They're completely different, if you can't tell the difference at a glance then you sir are either blind or downloaded 1.2.2 twice! lol.

The changes are listed in the version details. Basically everything was overhauled.

@The_Last_Goblin_King i guess you didn't bother to read the 'lol' which was meant to inflect a teasing tone. but fine, take it that way.

@The_Last_Goblin_King wow dude you are the most sensitive goblin king i've ever met. a little chill would do you some good.

1.5 MP is just pretty big for a 16 GB card, that's all. Maybe not if you are genning 5 seconds, but my guess is you're going for 10 or more. It's overwhelming your system. The v1.5 version has more VRAM options enabled and configured by default. Make sure to update the Latent Upscaler node too, as it also has a VRAM assisting option now as well. You could also consider the startup flags --disable-pinned-memory (i think it's that) and --reserve-vram 1

For now, I removed upscaling Video Continuations because simply put, there was a major issue with it; it produced flickers due to the latent being filled with two guides covering the same starting frames but at different sizes. You may have not noticed it if you do 2d/cgi stuff, but with live action it was pretty noticeable.

And to be clear, if you were able to gen at 1.5 on final upscale sampler, you absolutely can on 1st pass sampler as well. They are literally both doing the same exact thing... sampling.

@teiji25762 I knew what you meant. And I'm saying something that's purely logical, so let me break it down for you one more time:

When you successfully render X steps on the final pass of my workflow at 1.5 MP
It is EXACTLY the same as rendering X steps on the first pass of my workflow at 1.5 MP

The two processes are identical.

Adding more steps doesn't increase VRAM usage.

THEREFORE, you can gen the entire 1st pass at 1.5 MP. Do you understand now?

@freewellgear The main use-case of the workflow is the seed hunting (3 previews) + pick one for upscale finalization. Video Continuation is just an option I added in, and yes, it is limited to the normal straightforward workflow style of 1 gen = 1 output. You can watch each step through the previewer, assuming you have --preview-method latent2rgb as a start up flag, or use the Model Previewer option in the workflow. That's how you can see if your gen is proceeding correctly or not with that Single Pass option.

@ThanksHERPECOL Specifically the transition from overlay to the new gen. Otherwise I would just remove the overlay frames and replace with the OG video's frames.

the prompt IDE really helps with structuring and ease of use. but yeah, no more 1st pass degradation.

@sergiogmx minimax is quite smart so I just use a compact 3-shot: full body frontal, full body profile, close up 3/4ths. If your char has specific patterns on their back you can add a behind shot, but a lot of things can be described just as easily in their subject_definitions section.

Ha, yes, the workflow looks like a ref2va workflow but indeed, it being pre-programmed for the fl2va model is intentional, because as you discovered they function identically for the most part--with fl2va looking, in my opinion, better for reference work by a considerable margin.

With bf vs fp, you know I really hadn't considered that here. When I first installed the Latent Upscaler it didn't have that fp16 option toggle. But you're right, I should try it. Precision in the upscaling process could be a uniquely important factor.

I will try the sgm uniform scheduler, thanks for the recommend! As a content creator I have to be careful what I stick people with, an er_sde/beta is a community fav. I also really like Seeds 2, definitely have enjoyed it.

Thank you sir! I've noticed that Eros in particular has awful sound. For me it was simply unusable so I really had to just stick with base model.

Upscale wasn't seamless for vid continuation after upscale, simply put. It flickered due the 2nd pinning of guidance frames in the upscale pass. This version is truly seamless.

@bwcproductions643 You can use the ref2va sections, but it's trickier. The main tips:
1. No Turbo LoRAs. All of them seem to promote hard cuts.
2. "a single continuous shot from beginning to end." (as you know)
3. describe the overlap frames. 22 or 39 is 0.95-1.7~ seconds, and that time unless accounted for won't lead to a continuous shot. So if your preview shows in that 1.7~ seconds Hermione Granger drinking a potion from the side in her dorm, you might add after the continuous shot line, "a medium close up profile view shot of the girl with brown hair drinking from a glass potion vial in her bedroom. afterwards, she ..."
Think of the description line as an anchor that the model can read and go "Oh, okay, that's what's happening." Because from the model's perspective, the overlap frames are forced upon it; it doesn't want to be showing those frames, it wants to do its own thing.

