My ComfyUI workflows for using Wan 2.2

The end-frame (last frame) is the last frame, you can not inject it in the middle. If you want the last frame on second 5, you have to do a 5s video.
That's why it is called FLF2V -> First Last Frame to video.

Hey! You found a bug, it was running on default 81 frames (5s), I fixed the noodle connection. Thank you for pointing that out. v1.5 got the fix.

SVI will run with cfg 1, because of the guider, it is build like that. At least that's what they suggested. So no need to change that. NAG is not tested.

You use C-SVI, right?

@darksidewalker 

yep. the newest 1.5 vers.

@Ajaxdiffusion Would you mind join my discord and post what you are doing, so that I can elaborate and maybe fix, if there is an issue? :)

@darksidewalker hmm.. I'd need to create a new DC account, just killed my last one for different reasons ^^

But i just tested it again and now it..works? ._.
sorry for the time waste lol - no clue why it does work now

anyway..back to my first comment:
"Hey! Love the workflow - genuinely the best WAN workflow" :D

@Ajaxdiffusion well, glad it works for you now!

You have to install sage-attention before you can use it. The settings are not static, you can change them to your liking. Maybe you did activate some features that OOM you?

Also make sure to place the model files into the correct 'unet' folder now

@darksidewalker sage attention was installed properly before and works in other workflows. I didn't activate anything extra. In 1.4 i could make 0.83MP videos (yes i reduced it from 0.85 manually), in 1.5 i could only make 0.72MP, but this took 10x longer (with sage attention). And 1.6 - well i get this error, and it seems i'm not the only one, and I also updated ComfyUI and all custom nodes.

@vicaut Well, I identified tho problem with using gguf, with safetensors this does not happen.
I'll look into it.

@darksidewalker ok, i use GGUF :)

@vicaut fixed in C-SVI v1.7

@darksidewalker i have a sage attention error now. I set sage to auto: "Error running sage attention: Allocation on device , using pytorch attention instead." Sage attention 2.2.0 is properly installed (as always). The message appears at 0.72MP and above. But even at 0.6MP the speed is extremely slow with GGUF. I don't know what changed since v1.4, but GGUF is broken since v1.5. For me it is only working until 0.4MP with good speed. Did you change something in the memory management?

@vicaut You wont get the error if sage-attn would be correctly installed.
"Error running sage attention: Allocation on device , using pytorch attention instead." Is not an error from the workflow, it is an error from your backend.

The speed is low, because sage-attn fails and doing fallback operations, what is slowing down.

I can not change the memory management of comfyui with a workflow and did not use nodes that would unload or clean memory.

@darksidewalker as i wrote the error appears only at "higher" Megapixel settings for me. At v1.4 GGUF your workflow made an iteration in about 150 sec at 0.82MP, now i need more than 1000 sec at 0.6MP, 0.4MP is working fine with sage attention at 50 Sec/it.

@vicaut Because sage-attn / pytorch fails to allocate memory.
As I said, I did not change anything that would effect that.

@darksidewalker i also tried other workflows like https://civitai.com/models/2079192?modelVersionId=2606468. No special sage node there (i use it globally with --use-sage-attention), and i get the 150 s/it at about 0.8MP there with just your GGUF 8.1 and the SVI lora. It seems your workflow has startened to eat up memory.

I have sage-attention 2 installed, workflow C-SVI v1.7 and tested both gguf and safetensors with 1.05MP and got on both ~86s/it
No errors, no slowdown.

I can not recreate what your problem is, sorry.
-------------------------------------------------------------------------------
gguf qtypes: F32 (2), F16 (693), Q8_0 (400)

model weight dtype torch.float16, manual cast: None

model_type FLOW

Requested to load WanTEModel

loaded completely; 11425.55 MB usable, 6419.48 MB loaded, full load: True

[02:52<01:26, 86.10s/it] Interrupting prompt
--------------------------------------------------------------------------------------

@vicaut The Workflow you mentioned has sage-attention nodes inside. The same I did use before v1.7, he just disabled them inside the subgraph.

@darksidewalker yes i know. They are disabled, and i haven't enabled them. 

@vicaut Just an advice, normally you do not want sage-attn enabled globally in comfy, because it can mess with models that does not support it. Try always just using nodes to enable it.

@darksidewalker well it seems you have enough memory for this setting. But if memory gets low (I have 8GB VRAM) there is a difference. Raise your MP even further to maybe simulate my memory problems.

@vicaut On 1.05MP is starts low VRAM patching and my VRAM is completely consumed.

On higher resolutions I get OOM with sage-attn, because I do not offload to disk and I won't do it, that would degrade my SSD.

How should I do something about that, as I said, as long as I run it inside the normal specifications I can not see any problem. If I could recreate it somehow, I would search a solution.

So if anyone got an idea. Let me know.

@darksidewalker all i can say is that i just did a 10 second 0.82MP Video with the other workflow in about 35 minutes (no sage nodes activated), and i also could do this until your SVI v1.4 workflow, but since v1.5 your GGUF Q8 checkpoint @ 0.82MP is impossible on my 8GB VRAM + 32 GB RAM + SSD system. Whatever you changed from 1.4 to 1.5 makes it impossible for me to make videos at this resolution any more, i only can do to about 0.5MP. Also the movement got slow with the predefined settings (euler simple, 3 hi + 3 low steps, shift 8), but normal movement with the other workflow using exactly the same values, checkpoint, lora and prompts (only color shift after each segment there, thats why i don't want to use the other workflow).

@vicaut 
I tested v1.3 and v1.7, both with 1.05MP, same settings, same image. And I found your problem.
In v1.3 the frames logic was broken, resulting in fixed 81 frames, regardless the seconds or setting.
That is the point: 5s with 24 fps on v1.7 are by far more frames than just 81, that's why you need more VRAM and more time, adding up on higher resolutions.
If you want your old higher resolutions you need to lower seconds, fps or resolution.
It is working as intended.
You can proof this by looking on the frame-count of the produced video.

The slowdown you face is like this: 24fps in 3s (81 frames) is a huge speed-up compared to 24 fps in 5s.

@darksidewalker i reduced the FPS to 16 and it seems to work now (with the GGUF workaround), thanks for your effort!

There is no spacial reason, it just happened by growing.

I'm not aware of the restrictions of 'always one', I'll have a look, thanks for the hint.

If you like you could post your idea in my discord as json so I can check?

Thank you for the valuable contribution! I added the refined logic.

You can do this inside the backend (if kijai loader is used) or globally on comfyui itself.

i have the same is the GGUF switch that does that, I dont know how to fix it properly but if u go inside the graph and create a load gguf, link them directly to the sage attention it works

Use C-AiO v3.5 it is fixed there

@darksidewalker same problem in C-AiO v3.5

@balthazaredison551 Cant reproduce, more infos needed.

@darksidewalker yeah i don't know, I opened the subgraph and didn't see anything amiss. Using @floralis 's method seems to be a functioning workaround.

