64s
- LightX2V fl2v Turbo 8step v1.0 ComfyUI bf16 and Spectrum, 14 Steps: 351.99s