Yeah, I held off on implementing it because too many people were OOM'ing with it and it was clearly just confusing people. I talked to the developer and he said he was gonna be streamlining it a bit. He's probably done, and there's probably a good implementation now, but I just haven't gotten to it quite yet.

@silesia335 Mm, yeah it sort of depends on what kind of scene you're trying to gen too. Artifacts appear at 0.2-0.4 MP frequently, I agree. If you're doing action I actually recommend higher megapixels for the 1st pass, like 0.5 MP @ 10 steps. That way the details are much more locked in by time it gets to final upscale. The two stages each have their own push and pull that I've sort of gotten used to.

Zoomed in, huh? Are you selecting the right aspect ratio on the green Video Settings node area? The key to continuing gens is to make sure your video aspect ratio is the same as your input, and to always be describing the continuation in your prompt. i.e.: what is happening in the overlap frames of your shot. The phrase `a single continuous shot from start to finish' comes up a lot in my prompts. Reminder, video continuation is ONLY for continuous shots. If you're just trying to link scenes together, a video editor is the right path.

@djm2000 Crop is the correct selection. It won't do anything if your aspect ratio remains the same. So if you're genning Widescreen 16:9, it doesn't matter what res you gen at as long as your Continue OG video is 16:9 as well. I wouldn't ever change it from crop. The reason you're getting a zoom is because of the model's interpretation of your prompt. Your prompt is allowing the scene to change, and usually that is combatted with the aforementioned a continuous single shot from beginning to end and then a description of what's happening during your overlap frames so that the model knows what it's continuing. For instance, if you're continuing a video of a girl playing a guitar on stage, it might look like: a continuous single shot from beginning to end. a close up of a girl with blonde hair is strumming her guitar on stage, she has a confident smile. which would describe the frames in the overlap. then you can prompt what actually happens. from the model's perspective, the overlap frames are forced on it, but it still has that 1-1.5 seconds of unaccounted-for time if you're not prompting carefully to explain what it's continuing from. I've seen it change camera angles, zoom out, or zoom in unless you're careful.

@SinglePage One thing I didn't seem to mention here is that turbo loras tend to create camera changes more than anything. I often have to disable them when using continuation 'cause otherwise I get a zoom in or angle change.

It is indeed in the manager. Not sure why you're unable to see it yourself there. Here's the link:
https://github.com/GACLove/ComfyUI-VFI

If Single Pass mode is on, Video Continuation mode is on, and you've clicked the "Use Preview" #2 and #3 toggles OFF (making sure #1 is on), the workflow will run just fine.

LOL if you're talking about Effy, the silverhaired girl, that's all post editing. I wash her out and put on glow to simulate the cellphone camera. Thanks for calling it not cinematic; that's the point in my case!

The reality is, you and any other user can switch to the ref2va model at any time in my workflow. I prefer the visual/audio clarity of fl2va, but I know some prefer ref2va. They both function identically as far as being able to use vid/audio references, with ref2va fumbling the ball more often for me.

Prompt for more to happen. Slow-mo only happens when you're not asking the characters to do enough with the length of the video. Asking a character to walk into the frame and stand, giving it 10 seconds, you'll get a slow-mo of them doing it instead of the expected "they walk in normal speed and then just stand there for 8 seconds". You can also use timestamps like:
At 00:00.000: <Subject 1> walks gracefully through the doorway into the room.
At 00:03.000: <Subject 1> stops walking and is standing in front of the camera. She then begins to make silly poses for the camera.

Basically, you can't have slow-mo if you constrain the time between actions. It will try to accomplish what you need in the time you give it, meaning if you give her 1 second, it'll appear too fast. Hope that helps.

I have no idea, I've never used RunPod. I hope someone else can chime in.

If you look to the left of the Seamless Video Continuation group there's a big white box with step by step instructions.