@balthazaredison551 3.5 is fixed on my end

@darksidewalker nice models a wf btw. thanks for making a gguf version

I have the same error (with GGUF) in SVI 1.8, not tested safetensors. Plugging 2 GGUF nodes directly to Sage works as suggested.

I have the same in 3.8. Will check plugging 2 GGUF nodes directly to Sage helps.

Looks like something with Any Switch (rgthree) not passing model but string.

Prompt executed in 0.12 seconds

got prompt

!!! Exception during processing !!! 'str' object has no attribute 'clone'

Traceback (most recent call last):

File "C:\SM\Data\Packages\ComfyUI\execution.py", line 518, 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:\SM\Data\Packages\ComfyUI\execution.py", line 329, in get_output_data

return_values = await asyncmap_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:\SM\Data\Packages\ComfyUI\custom_nodes\comfyui-lora-manager\py\metadata_collector\metadata_hook.py", line 165, in async_map_node_over_list_with_metadata

results = await original_map_node_over_list(

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

File "C:\SM\Data\Packages\ComfyUI\execution.py", line 303, in asyncmap_node_over_list

await process_inputs(input_dict, i)

File "C:\SM\Data\Packages\ComfyUI\execution.py", line 291, in process_inputs

result = f(**inputs)

^^^^^^^^^^^

File "C:\SM\Data\Packages\ComfyUI\comfy_extras\nodes_model_advanced.py", line 127, in patch

m = model.clone()

^^^^^^^^^^^

AttributeError: 'str' object has no attribute 'clone'

@lessi Yes. I identified a problem with the anyswitches from rgthree. I'm changing the loader logic asap. This only happens with gguf.

C-AiO v3.9 should fix that

Color shift has multiple sources:

* Too much frames / too long

* First and last frame very different

* Different resolutions

Color match needs much resources, if you oom you need to reduce video length or resolution

I had brightness/color flash at the end of the vids as well on FLF2V, turned out my ffmpeg install was broken, uninstalling and reinstalling fixed it for me.

Thanks, as I do not use T2V I won't make one, because I only make things I actually use my own.

check fps in basic settings

On my testing SVI works good with default 1.0

You have to activate the features on the feature switch, like it is described in the notes.
You misinterpreting the nodes, the logic works fine. They grey fields won't change.
You can use the 'preview as text' nodes next to the logic to display the status and see the logic changing if you want. Just click the preview node and run the node only.

Edit: My ComfyUi version is 0.10.0 all custom_nodes latest

@darksidewalker Yes I know to activate them on the front with the switches. I had a working build of this workflow working perfectly and then updated something somewhere which broke the if/else nodes for me. I was pretty sure the color correction wasnt working so wired a black image to the reference for a control test and sure enough it doesnt color correct unless I disable the if/else that feeds into it (in which case it always color corrects). Thanks for your help, the workflow is great!

@darksidewalker Is color correct meant to only work with perfect loop? It should be selectable in the non flf2v modes right? With loop mode disabled the loop mode split logic if/else never goes true so it always passes through the images that bypassed the color correction. See the image. https://ibb.co/HfxPn3Js

No idea why my setup is doing or behaving like this... Im thinking something is scaling different or some new update of something is changing the behavior of the nodes.... 

Edit - fixed this one, the input of the boolean should have been the anyswitch 1098 but was wired to boolean #1100 forcing it to always be false. This was on a fresh workflow I downloaded today for sanity. What's causing these breakages on fresh workflows for me, any ideas? This is an eldritch bug to fix, I can feel myself taking psychic damage.

@MiddlingMaker You really found a bug!
I did wire something wrong by implementing the logic, it was skipping color-match on some combinations. v3.7 should fix that.
Thank you for elaborating this!

I hope your psy-dmg is prevented with the v3.7 update :)

@darksidewalker Put a preview image on either side of the color corrector and force the color switch to sanity check with a dark image where the correction will be obvious... I swear the if/elses arent respecting their inputs. I had to modify my custom nodes and force them to accept disabled inputs. I kept getting an error that the if/elses couldnt run with a disabled input. Many, many hours later I custom coded my nodes to finally accept it properly (but still Im not convinced this isn't something with my setup).

Fix for this was to wire the #1099 on_false both with the color corrected output from #977, instead of the non-corrected output from before it. This feeds the "corrected" output into #1099 every time, but when the color correction section is disabled it just passes through the uncorrected frames.

Thanks for the sanity check, hopefully this was helpful and not burdensome and thanks for all your feedback. I'm learning a lot and having a blast tinkering away

@MiddlingMaker No trouble here. Awesome that you take the time to elaborate all this.
Helped me a lot to find this nasty bug. v3.7 works as intended, you can sanity check this the way you did with preview images :)

@darksidewalker One more bug I noticed on v3.6 (still present on 3.7), the GGUF mode wont change, pressing the toggle to on immediately disables it. Fixed by moving the overlapping nodes at the very beginning of the subgraph, one of the areas got moved or scaled so the disable/enable wasnt applying properly. Resizing the area fixes the issue. Still really new to ComfyUIs arcane installations - this could be something on my end scaling things improperly.

Default layout on fresh workflow (broken) - https://ibb.co/PvKpLHzC
Moved and resized (fixed) - https://ibb.co/0jz5yPRv

@MiddlingMaker They are not overlapping for me.
The screenshot you uploaded is not from v3.7

@darksidewalker hmmm, Ive verified it with another fresh download and import just now it always moves them on my setup. The ghost hunt continues :D

@MiddlingMaker The screenshot shows fantasy-talking nodes. I did not use them, I only used the native comfyui loaders, mine are not from a custom node, maybe your custom nodes override something.

@darksidewalker omg this solved the issues, THANK YOU. Something was installing FantasyTalk and a bunch of other nodes, for some reason it always thinks the native GGUF loader doesnt exist and pulls more custom nodes. Removed a bunch of custom nodes that were in the directory but not showing in manager and the workflow works perfectly!

I just downloaded it, run flf2v and everything works as expected. I do not know, what your problem could be.
Maybe you could join my discord and post a screenshot what you set. And please make sure to use v3.7

@darksidewalker I upload workflow to grok and ask why, this is grok's answer but I can‘t understand:
The error message "Loop (911,436) - not submitting workflow" indicates that ComfyUI has detected a circular dependency (loop) in your workflow graph between nodes 911 and 436. This prevents the workflow from being queued or executed, as ComfyUI's validation logic blocks cycles to avoid infinite loops or undefined behavior during processing.

Why This Happens

Node 436: This is a "Power Lora Loader (rgthree)" node (titled "Loras HIGH"). It loads LoRAs onto a model and CLIP, producing outputs like MODEL and CLIP.

Node 911: This is a subgraph node (custom type "4123daa5-14f2-4f1e-9343-aa619fdd9f9b", representing the "Backend and Settings" subgraph). It handles core logic like model loading, sampling, and video generation.