It's up to you. I've run 1080p movies through it, I've run 360p youtube vids from 20 years ago through it. It's single pass only, so whatever you set the aspect ratio / megapixels to, that's what you're gonna get. I do have some recommendations though. Try to shoot for no turbo LoRA if you want it to be a truly seamless transition. Assuming it's a single shot, you'd start your prompt with a single continuous shot from beginning to end and describe the scene that is happening, because from the model's perspective, the overlapping frames are part of the gen, and the prompt needs to guide the gen from that to where you want the video to go. I recommend just 10 steps, no lora, to match most 'medium' quality videos. Super high quality videos you'll want to shoot for 12-14. But very rarely do you actually need the full amount of steps to resolve a continuation.

@djm2000 if you use a turbo lora you'll get zoom ins. if you don't say 'a single continuous shot from beginning to end' at the start of your prompt, you can get angle changes and zoom ins. also just to be clear, the overlap frames follow the formula 17n+5, so 5, 22, 39, 56. You listed 5, 25, 32 which would cause big problems. I'm only mentioning it in case you don't know and are actually using those values.

Hmm, I will update comfy/nodes as see if something changed. I'll report back.

Hi there, I updated comfy + all the nodes, and the workflow is still running fine for me. The AV Separate node is from the LTX node pack (the actual node itself is called LTXVSeparateAVLatent). Are you running an up to date version of it? And you're also not using any weird versions of the model right? Like some fasth3 + custom text encoder combo?

@equinoxe_c64 You're alright, it's a complicated workflow. Don't worry about it.

Test it out first, run 0.5 MP 1st pass, upscale to something like 1.2-1.5 MP, and see what times you're getting for like 5 seconds, 8 seconds, 12 seconds. If all good, you can do 15 seconds at that size. It's a pretty flexible workflow so mess with things and see what works best for you. Some people like the 3 previews to be really fast, so they do 0.3 MP there @ 8 steps, and upscale to 1.0 MP with like 10 steps.

Can you tell me what you're doing? Regular gen, Vid continuation? And what part in the process does this error pop up? Before sampling starts? As you try to upscale? Need more info to diagnose. But I can say it's very likely just a misconfiguration. NoneType can come from not having a node pack installed, it can also come from empty data being passed to a node. Usually it's just an improperly clicked toggle or a node pack needing to be installed.

@tibbydapug232 Interesting, the error is propagating from kjnodes, which with the error implies that one of the image locations isn't supplying data for stitching. I'm assuming you followed all directions and have turned off previews #2 & #3, and turned on Single Pass in the options. Given that assumption, is it possible that your OG video's time-span selection (the blue bar) is smaller than your set number of Overlap frames?

@lajidef734122 Likely you have the startup flag "--no-cache" enabled, or simply DID change something that would affect stage 1 without realizing it. Comfy doesn't regenerate unless there's a reason to, it's just not how it's designed.

Hi @KindyT -- this is not a workflow for an 8gb vram card. I have not optimized it for that low VRAM. I do have some VRAM saving options but 8gb is probably outside of the ballpark for them to really help.

As I said to @lajidef734122 typically this comes down to user error or settings. Make sure you don't use the start-up flag "--no-cache", as that instructs ComfyUI to never save latents to memory. It's a very common startup flag for low vram users, so I wouldn't be surprised if that's what's affecting you. But personally, I'd leave that on and use a workflow designed for low vram.

@KindyT Well if it isn't --no-cache and you're sure it's not user error, you're one of the very rare individuals where I'm not sure what actually is happening. You run comfy desktop, or portable? Comfy's supposed to save the latent in memory, which is what allows it to carry over from 1st to 2nd stage between the two runs. I assume your "Manual Random Seed" nodes are still both set to 'Fixed' right? I've seen people accidentally/automatically switch them to "Randomize" which causes of course regens every time. I suppose one optimization if you're stuck in regen-world, would be to click disable and turn off the two previews you didn't want to use, so that way you only ever have to regen the 1 latent you chose, and then it goes straight to upscale without regenning the 2 rejected previews.

@KindyT It sounds like you're likely encountering some user error in that 3rd segment. The workflow has no knowledge of previous segments. It just sticks 22-or-whatever overlap frames onto the beginning of your gen and tells the model "continue this". Verify everything's correctly set up, especially with the blue timeline selector in the vid loader. To be honest though, it sounds like you could do yourself a favor and get a solid video editor like Davinci Resolve (free). If you're doing a bunch of shot continuations in a row it's just easier to not have comfyUI stitch them together over and over and just toss them together in the video editor. Video editors are how you'd make real high quality production output anyway.

Either way, glad you're up and running!

I haven't had the chance to investigate it much, besides seeing others discussing it in the main h3 banodoco discord.

If you're running a 5090 (even on a laptop) something else is wrong. Should not be taking an hour. And you should always be using Comfy Kitchen, it's basically free speed. I will be updating the workflow tomorrow to switch to the new built in core node Block Attention, which will also help you speed up. It's Kijai's Sol Attention without requiring triton, it's really good. But seriously, make sure your comfy is all up to date, your nvidia drivers too, and make sure you're not running any programs that use vram. also consider starting up with --reserve-vram 1, which will help you not spike out into massive slowdowns due to memory mismanagement.

That node definitely does look clean! However I actually use the features in the Image Loader (such as being able to crop with click->drag rectangle, setting individual image max MP) and the audio/video loaders (being able to use the timeline selector to select beginning/end times). So while my setup looks a little clunkier and messier, it's got functionality that I use daily.

@kaileykk If it does have those features then I'll check it out. Everything else about it looked pretty solid. Thanks.

the only explanation I can think of is if your audio vae isn't properly selected. this is not an issue that can just 'happen' to the audio but not the video without some logical explanation. also I guess make sure you don't have 'force custom audio' option turned on near the Audio 1 ref section.

Is the sampler built into the MMH3 Ultimate Upscale? If so I have no interest in it. The actual latent upscale process isn't what causes people issues with my workflow; that part only takes 5-10 seconds on average. It's the sampling at the new latent size that causes OOM or slowdowns due to VRAM consumption. The SamplerCustomerAdvanced simply is having to reckon with a huge latent, often times bigger than the model was intended to run at.

I'm assuming your comfy is all up to date? You might wanna start using the startup flag --reserve-vram 1, which will prevent comfyUI from spiking vram into OOM when upscaling. Though this model is a beast, so 12GB VRAM and 32 system RAM means the 40+ GB of text encoder + model don't have a ton of room to wiggle around when it comes to upscaling in general.

Read the version info. It's part of the new comfyUI update.

@lukaskolembus245 Are you using the Desktop version of comfyUI? It might be that the node hasn't been pushed there yet. It was released yesterday for the Portable version of comfyUI that most people use, I do know that much.

@lukaskolembus245 I guess that explains it. Desktop users get updates a few days later typically. You'll just have to disable the option until comfy pushes the update to you. The WF is still perfectly usable, it'll just be a little slower for upscale phase.

@Sassyphrassy hi, you would use <Audio 1> to load whatever reference from the video your continuing that you need. It's not like just because you're doing a video continuation you can't use all the reference slots that minimax offers, 9+ images, 3+ audio are still good to go. You just need to make sure to prompt correctly so that the model knows how it's utilizing them. Also, you can set your overlap value. 22 is just the default since it's reasonable for most use cases. The formula is 17n+5. so 5 is the shortest, next is 22, 39, then 56. The more overlap frames the longer you gen will take though since they all have to be considered by the model.

Right, your AI is thinking that the guide's audio input from your OG video was overwriting <Audio 1> from the ref control area (I don't believe that's true) but I'm glad it's working for you. A way to accomplish the same thing would've been to just select "OG Video has NO Audio" in the Vid Continuation group. Then it wouldn't send the audio to the Guider nodes, which is the same solution your AI came up with.

Are you just trying to do a normal gen? Without context / more info I have no idea how to help you. If final audio isn't making it to upscale usually that would imply there's some option on that shouldn't be.

@Witschequast Oh awesome, glad you like it! Episode 2 is out as well, on my youtube page (@foxfuressence). And yea, Nodes 2.0 is a misfire.. 99% of comfy users have to turn it off to make stuff work.