The loop forms because these nodes are mutually dependent:

Node 436 outputs a CLIP to Node 911 (link 2087: 436 output 1 → 911 input 4).

Node 911 outputs a CLIP back to Node 436 (link 2124: 911 output 2 → 436 input 1).

Similarly, for the MODEL: Node 436 outputs to Node 911 (link 2125: 436 output 0 → 911 input 0), and Node 911 outputs back (link 2127: 911 output 3 → 436 input 0).

There's a similar loop involving Node 422 ("Loras LOW") and Node 911 for the low-noise model path.

This creates a classic cycle: Node 911 depends on Node 436/422, but Node 436/422 also depends on Node 911. ComfyUI aborts submission to prevent runtime issues.

@NAGISAZUKI Well Grok can not understand the workflow, that is nonsense what it wrote

Thank you very much!

The checkpoint switches are disabled till the maintainer of rgthree fixes the anyswitch for gguf loader.

You have to do it manually.

@darksidewalker ok. Maybe you should make a changelog so everyone can see what is intended and what is possibly a bug.

@vicaut There is. Reading is key ;)

@darksidewalker i mean every change from the start - a changelog history. You report the change from the last to the actual workflow, but if somebody makes an upgrade from an earlier version he only sees the last change - maybe also here on Civitai too?

@vicaut I won't do that ... its just workflows not a critical software.

@scizzit Make sure to use v2.2

Make sure to enable the feature

Color match is not always recommended. If it doesn't work for the video you are doing, disable it

You can copy the upscale part and build your own workflow, that would be the way.

Update the workflow and don't use 5s with 24fps

Well yeah, but it is in since weeks now😁

24 fps on 81 frames are ~ 3s
SVI 2.0 Pro wants exactly 81 frames, as written in the workflow. Or it will color shift or degrade.

Try C-I2V v1.7 and make sure to test with all features disabled first.
If this does not work, try the backend test and make 100% sure you installed all dependencies.

No Florence is trash for NSFW and not worth it, if you like it, just wire it in.

@darksidewalker Your not wrong there but it deffo helps with defining a character adding more details.

I tried Florence many times and it doesn't help. It doesn't understands explicit stuff and the description from the setting is in almost all circumstances not helping and wrong or too much, destroying the scene. 1-2 self made sentence are in 99% of all cases just better. Use it, but I'll not implement it, it made all tested results worse.

Not really. The NSFW t5 has no real impact compared and mmaudio has no use except some lewd noise. NAG is almost incompatible with svi and raises VRAM consumption massively.

But there is also a version with better precision: https://huggingface.co/zootkitty/nsfw_wan_umt5-xxl_bf16_fixed/tree/main

@vicaut I don't get what you want from me, if you want to use it download it and just use it.🤷

You can ignore that message, it is just from the comfyui native loader, no relevance

Nevermind I solved it.It's vae issue and It's fixed now.

@hellbergdk2 你好，我更改了VAE，但还是在运行了0.02s后结束，仍然显示

Failed to validate prompt for output 911:1180:

* easy ifElse 911:1169:

- Required input is missing: on_false

- Required input is missing: on_true

* easy ifElse 911:1168:

- Required input is missing: on_false

- Required input is missing: on_true

同时”SageAttention split logic“卡片中的两个Ease-Use的是否判断阶段显示红色报错。

请问你是怎么解决的呢

Workflow "C-CA" can be used to combine videos made with single i2v parts.
The eyes part, maybe your resolution is too low or you are using strong loras. Prompting the eyes help.

@darksidewalker Thank you very much.

You did not select any valid file for the checkpoint.

Hmm odd, this checkpoint worked on your old workflow. I've also tried your SynthQ8 Guff checkpoint and it shoots out this error:

AttributeError: 'str' object has no attribute 'clone'

The checkpoint works with my workflow. I don't know what you are doing.

I had this issue when I was using the C-Aio v3.8 workflow but upgrading to v4.2 fixed it. Hope it'll fix it for you too

Ah yes... turns out ComfyUI loaded the old version
Redownloaded the newest version and now it works

Sorry for the misunderstanding the error was related to the older version of this workflow

Works perfect now, love this workflow, its so clean

This idea might be worth to look into, I notices that behavior too, if the face is not always inside the segment it renders. I'll look if this is possible for me and how it works, if you manage to get it in, I would love to see a contribution on my discord if possible.

I've been experimenting adding the simple face restore workflow to each SVI 5-sec clip, and the issue I am running into is that the way the workflow is done it combines each video in sequence, so 4 videos would be 20 seconds (4x5s clips stitched together), but the issue with applying ReActor to each individual video node is that it overcooks the earlier videos because it renders the face restore fix over the older clips again and again. If there's a way to have the workflow individually do each 5 second clip and then apply face restore and then stitch them together and use the last frame of the previous frame (with face restore applied), then maybe that would help with keeping faces looking as close as possible to the source image.

You are missing dependencies. Make sure to install all custom nodes correctly and use the latest workflow.

Same issue, ive installed all custom nodes and using 4.2

C-I2V 1.7 works, I get the same issue as OP on C-AiO 4.2

@aawti I wonder, the loader logic on both is the same. Except I2V has no flf2v. Without screenshot of workflow parts it is difficult to imagine what is happening.

@darksidewalker Strangely, when I used the workflow provided with your example video, it worked correctly and reproduced the video effects.

Thank your for this kind words! I wish you awesome fun and luck with the seed-god ;)

That depends on your settings (resolution, steps, seconds, features ...) and how you set-up comfyui. The workflow itself has no requirements.
I read ppl running it with 32GB RAM and 8 GB VRAM.

@darksidewalker Hey thanks for the response! I think i figured out the issue last night. I was loading in the safe safetensor high low files, which i think require +28gb of vram, but i only have 24 gb of vram. I learned to use the different gguf files to save on vram.

@thespy that can not be true, the safetensors do not allocate 24 GB VRAM at the time given. Low and high are used independent and in succession. Also gguf will allocate more VRAM with Loras and be much slower. Just saying.

@darksidewalker Huh.... Do you know why the synthseduction high and low q5 gguf would work but not the safetensor of synthseduction causes a vram on the same settings? Im new to video generation so i assumed the safetensor was using more vram compared to the gguf

@thespy well Q5 need less VRAM, but is like for 8 GB VRAM cards.

@darksidewalker I see, how much vram is needed to run the safetensors for SynthSeduction? https://civitai.com/models/1981116/dasiwa-wan-22-i2v-14b-or-lightspeed-or-safetensors

Im running some tests to see if i can diagnose why the safetensors are causing a OOM on my end. If that doesnt work, could i DM you the workflow of yours i downloaded? I feel like im missing something obvious...

@thespy depends on the resolution of the video and length. 8-24 GB ...

@darksidewalker  I'm on the using 3 seconds at 480p (0.40 MP - SD Speed), with 4 steps, refiner at 2, and cfg of 1.
I feel like that shouldnt take 24gb of vram, and i make sure to close everything else out on my pc to rule that out.