Use a free video editor like Davinci Resolve to bring the two clips together and export them as one in a single video file. This is a filmmaker's workflow first and foremost, so I do sort of frequently advocate that people become familiar with video editors. The only reason my short films look any good is because of them.

@Dannyone Indeed. I do almost all of my scene composition in Davinci. As long as your clips are the same aspect ratio, they will be sized to whatever output you're making for your video in the video editor.

Hi, first off thanks for taking the time to explain your struggle rather than a 1 sentence comment that I'd struggle to help with.

>But what do I do after that if I want to continue the exact same scene?

You'd turn on Video Continuation, Single Pass, set your MP/length/step count accordingly. Load your video, select your overlap frames (5, 22, 39, 56), choose whether or not the vid has audio (if you say yes the gen will have access to it and learn sounds/character voices, etc. if no, you can use audio reference 1 for whatever). As far as prompting for continuity, I always start prompts with a single continuous shot from beginning to end then describe what the first overlap frames look like, what's happening, etc. If you use a turbo LoRA or don't do that, the model sometimes will cut to a new angle. So I always recommend not using Turbo. If you want to keep your graphical fidelity up, use a Last Frame in <Picture 2> so the model has a goal of where to go to return the vid to high quality. Otherwise with each continuation you're feeding AI into AI and the vid quality will degrade.

There is no upscaling for Seamless Video Continuation. What you set for MP in the 1st pass is what you get. It's perfectly fine to run 1-1.5 MP gens through single pass and achieve the same quality as upscales, it just takes longer. Video Continuations aren't meant to be chained a bunch. Very few cinematic filmmaking shots require a continuous shot more than 15 seconds long.

Generate → choose the best of the 3 → Upscale → Seamless Continuation
^ this is the only valid path.

Hope that helps.

@Dannyone You're fine, and to be clear I wasn't making a reference to you or anything. I get looots of 1-2 sentence "why is this not working?" questions that I can't answer without more information.

Really appreciate hearing that! A lot of the time I only get comments from people who are struggling and have questions. I know in the back of my mind there a lot of people are out there silently enjoying the workflow with no problems whatsoever lol

Apparently comfyUI didn't push the node to Desktop version users. Just Portable. Thankfully most people use the Portable version, but that doesn't help you guys. Thankfully the workflow will function with the node simply turned off.

@addison8406 hi, the default settings are supposed to be fine. try changing the drop down choice for both of the nodes in the wf to 'sol-attn' (both in the single pass and in the bottom right '[Attention] Block' groups. I really shouldn't have updated this workflow to the node yet-- I assumed because it was built into comfy now that it would work right out of the gate, which is historically a very bad assumption. They're just changing it under the hood and it's creating errors for us users.

@addison8406 what helped for me (i'm experiencing these errors too after all) is right click the nodes -> recreate node (reset widget values).

damn comfy, pushing this glitchy crap.

also, i updated the workflow to replace the values that caused the errors so this won't happen for future users. I just wish comfy wouldn't push broken nodes into their core like this. In the future I won't be so quick to add them just because they're supposedly 'built in' now.

Hi, to answer your questions, the default for 1st pass being set to 8 is exactly for the reason you suspect; because when you find a preview you like, you finalize it in upscale with another 8-12 steps.

And yes, the workflow has an option toggle: "Single Pass", where it skips the upscale pass and only uses the 1st preview sampler, sending whatever finishes there to any finalization options like RIFE and then saving.

As for quality, single pass will always be better. But it's also 2x longer if you do 20x steps at 1.0 mp rather than 10 steps at 0.5 MP and 10 steps upscale to 1.0 MP. It's a trade-off decision you have to make for yourself.

The only difference made between the versions is Block Attention, which uses the same process as the previous node if you use the drop down and switch to SLA (which is good for if you're using turbo loras). So your perception of the last version being faster is really either just that, or anecdotal.

You can enable the 'Finalization Only LoRA' option to use a Turbo LoRA only for the final upscale steps. It's good to not use turbo on 1st pass because the base model's motion and clarity tends to be better. Using a Turbo for upscale pass only speeds up how quickly you can upscale, without really affecting the motion of your original latent too much. It's a good trade-off for quality and speed for some people.