@thespy You are right that shouldn't need 24GB VRAM, there is something off. I have 16 GB VRAM and can run 1.05MP with 5s on 16fps

@thespy If you like join my discord and we can have a look on the help channel.

@darksidewalker Yeah i can, my discord got hacked a week ago so ill try to get that back online through support. Thank you!

For others: i solved the VRAM issue by upgrading to pytorch 2.10 with cuda 130 ( was on 128). I was reading the logs and it said cuda130 had vram optimization and that got me over the OOM issue. The next issue i had was i was running out of RAM. Strangely that was happening after the Ksampler (low) finished and the clean vram node started. I did a test where i removed clean vram node and my regular RAM no longer went over 32GB. The last thing i did to fix it was i launched comfyUI with the params: comfy launch -- --lowvram --use-sage-attention --cache-none

I think you are missing custom_nodes or they are not installed correctly.

You do not need to unpack it, as described on the notes, just press "3". :)
It is inside the graph, because it is a one time setting, you normally do not alter after setup.

@darksidewalker I'm just blind apparently lol

Yes it shows there in the node, but I don't use runpod, so I don't know this work there.

No worries, thank you for the reply. Keep up the great work.

@Fatty_McDuffins you could try to simply put a "LastFrameExtractor" Node on your last VAE node's image output (the point at where your final images are rendered) and put a preview image node on the extractor node.

it doesnt save the image in a folder, only temporary in the UI. just save it and reuse it for your next run. thats how I do it ^^

hit me up rq if u need help to set it up

I looked a bit more around and looks like i fixed it. I don't know if it's supposed to be like that but i just changed the wiring in the "basic settings" graph from "after_intermediate" - clear vram/cache - color match to "extended images" - clear vram/cache - color match.
So i had to get the extended images and not the output from the "Video Intermediate" video combine.

@H___  you are right, I got some wires mixed ;) Will upload a corrected version asap.

C-SVI v2.4 should fix that

You did not set the checkpoint file correctly

@darksidewalker like theyre in the wrong folder? I had them in diffusion_models I believe

@Overondid the folders should be used like described in the workflow notes and properly selected, like described in the workflow notes.

@darksidewalker thank you so much for your help! I was able to bypass this issue. Of course now I am having another issue: The error reads: Power Lora Loader (rgthree) Error while deserializing header: header too large. Any ideas? I have tried updating, restarting, reinstalling the rgthree node to no avail. lmk if you want me to post the whole log. Thank you again for your time

@Overondid The lora you load is faulty

@darksidewalker great info, Ill see what i can do

@darksidewalker seems I had to redownload the SVI loras I downloaded. Thanks again!

Apologies, the workflow is defective. I'll look into it!

Please use C-CA v1.8

I saw this in other comments, and it solved my problem.

For anyone encountering this issue, it's because we're on a Windows system, while in Linux, the "/" is reversed, so you have to manually change it and find the vae dropdown menu. This workflow uses a bookmark-like system, so look at the top middle, click the bookmark symbol, then click the checkpoint section, and manually change vae there.

@laogetmd862 Well, you can just click on the front workflow settings to choose the path.

Please read the workflow notes where to put the files.

You can use whatever you want 5-10 are valid

same trouble

I have the same problem. It was working fine yesterday

It is working on my side and I did not change the logic there, so it is possible a problem on your end.

@darksidewalker Having the same issue Failed to validate prompt for output 958:

* easy ifElse 911:1003:

- Required input is missing: on_false

Output will be ignored

Failed to validate prompt for output 644:

Output will be ignored

Prompt executed in 0.62 seconds- i just loaded it up and put all the files in - cant figure out why - LTX2 again tonight I guess until there is a fix

@OpNickerz Really no clue, I'm on comfyui 0.11.1 and pytorch2.10+cu130 and all custom nodes updates and everything is working.

If anyone find the answer, please share

May you try C-AiO v4.5?

@darksidewalker 4.5 is work fine for me

@darksidewalker you are a hero - thank you

It is working on my side and I did not change the logic there, so it is possible a problem on your end.

@darksidewalker I thought its an error with custom nodes and tried switching to earlier versions. No luck. I2V-17 Workflow working fine though

@XtreamRish Because that nodes are missing there, I assume.

Same issue here, Failed to validate prompt for output 958:

* easy ifElse 911:1003:

- Required input is missing: on_false

Output will be ignored

Failed to validate prompt for output 644:

Output will be ignored


I2V and FLF2V split logic node id #1003 errored with red box
LORAS + BASIC SETTINGS node id #911 red boxed also, verified all paths correct for models etc. Note that VAE not selectable from here, only in subgraph, intended?

Have copied/pasted node #911 from alternate workflow, C-I2V-17 and reconnected all outputs etc. to that node pasted into C-AiO-44 and it works fine. So the absence of VAE and CLIP selections from #911 in C-AiO-44 is perhaps the issue.

@PD63 If you did that you transformed AiO into I2V alone and replaced the complete backend ... that's not a fix. So you could just use C-I2V... that makes no sense.

@darksidewalker right-o, back to you to fix it then, it is certainly broken. Was trying to help. Good luck.

@PD63 Well you say that, but for me it is working 100%. I have no hint what is causing this for some ppl.

May you try C-AiO v4.5?

The workflow has been updated to v4.5.
It fixed the issue for me. Try downloading the new one and see if the issue still persists :)

@darksidewalker Yea, I get you, need something broken to be able to fix. Understood. Working now in v4.5, so once again proving you are the best 👍

Workflow is updated and everything is working now. Thank you very much for the fast fix @darksidewalker

May you try C-AiO v4.5?

The workflow has been updated to v4.5.
It fixed the issue for me. Try downloading the new one and see if the issue still persists :)

What workflow? There are 7 ...

@darksidewalker I have the same error but I don't know if it's the same issue. Both the AIO and the I2V. Actually, I used your image workflow from the V9 SynthSeduction model first, which is the AIO too right? I don't really know what I'm doing but there seems to be a purple border on each leading node in the backend. There seems to be an issue with rgthree and Fast Group Bypasser nodes too. Mine loads in with blank/broken nodes and on reload only populates the Easy Switches all janky. I don't know if that's related. I'm able to run another Wan 2.2 workflow for I2V so I suppose it's got to be a custom node somewhere causing the issue.

At the very least, I think Wisels and I have some sort of dependency missing. Either the wrong version of something for rgthree/ other customs nodes or something else.

"Loop detected 911, 436" seems to point to "back end" and "Loras High" nodes if that's any help, but I think it's just borked on our end.

I'll eventually go learn what this backend feature/node is, as it's new to me.


Filepaths done
Comfyui 0.11.1
PyTorch 2.9.1+128cu

@0Rookie0 i have the same issues with the AiO and the I2V did you figure it out ?