The errors are because the node is brand new in comfy and there's problems with it. I simply got excited thinking 'yay, people will be so happy to have the speedup built into comfy instead of being a separate node', without thinking that comfy could fuck up their own implementation. It was my bad.

Very likely so. I did make one for v1.0 but the workflow has changed a lot since then. I'd make one today but I'm knee deep in finishing a different video...

Very few workflows are like mine. Your AI simply has no frame of reference to judge it. What other workflows do you know that you run twice to get a result? The first run produces 3 sample previews, and the 2nd run takes your chosen sample preview to upscale finalization. On top of that it has several complicated features like seamless video extension with frame blending. So yeah, to a GPT I'm sure it looks crazy under the hood, but it's organized and functional and suits every need I have as an AI filmmaker.

@Sassyphrassy it's fine, at least he didn't make a 2 minute hit-piece video like the last guy did lol

afaik no model fully beats color shift -- it's a problem for sure. Typically I just adjust in my video editing software. Most of the time it involves desaturating slightly, and changing the shadow/highlights level (contrast, really).

It was upscaled yes. I think 16 GB can handle 1.5 MP, but it'll be really slow because even on 24GB 1.5 MP is pretty slow at 10s

usually it's just a misconfiguration. make sure your model loaders are loading the right files, that you've followed vid continuation instructions (i.e. single pass is on, upscale is toggled off, you have one option selected for if the video has audio or not, etc.)

Not really sure then, unless I saw your actual settings in the workflow. I've never gotten this error. Does a different video give the same error? And you only have 1 audio option selected correct (either has audio, or doesn't). I had a situation earlier today where a video's specific audio type messed up my gens. You could play/hear the video in the previewer, but if you listened to it with audio preview it was all slo-mo and muffled, and caused errors. So sometimes it's just the video format that comfy can't handle. Just trying to think of everything here since I can't sit down at your computer and troubleshoot myself.

@tibbydapug232 Right click the Block Attention nodes (there are two in the WF, one in single pass group and the other under the final pass upscale group), click -> Recreate Node -> Reset Widget Values.

ComfyUI screwed up their release of the node and changes settings randomly.

@snake2025 it's needed sometimes... everybody needs a refresh occasionally. been a year for me, somehow with everything i've done i still haven't managed to break my install. rolling that dice every day though. good you have everything working now.

@MoonDroch just to be clear that node does not exist in my workflow... are you just saying having it installed caused that error for you, even if not present in mine?

If options that shouldn't be enabled like Vid Continuation, Single Pass, or Force Custom Audio are on (without meaning to be and loading an audio file from Audio 1), those can affect final output audio on normal gens.

Because comfy-kitchen attention is the same speed up & cost basis, but doesn't require any python installs of sage attention / triton. And since they're mutually exclusive, I removed the more complicated of the two options. AFAIK there would be no valid reason to use sage over comfy-kitchen attention.

You have to right click -> recreate node -> reset widget values. ComfyUI screwed up with their update.

Yeah, I'm aware. Guy seems to have his own whole head-canon about events that I don't recall. I just think when you get big enough or make something popular enough it starts to attract flies, really when it comes down to it.

Right, as @zachet39 said, if you're missing MinimaxH3 nodes then you're definitely running an outdated comfyUI.

@johhnnymann1 There are little notes above each option at the top of the workflow. You start on stage 1, gen up to 3 previews, after it finished comfy will go to "Idle", and without changing anything about your prompt, select the toggle of the preview you like, enable Upscale Pass, and then click the Red 'manual random seed' button to the right of that over the Upscale Pass section. The green "manual random seed" button controls 1st stage, red controls Upscale stage.

@tibbydapug232 @equinoxe_c64 right click the two Block Attention nodes in the wf (there's one in Single Pass group and one on the bottom right near the Upscale phase), and click -> Recreate Node -> Reset Widget values. That's what fixed it for me.

USUALLY this is user error-- simply not realizing you changed something, whether it be removing a letter in the prompt, changing the crop on an image, ANYTHING that would tell comfyUI that something has intrinsically changed about the 1st pass. The other common reason is the startup flag "--no-cache", which tells comfy to never store data between runs. I recommend you test with someone really really simple, just a 2-3 sentence T2V prompt with simple settings; gen the 1st 3 previews, don't touch anything, just click enable on upscale pass and then hit run again. You'll see each of the 3 preview gens run a quick decode pass on the VHS Video preview, but it only takes a few seconds before it should pass to upscale phase.

I think in most cases you know what kind of aspect ratio you're trying to produce videos in, right? Like if I'm making a short film it's usually going to be in 9:16 or 16:9. I also just dunno of a node that does this, and I do have to be very careful about what custom nodes I add for this workflow because people start to complain if there's too many random ones required just to start using the workflow.

Mm, expressive voice rather than AI voice is a pretty big topic, but loosely, using voice references is the key I've found so far. 45 seconds of a good natural voice actor that isn't speaking right up against the mic, assign it as a vocal-timbre reference in retention_analsysis to your subject, and then in detailed_description you'd have them speak like: "<Subject 1> says using <Audio 1> as her vocal-timbre reference: <d>[English] I can't believe you let the muffins get cold.</d> her tone is playful but grumpy, her expression pouty and her eyes narrow." something like that.

Lack of sound probably has nothing to do with your prompt, you just make to make sure only 1 option in the Video Continuation section for audio is selected: -EITHER- "OG video has audio" or "OG video has NO audio".

Make sure Single Pass is on, and the other two previews are turned off (#2 and #3 each have toggles). Lastly, make sure Final Upscale Pass toggle is off. The brown wavy noise is what happens usually when a latent that doesn't exist is being sampled. So it's definitely just a configuration error.

I noticed too that on occasion, a video that's not encoded with a common codec or is of a specific file type can sometimes not load. I was trying to extend a .webm the other day that had corrupted audio and I thought my workflow was broken until I realized.

Anyway, best way to check would be to re-export your video in a video editor or ffmpeg or whatever you use. make it a generic .mp4 with the typical h264 encoder.

also, yes, this wf was not designed for anything less than 16 gb vram.. i'd wager your upscale issue has to do with that more than anything. even my 24 GB VRAM can't upscale more than 10 seconds of 1080p.

You can definitely use this wf with 12 GB VRAM. Just don't expect to be able to do 1080p, really.

Err, you're upscaling TO 0.4 mp? You mean that your 1st pass is 0.4 MP and that you're upscaling to something else, right? The workflow's default is to upscale to 1.0 MP.

This sounds like more of a node issue than an actual memory issue. You verified that your Upscale custom node is up to date? This one: https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

Also, since the error mentions 'attention' too, I will also note that if you're not on the newest comfy portable update, the Block Attention node probably won't work. They JUST added it to comfy and it's been a major problem for comfy-desktop users so far since they straight up don't have the update yet for it.

You can still use the workflow by disabling the Block Attention option, it's just kind of slow without it.

I assume you mean using audio vocal references for video continuations right? Yes--If your video has a voice in it, it will by default copy that voice for the same character assuming your overlap has their voice signal within its range. But if you want to supply a reference, do it in <Audio 2> because I feel like <Audio 1> gets overwritten by the OG video's audio index. Make sure you describe the character as speaking with <Audio 2> as her vocal-timbre reference. In summary also you'd want to have [reference generation + audio reference], all basic stuff from the prompting guide.

It's slow because Minimax has to look at all the reference frames of the video you're continuing. If you set overlap to 5 frames, it'll be faster but have less context. The default of 22 is a good medium group for most things. But sometimes I do 39 or even 56 depending on use-case.

Well I don't know what your mem usage is in between swapping, usually huge balloon-ups in s/it come from lack of availability of resources.

Sol attn doesn't apply to the first 1-2 steps also, so if you canceled after 1 step seeing 400s/it then that didn't factor the speed up in. But still, obv 400s is wrong. My suspicion is if you restarted comfy in between your tests and made sure no other system app was pulling resources, your speeds would be in line. The only change between 1.5 and 1.6 is the Attention node switching from Zironic's SLA to the comfy UI built-in Sol/SLA version.