@darksidewalker it was in the 4.4 workflow now in the 4.9 the loop is 26,29

@Wisels I don't know how to solve your problem. Maybe try my installer, it should solve all dependencies. If you like.

@0Rookie0 Did you ever figure it out? I'm having the same issues.

What do you mean by landscape?
The workflow does not compress anything, it calculates fitting mega-pixel and scales the image perfectly inside. There is no possible known way to me this could produce borders.
As example even 21:9 makes absolute perfect videos, without any deforming, stretching or borders.:
https://civitai.com/posts/26296658

@darksidewalker I did a test to be more precise and the problem is that:

When I use a base image of a perfect 16:9 format 1920x1080

I selected SD+ quality (without the upscale), I get a 1040x592 video. so it doesn't "create" border but when I put it in capcut it doesnt fit well the 16:9. And if I continue to extend:

I extract the last image and for the 2nd video I get a 1024x576 video.

And then every time I will extend It will alternate between a video in 1040x592 and 1024x576. And in capcut that doesnt looks great.

Forcing the resolution manually fix the problem so I will stick to the old workflow for now but here is an exemple:

https://civitai.com/posts/26301367

@ShortLove I see. The problem is you are using different resolutions for the initial input and successive inputs. That will lead to different outcomes.
I understand that this is not what you want to do, but that is not a problem of the workflow itself. And may not change, since I really like the automatic resolution optimization.
I did not invent the workflow to be compatible with capcut.
It is made in mind, to input an image and get the best possible result as single video.
For longer videos there are better alternatives like SVI 2.0, so you do not need to chain manually.

That said, I got an idea, for better chaining or unfiltered resolution, a switch for 1:1 (no scaling) resolution. Maybe I'm going to implement that for such things.

C-CA workflow for joining

@darksidewalker Thanks. Will give it a go.

I don't know what I did, but it suddenly started working.😯

Thank you!
Indeed some settings got hidden,... But I'll correct that and the node description for the loras.
A flash attention toggle is not inside, but you can just enable it globally in comfy.

C-AIO v4.8 should fix that hidden settings

1# Update to latest C-AiO v4.8, fps setting was missing because of a bug

2# 10s are too much for WAN 2.2 without spacial setup like SVI

Great, thank you!

What duration do you recommend for video generations? 5s for best quality? And then I guess I can export last frame to generate the next segment and stitch them together. I'll read up on how to do that, but any advice appreciated.

Update: Oh, i was just reading about SVI - so I'll try that too and see if it works well. Any recommended SVI LoRas to try?

Love what you're doing here btw!

@hansomething I have the workflow C-SVI, you can start there. The loras are described and linked.
5s is the standard WAN 2.2 supports.
I also made a workflow for combining multiple single i2v videos C-CA.

So just have a look and thank you :)

The wf you mention is for combination or adding audio, it will not create any video. So I cannot imagine what are you doing.

@darksidewalker @darksidewalker 

Of course not. I mistaken those names. sorry. its DaSiWa WAN 2.2 i2v FastFidelity C-AiO-46. Spent another 4 hours try to solve with gpt why and how to fix those artifacts which add like lighting to 1-3 frames near before loop end... It is like it doing for no reason.

@Onheek I have not a clue, but did you try to update the workflow to the latest v4.8?

@darksidewalker I just actually found it.. It was the last thing on my mind.. so I test it at last... It making the model. SyntSeduction9. When I use Tasty sin 8.1 or other model, no more artefacts when check FLF2V.... after I downloaded the official workflow from FLF2V and it did the same thing, it occurred to me that only think I didn't try to swap/change is the same model DaSiWanv9 .... it doesn't make sense, but it works now. Thanks :D

@darksidewalker + update + Confirmed. I tested even new WF and yes. Confirmed. FLF2V making lighting artifacts ca 5 frames before end in about 75% videos, and its only model SyntSeduction9... older model 8.1 is ok, no white/lighting artifacts. Just for info... anyway, thanks for answer and those amazing works

I can confirm the behavior now, but why the issue may happen... still lies in darkness to me.

Any update on this?

@GaXxO11 No clue, I tired with just i2v and loop and it does not happen, with flf2v it sometimes happen. maybe a sampler thing, I did not test. I almost never do loops.

@darksidewalker Okay, luckily sometimes it does work, so you just have to keep trying and trying. Thank you.

i have the same problems.

It startet with wf 2.4 i think.

Currently im using 2.3 without issues.

I always do the same.

Settings like Vae,checkpoint,clipencoder,video output nvenc av1.lora(svi high/low every stage).i always used all wf since 1.8 never had issues and i know what to set up each "new" worklfow.

I tried gguf and safetensor. There is something wrong with this svi wf but i dont have the knowledge or time to find the issue.

Comfiui was updated to 11.1 and a new 12.0 is out today and i dont know if the problem was comfiu itself i will try it on the weekend.

@hartweizen Any hint or help would be of value to me. I can not recreate the issue on my end and I can say the intermediate video is not the problem, it just greps the image stream before any secondary feature will be applied. It is just a standalone output before running other features.
I updated to comfyui v0.12.2 and will try again.

On comfyui comfyui v0.12.2 I can now reproduce the problem, but I'm not yet sure what causes it or if I can do anything about that...

@darksidewalker I couldn't find a solution but u can reupload workflow 2.3.that workflow definitely works with the latest comfiui.

I did compare wf2.3 with 2.5 and did try few things but it is maybe a hight/width issue.

I asked google gemini but only thing could be the resolution option.

Source image(anchor) to prev samples mismatch.

But it did work with workflow 2.3....

I will try again if i have time.And thank for youre hard work ;)

Oh, I'm seeing the same issue. Following this thread to hopefully know when there's a solution

C-SVI v2.6 should be fixed from ghosting.

@darksidewalker I will try it in ~10hours.what was the issue?!

@hartweizen clip and condition racing by the global prompt injection ... I assume.

@darksidewalker 

I tested it. I'm using Comfyui 11, since I'm on MimicPC. The ghosting seems to be gone. Great job on that!
At first the colors were al washed up and the image degraded over the 5 seconds... Then I noticed the SVI lora is not there by default. Once I put it, all is working.
I'm thinking on doing some quality of life changes, like using the rgthree seed instead of the eay one. That other one gives more control I think. I'll share it with you once I finish the changes.

You ether did not install the custom nodes or you did not download the model files, I guess.

I'm not sure how the workflow could influence the camera motion, at least I'm not aware of anything that could do that.

@darksidewalker I actually have a similar problem, all my SVI videos camera slowly shift left each segment as it generates. I tried prompts to control the camera, but they didn't work. It doesn't really bother me much, but I thought you should know. The problem persists with your latest SVI 2.6 workflow.

@ferelik440 are you sure all Loras are setup on every segment?

@darksidewalker I also encountered the same problem. When using the V9 version, the camera would always automatically shift to the left. V8.1 doesn't have this problem.

@darksidewalker Me too, which is why i was confused. But I can run the exact same thing on the older workflow and the camera movement is normal.