The main one is turbo LoRAs. They change angles more than anything. So:
1. No Turbo
2. Start prompts/first thing in summary: "A continuous shot from beginning to end."
3. Describe what is happening during the overlap frames. If you're extending a video of Hermione drinking a potion in her dorm room, you describe the angle and shot composition, the brown wavy haired girl drinking a potion. Sometimes the overlap is 1~ second long. If the model doesn't account for that second in the prompt, it can get confused and just try to leap to something else.

I have considered it. There are a couple of things I'm weighing right now for the 2.0 release. De-roping option, a new (normal) ref handler.

Hi, I have seen others get this error but I have never been able to recreate it myself. In a couple of cases it came down to that users were loading different versions of the model (finetunes), or didn't have the correct audio VAE selected (like they had the video VAE loaded in the audio VAE slot or something like that mixed up). Since you seemed to have genned the first part okay, I'm not really sure why you'd get a mismatch error for vid continuation assuming you have followed instructions (single pass ON, chose 'OG video has audio' and disabled the other in vid continuation section, disabled Previews #2, and #3). As long as your comfy is up to date, your nodes up to date... should be running with no problems. I wish I could recreate this error somehow, because I sure would like to solve it for others who experience it.

@arcusprints Really appreciate that! If you find anything you know how to reach me.

Hi, thank you for the kind words -- and yes, I am very much seriously considering adding RefMod functionality to the next release. I am in the middle of projects/video production right now and have simply just been USING the workflow rather than developing it for the past few days. It'll probably be another few days before I can release a new version.

Hi, in researching VDN, it is not the direction I want to take my workflow at the moment. However, with the custom comfyUI nodes for it and the flexibility of toggles in my workflow (being able to turn off all the accelerators and such), I do not suspect it would be difficult for you to add. Most of the changes would be just at the model loader level.

I will check it out. Thank you for the suggestion!

Hi there.
1. Yes, the whole point of video continuation is to continue a video. You can always use this workflow to gen normally. Video continuation is just an extra option.

2. Overlap frames is how many frames you give to the model to study for continuation. The model only accepts multiples of 17n+5, so 5, 22, 39, 56. Meanwhile, the blue handle bars on the selector control the final resulting stitched video. The right handlebar determines the first frame the new gen will continue from, and the left handlebar controls when the video starts when all is said and done and the WF stitches the two videos together.

Hi, since others are asking about this too it is something I plan on seriously considering for the next release.

In classic form, ComfyUI is changing things internally that make things very difficult for users. Last week it was the compiler, and maybe it still is this week, but yeah... just gotta hang tight until the storm passes. I'm being careful not to release any new versions until they get their shit together with this. It's not like many of us have RAM/VRAM to spare with a model like h3 lol

Best way is to set the "<Picture 2> is LAST FRAME" option. It pins the high quality image of your last frame into the latent and will force the model to repair or maintain quality, since the model's primary goal is to get to the final frame. If that final frame is a high fidelity shot of your character for instance, it will be high fidelity by time the gen ends, allowing you to extend infinitely as long as you can keep producing high quality end frames. It's a pain but it works.

LOL how this never occurred to me is crazy.

Hi, yes, n-sfw works (you'll see tons of people have posted N.SFW examples in the gallery at the bottom of the page)

And, I see now that the maker of the latent upscale node has changed a bunch of things as of 8 hours ago... so, you will have to bear with me to make sure everything's still working. but for now, yes, you should be able to use https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/blob/main/minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors

@maaade that's sort of a vague question and i still don't know how to answer it. best way to get clean, clear gens is to use single pass at 1.0+ MP, no block attention (only comfy-kitchen attention).

Hi, thank you. If you can make it customizable as to which group it triggers, that could be interesting programming logic for future versions. Custom nodes are a big ask for people, and this workflow already has a lot of them, so I can't just add in stuff without it being a really big deal. But if there are ways to handle resets throughout the entire workflow on start/end actions, that would be neat to explore.

I am sorry, I don't know much about the sub 12GB VRAM territory. Likely my workflow would not be a good fit as it doesn't specialize in low VRAM setups.