This could be completely wrong because I'm not super knowledgeable with comfy, but the only thing I could think is maybe it's something to do with the auto resolution calculating system/nodes you have on the newer workflow, since it gets progressively worse the higher the resolution.

I got what you facing there, that is happening with SVI slightly zooming and because SVI starts from the last couple frames it shifts and shifts.

But only if I prompt nothing ... If I put any basic prompt and person description it stays on the person.

@darksidewalker Oh yea, I keep it simple, only use one or two Loras. smoothxxx or dr34ml4y.

It is hardcoded 81 frames, because it is a requirement from the SVI team.
You are right 24fps with 81frames ~3s
16fps with 81frames ~5s

Yes, update your comfyui

@darksidewalker  but now it just crashes midway through the actual generation
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16

Found quantization metadata version 1

Using MixedPrecisionOps for text encoder

Requested to load WanTEModel

loaded completely; 6419.48 MB loaded, full load: True

CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16

and then just stops and the job is removed but comfui stays open, i have 64gb of ram and a 3090 so?

I can not tell you what is wrong on your configuration or what workflow you use. The model works fine.

@darksidewalker Someone said "For comfyUI portable, updating ComfyUI is not enough! The embedded python version must be 3.12, which is never updated via ComfyUI manager, for example." Is this only for the desktop version? I keep running into problems with the portable. I want to use SageAttention. Should I install the missing custom nodes manually from your links, or use ComfyUI Manager?

@hakusei837
1# Installing a correct comfy + sage-attention is not easy, no way I write a guide here.

2# Updating custom_nodes work both ways.

3# Python 3.12 is needed for some custom nodes and some bleeding edge features need 3.13 - So there is no way to tell you what is best for your use case and what you try to do.
Also, yes, comfy will not update python itself.

FLF2V + Perfect Loop

if u placed the model right now while comfiu is open u have to klick the R button for refresh.
if its stil not there u placed them wrong.u need to place checkpoints(safetensor/gguf) in the unet folder not in checkpoint folder.

@hartweizen unet is the problem, thank you!

As the default is fixed seeds it will generate the same thing without resampling if you just don't change anything you can add step by step and sampler, if you do not change the settings it will do the same and add the next. So you can only change the last till you are satisfied with the prompt.

@darksidewalker Oh my, this is so obvious. I never actually tried not changing anything in the prompt haha. Thanks for the help. And many cudos on the fantastic workflows, checkpoints, loras, ...

als google gemini ki for 2x upscaler for svi or wan 2.2 i2v and how to install. be aware that u might ran out of memory if u are upscaling 2 or more stages in svi pro workflow as example.

In the description notes inside the workflow is a link to modeldb, you have to download one or from civitai, if available. The predefined model is also not downloaded, it is there because I use this. The downloaded model has to be in the folder "upscaler_models"

You are talking about C-SVI workflow?

@darksidewalker yes :) i did try few times now and refiner 1 is better for the face stability.if i use 2 or more it shifts over time after 2 or 3 stages in svi workflow. try it out :) and beta seems to be sharper for faces.

@hartweizen I'll try!
The implemented upscale are already very resource friendly, but if you download an other model from modeldb, you may save some VRAM, but the lanc-upscaler is already the cheapest you can do, if this OOMs all models will probably.
1.05MP are already very much for SVI to handle, they recommend 480p, what is like 0.4MP.

@hartweizen Can you provide a link to Refiner 1 or the full name so I can download it?plz

@mmikemiller823390 use the svi workflow and change the settings and change .fps24,shift high/low 8,sampler:euler,scheduler:beta,stepstotal:4,steprefiner:1.
darksidewalker uses 4-6 total and 2-3 refiner and i use 4/1.render times are faster(because less steps).

@hartweizen you use 4 step for high noise and 1 step fot low noise? Right? And euler/beta 24 fps shift 8, right?

@avavryns165 the sampler where u can change the fps.there is a total step value and refiner value.set steps-total to 4 and refiner to 1.the refiner literally adds new details to the face on every stage thats why faces shift over time.if u set refiner to 1 he keeps the source image.and change scheduler to Beta that also helped for sharper faces.

@hartweizen Sorry for the stupid questions. I understand the mechanics of work. The only thing I don't understand is maybe stupid me =) where in the samplers is the time meter "refiner"?

@avavryns165 if u open the workflow u should already looking at the sampler.its in the big blue box(basic settings) and below those word thats where u change all the settings in the backend.u should see fps,shift high,shift low,samplername,scheduler,steps total,step refiner. just klick and change them.

@hartweizen thx

@hartweizen Thank you. Indeed, the result has become better, and the generation speed has increased. Is there still a problem with color match. After 30 seconds of total video in HD resolution, when passing through the color correction node, everything freezes and gives an error in local comfyui.

@avavryns165 30s are many frames to color-match, I bet you just OOM

@darksidewalker And if you insert a color match node every 10 seconds. And as a result, there will be at least 2 nodes of video for 30 seconds. Will it work well? Or how best to use. Otherwise, by 30 seconds, the color fades a lot relative to the reference.

@avavryns165 you would have to build it between every sampler node, but I assume it will not work well for seamless transition if colors are altered in between.

@avavryns165 It will colorshift i did try with 6 stages and fhd2.07 once but the results are not worth it.OoM is always the main problem.only higher resolution compensate colorshift the most but the video will have less motion....

If u have nvidia rtx5xxx use sage attention 2.2 with cuda 13.0.Sage attention 3.0 doesn't work with wan2.2.

Maybe if wan 2.5 (local) releases we have less issues,but nothing official so far if and when it will happen....

Only if you implement this yourself, at this time.

Also when I start the workflow without perfect loop and upscaler, it's just stuck on the "Backend and Settings" node and it stays on 0%

UPD:

I've managed to fix the issue with Perfect Loop and Upscaler. It turned out White Rabbit hasn't installed "torchlanc" with it

Now only the only issue remaining is that the workflow itself is stuck -_-

Maybe your comfy is just broken or missing dependencies.
If you like you could try my installer:
https://civitai.com/models/2364056/tool-dasiwa-comfyui-installer

@darksidewalker so I just delete the current ComfyUI with the Windows built-in uninstaller and install your version?

@darksidewalker It's working with your installer. Thank you very much! You've saved me a ton of time and nerves!

Enable the switch for gguf, select the checkpoints, read the notes. GL 👍

@darksidewalker Hello, I placed the gguf model in the (ComfyUI \ models \ diffusionmodels) folder, but it seems that the node did not read the gguf model, only the safetensors model. I have installed the gguf plugin and set Fast Groups Bypass (rgthree) to gguf, but restarting Comfyui still cannot find the gguf model in the options😭

@voidyear unet folder. Just read the notes.

@darksidewalker My mistake, thank you very much for your reply！😃

@darksidewalker I put both safetensors and gguf models into the UNet folder according to the comments, which makes both types of models unreadable. In the diffusionModels folder, only safetensors can be read, which is too strange🤔

@darksidewalker Maybe it's because my desktop comfyui shared the model folder path from other files. I'm trying to solve this problem, sorry to bother you

@voidyear you are doing something wrong, all model need to be in unet for unet-loaders in my workflow.

@darksidewalker I created a separate GGUFLoaderKJ node and found that it can read the gguf model, but I was unable to successfully read the gguf model from your workflow's (Backend and Settings) node

@darksidewalker I found that it is because UNetloader is used in Use SAFETENSOR, and this node seems to have a defect that cannot read gguf. Only Unet Loader (GGUF) can read gguf

@darksidewalker The problem has been solved. The gguf model parameters are on top of the node, not below, which caused me to not find them....

@voidyear Okay, so the gui must have moved them somehow, normally they are below.
But please dl C-AiO v4.9, because it fixed a lora-model wire.

@darksidewalker ok,thanks!

Yeah, thank you for reporting! No clue how this could happen. Fixed in v4.9

Yes that is normal. It will be 10 if you reduce the fps to 16.

@darksidewalker ok but action will be slower I guess?

@yannberg240398 Sure, or you go with 24 fps and get 6s

My understanding is that Dasiwa fixed the WF for 81 frames / 24 fps as per SVI official requirements. So every sampler is 3.75 secs.

@Elmer588 Yes, that's correct.

I see you seem new to WAN 2.2, I suggest you to read my guide and the notes inside the workflow.

It's loading the WAN 2.1 VAE, what is correct.

@darksidewalker I read it, i even added correct model to vae_approx folder for wan21 - lighttaew2_1.safetensors as it asked for, but in preview i get just single image not video. How do I fix that ? i ask about that, otherwise model is great and fast, and i get nice outputs but it would be nice to see current preview before final video.

I don't know what i'm doing wrong.

Thank you

@spavlek9491 don't use lighttaew2_1.safetensors, it's trash, just use the regular one

@darksidewalker but it asks me for this. if i don't include it then i can't see anything. can you please point me to a file where to find ? taew2_1.safetensors i have this also in my vae_approx folder.

Warning: TAESD previews enabled, but could not find models/vae_approx/lighttaew2_1

i get this error

Thank you

@spavlek9491 you may read the wiki or docs of comfy for that feature, that's not a workflow feature, that's a global thing you have to do for comfyui.

Than install all dependencies or use my comfyui installer

Just choose your output folder.

@darksidewalker I'm sorry but I'm not sure I understand here.

The video I created with AiO was saved to 'ComfyUI/output/video/2026-02-12/' with the name '233850_00001.webm'

So for C-CA I need to write 'ComfyUI/output/video/2026-02-12/233850_00001.webm' into the video_audio field, right?

This also throws a 'not a valid path' error.

@Milvidi You do not have to write, just click the field and select your path

@darksidewalker I can't seem to 'select' anything. When I click the video_audio field a 'Path' screen pops up with your default 'output/video' in the textbox, and a OK button next to it.

So here I have to either type or copy the path in the textbox, then it'll throw the error when I try to run.

@Milvidi When I click into it I can empty it and select the path

If you referring to the previews videos, they always just show the segment, the complete video is rendered when all segments are rendered.

@darksidewalker yes, I mean the final long video, for some reason it shows the jump. I know I must be doing something wrong because your other workflows worked great, but I can't find what :(

Not possible to get the exact same. Every setting will change outcome a bit

@darksidewalker ok! thanks man.

Yeah, only exact same settings and seed will produce the same outcome.

@darksidewalker Every settings is same. I even changed "control after generate" to "fixed" in the "Backend and Settings" tab.
After multiple tries I've noticed, that stage-1 preview 360p quality has amazing good pace of animation, understanding the prompt very well.
But when I change the quality to stage-9 HQ 900p quality (while keeping all the other settings same , seed same ) the animation is slow and different.

@learnrijo As I said only exact same settings will produce exact same output. Resolution is a setting, that is valid for all video generations.
Higher resolutions on WAN 2.2 will always cause slowdowns, it is always this way.

Add focus and or camera description to the prompt to prevent that, could help.

Many thanks

I managed to improve the issue by switching the sampler_name from euler to uni_pc_bh2. Although there’s still a slight drift to the right, it’s much less noticeable now. Previously, even with a fixed camera, the character would constantly drift toward the right side of the frame, which was especially prominent when generating long videos

"static camera. still camera." produces absolute static camera for me.

Yes, but you cans et whatever you want

3 hi and 3 low are even better

@vicaut Depends, what you do. more steps high than low is good for anime sometimes.

@vicaut it was my suggestion. refiner higher then 1 alters faces overtime in svi.u can always change what u prefer.currently im using 24fps,8,0,euler,sgm_uniform,4,1.For anime its diffrent.faces will still alter if they are out of the frame for to long but its way more stable with refiner 1.

yeah, @hartweizen is right, his suggestion works. Also I did some experiments with samplers and schedulers, they have major impact on outcome and motions in different situations and styles, so experiments are valid.

It can also help to add in every segment something like: "(consistent face:1.1), same identity, (same facial structure:1.1), same proportions, same hairstyle, same makeup, same expression style, identical eyes, identical lips, identical teeth, identical nose shape"

Because flf2v needs more VRAM than i2v

@darksidewalker that not true, i can use other WF without any issues.. I asked chatgpt, the problem is on whiterabbit nodes.

here the answer: unloading the model before running WhiteRabbit can absolutely help.

For most GPUs under 16GB:
✔ Generate
✔ Save
✔ Unload
✔ Interpolate

can u look at this?

@RavirKun Well, if ChatGPT knows the answer, just do it :)
But I can say you, the answer is not correct anymore, because the clean/unload nodes harm the comfyui memory management and are leading to errors as of comfyui v0.4 and higher.
But if you want, just wire these nodes in between yourself.
You can also tune the whiterabbit nodes for lower batches, tune as you like.

C-SVI has no loop_count 🤔

The loop_count is in the Video Combine nodes.

They have no effect on the final video. The images are passed directly from one extension to the next and the intermediate saved parts aren't referenced at all.

The video combine nodes by the samplers are there as reference so you can check the results while the next part is running. They also provide a 'backup' so you don't lose everything if the workflow fails part way through (out of memory, etc). Increasing loop_count there just makes the backup video loop and has no effect on the continuing workflow.

The video combine node up top on the left is the pre-final intermediate node which saves the video before upscaling/colour grading/interpolation. Again, looping that just loops the backup video and does nothing to the result.

All intermediate parts are deleted at the end of a successful run.

The final Video Combine node on the right is the end result. Increasing loop_count on that will... just make the final video loop. Most video players will loop automatically so... yeah.

if u use his workflow for the first time.u always have to change safetensor or gguf files high/Low(unetfolder),clip,vae 2.1.otherwise u will not be able to use it.all files in his workflow a linked to his computers filepath and not yours.

@hartweizen I did that and I'm still getting this loop error.

@hartweizen yes, I was able to see that and made sure each was pointing to something in my file path and I am still getting this loop as well.

just use AiO version thats always up to date or svi for longer version.AiO is the same but more features.

@hartweizen I get the same issues with the AiO version as well, everything updated.

@BinaryBottleBake wich comfiui version are u using? im using 13.0 and there is a version 14 and 14.1 im using windows installer version.maybe its a new bug for 14.x and check for update for modules(custom node manager).did this issue appear since 24 hours if yes then comfiui 14 is bugged again...

@hartweizen Using 14.1 right now, and all my nodes recently updated as well. Keep getting Loop (911, 436) backend and high lora node

@BinaryBottleBake Using my comfyui installer, latest comfyui 0.14.1, latest nodes, torch 2.10.0+cu130, workfow C-AiO v4.9 ; everything works.
Maybe you installed custom nodes that are incompatible.

For anyone curious and sees this thread, it definitely seemed to be some custom nodes that were causing some conflict. I removed all nodes and reinstalled comfy and after adding the requirements this became usable for me. It's a great workflow so it was worth getting rid of the extra nodes I don't use anyway.

@BinaryBottleBake bro, have you solved this yet? I had the same problem and fixed it by going into ComfyUI settings, searching for 'loop', and turning off 'Check for loops before submitting'.

@BinaryBottleBake I tried everything — updated ComfyUI, switched PyTorch, updated all nodes, searched everywhere — and it ended up being some dumb setting in ComfyUI. Can't believe it

Every setting that is used for generations will change the outcome, only post processing like interpolation will not change the outcome. It is all adding to the noise. Only if all influencing settings stay the same you will get the same.

@darksidewalker Thanks for the reply. Than it was just a coincidence that sometimes the output stays almost the same I guess.

The math is right is you set it like this: 16 fps x 81 frames = 5s; 5s x 10 sampler = 50s.
It all depends on your settings.

The error is when you try to combine videos without audio, you would have to skip the audio combine if there there is no audio in.

@darksidewalker then how to add audio to the 5 sec videos you create with the first FastFidelity C-AiO workflow

@Yinrex 1st do audio with foley / mmaudio, than combine. Or disable the "with audio combine" switch to combine without audio.

use AiO version.i2v version is older and aio has everything u need.otherwise svi workflow for longer videos(3.3sec per section).

I did change that, because it was hard to determine what was happening in each segment when I just want to change one segment without touching the others.
That way it is far more easy to change just one segment, as they act as preview to that segment, to a desired outcome without touching the others.
If you want it the other way with successive complete rendered videos, you will have to change the wires of every preview video to the "extended images".

@darksidewalker thanks

Can you also tell me which one of the 7 workflows you are referring to?

@darksidewalker Oh my bad, the SVI 2.7 one

i too am just getting 3seconds per sampler. SVI 2.7

do we need to set frames to 120

@ferelik440 I tested both 16 fps and 24 fps and got correct times. I really don't know what the issue could be.

@Kingp0dd529t absolutely not. SVI wants 81 frames.

@darksidewalker i tried 16fps, i got 5 seconds. i put it back to 24fps, i got 3 secs

@Kingp0dd529t then yours is working, what's wrong with mine🤔? maybe I will try it on a clean comfyui instance.

@darksidewalker so does that mean that the 3second output per sampler is as expected?

@Kingp0dd529t 3,3seconds yes. 81fps divided by 24fps.if u try 16 fps u will have many issues.thats why 81fps is fixed.only in the normal Aio workflow u can change fps and frames but its will start looping with 16fps after 6- or 7 seconds.wan 2.5(no local version out yet).wan2.2 was made for 5 seconds low resolution and wan 2.5 for 10 seconds higher resolution.

Yes @Kingp0dd529t it is like @hartweizen said.SVI wants 24 fps and 81 frames. 

thank you to both of you! love the work!

i have a question, in the official workflow of svi 2pro, they use 16fps https://github.com/vita-epfl/Stable-Video-Infinity/blob/svi_wan22/comfyui_workflow/SVI-Wan22-1210-4-Clips.json

@Kingp0dd529t you can do 16, but following the discussions on their GitHub they recommended 24

only 1-2 seconds (rtx5090). exept u change something in sampler 3 as example but u already generated sampler 6,then u start again at 3 a new processing starts for 4,5,6. only work with 1 workflow not 2 workflow at the same time.never change antyhing in the samplers before.

@hartweizen that's weird, i already generated sampler 1 and 2. now i am changing the prompt for sampler 3, i don't change anything in sampler 1 & 2 (or any other nodes) aside from prompt text for sampler 3. but it takes minutes before it reaches the sampler 3 boxes. for reference, sampler 1 only takes 70 seconds, but when i edit sampler 3, it takes 300 seconds.

@Kingp0dd529t if im at sampler 8 cancle the job, because i didnt like what i saw at 7.i change sampler 7 and press run.he then checks all sampler video outputs before then starts rendering.i takes 1-2 sec per sampler. im using nvenc-av1(nvidia gpu) thats a lot faster then webm or whatever the standard setting is. video output always takes seconds. i dont know if there is a video output thats way faster for amd.but if there is one u have to change video output on all samplers(under each sampler) and both video output near image loader.

@hartweizen thanks for the tip, i will try nvenc-av1 format

You may be right. I'll look into it :)

C-AiO v5.0 revised that logic. Thx for the hint!

This is not normal, but I need more information please:
1# comfyui version, pytorch version, python version?
2# costom_nodes updated?
3# Workflow version

@darksidewalker pytorch version: 2.9.1+cu126
Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
ComfyUI version: 0.11.1
ComfyUI frontend version: 1.37.11

yes updated

C-SVI v2.8

@RavirKun As I can see you running an older version of comfyui.
Please update it and try t also update to pytorch 2.10+cu130

Yeah, there is a math error, I'll fix this asap.

Got it, thanks!

I noticed there's a 'Preview images' node sitting at the bottom of the Markdown Note. Did you forget to remove it?

@o622145789634805 I can not spot that node. 🤷

@darksidewalker I opened DaSiWa WAN 2.2 i2v FastFidelity C-SVI-28.json with Notepad++ and found a node with id=374 named previewimage, but it is not being used.

@o622145789634805 Thx, found it and removed it.

It mentions in the workflow notes that FPS is automatically calculated based on length. But it also mentions that MMaudio requires 24FPS and 5 second length. Default FPS is 16 though. Found the FPS node in the backend but I'm unwilling to change anything back there for fear of buggering something else.

Was a bug, comfy deleted the value, just load the recent one, should be right back in that gui

@darksidewalker awesome; great work as always

manually download rife49 pth then put it on comfyui custom node > frame interpolation > ckpt. paste it here and refrersh browser
work for me

I know. Will be back when I'm finished with it :)

I think this was a math error, tried to fix it in the recent workflow, maybe just redownload it 👍😄