@grbear750611 Thank you so much.

While creating this version I dived a little bit deeper into the ltx-2.3 specifications and what I finally got was: use 16:9/9:16 aspect ratios only and images sizes divisible by 32. So one of the aims was to make sure every generation runs under these conditions only. And after my first tests I am pretty sure we will get better outputs now.

So, short answer for any other aspect ratios is: no, these limitations are by "design".

@Cosmicv As mentioned above, I strictly "designed" it to get rid of the previous issues with image quality. But according to my first tests it is no problem to use rectangel images/videos somewhere around 16:9 as long as the main part of the image is not too closeup.

@grbear750611 Yes, you can still use v2 if it worked well for you, cause there is no change in the generation process itself. Just keep in mind to use the latest upscale model.

I had not the time nor the resourcess to run serious side-by-side tests between workflow versions 2.0 and 3.0. So it is all quite subjectif. My personal "feeling" 🙄 after some more tests is, that v3.0 gives better video quality and better prompt following in general. But let`s wait and see, what experiences others will have.

@RishanMasters Look here.

@charlesdelavigere743 Thank you. Let me know your experiances after testing, cause my own test ressources are very limited. But according to my first quick-and-dirty tests it seems, even prompt following is better now.

Btw. the Ukulele video is really too cool 😂🤣Would no wonder, if it will become the summer hit of this year 🙂

@charlesdelavigere743 Ups, overlapping comments....

There should be no differences in generation time, except comparing different resolutions. Cause did not touched the generation process in general.

@dirtysem Hi -thank you. Let me know, if you find any bugs.

@dirtysem Yes I know, several users want other aspect ratios, but I exclude it by "design" 😉Please hve a look at the discussion here.

@wolfcat2 Did you choosed the right options only?? Please start simple with T2V first. If this works, go the next step. If you still get errors, please provide usefull informations: Have you followed my Troubleshooting guide, wich option in use, etc.

@wolfcat2 Ok. What`s about the slider nodes in your workflow? Can you see them? If not, look here.

@wolfcat2 Did you solved the node conflict by disabling mixlab nodes, like mentioned in my link???

@wolfcat2 Uhh - don`t play in the scripts. Simply disable mixlab via the manager. If you need it someone, you can simply enable it agein.

@wolfcat2 Sorry, I can`t help you with your system. Maybe the easiest way for you is a fresh Comfyui-Easy-Install installation, just for video generation (see my short guide at my Wan model for help). It will take you around 30 minutes and you are up and running.

@wolfcat2 As descriped in my short guide you can install it parallel. Believe me, it is the most lazy way of all 😉

@hid91herring074395 Seems you have node conflicts in our comfyui system. Anything updated according to my Troubleshooting Guide? No errors during updates??

Generally all nodes I use in my workflows are installable via the Manager. In the Manager search for: "Custom Switch" to get the Orchestrator nodes (or see github).

"LayerUtility: AnyRerouter" is strange, cause I did not use any LayerUtility (as far as I remember).

As I often mentioned here: take into account to install a seperate comfyui for video generation only with Comfyui-Easy-Install (see my Wan workflow for a short guide). It takes around 30 minutes and 2 mouse-clicks and you are up and running.

@hid91herring074395 Sorry, I don`t got you. What have you done to solve the issue??

@Gavr728 To use I2v/V2V is no problem as long as your input is rectangle and your subject/object is mainly centered. It just gets automatically croped. For the aspect ratio discussion have a look here.

@dirtysem But that's getting a bit pedantic 😂🙂 No, serious - I got you. I really did not tested yet, what the crop node will do with a square image 😉 Easiest way to be sure, would be to crop the image manually in advance - also for ractangle images, where the main part is not centered.

@Gavr728 You simply can edit the slider nodes (right click -> properties). Set Resize by longer edge = 2048 should work. It will generate 2048 x 1152 automatically. But you should manually edit the values for the LTXVPreprocess too - from 1536 to 2024 (open subgraph, see top centre). Let me know if it works.

@Gavr728 What gpu/vram/ram do you use? Do you still use the gguf models? I have no experiance with higher hardware. LTX-2.3 will work up to 4k. If 2k at least delivers better quality (and I`m pretty sure it should) I would implement the settings for the higher resolutions in the next update.

@Gavr728 Ah, thank you. RTX 3090 16gb vram is a common league, as I see here from the comments.

One more question: did you compared your resolution 2560x1408 with the standard 16:9 2048 x 1152 resolution? Cause there were a lot of image quality issues with my previous workflows, were I "allowed" any aspect ratios.

@Gavr728 Thank you. But I`m a little bit confused now, If I understand it right:

?? 1536 x 864 = 16:9 (highest resolution I set in the workflow for 12 gb vram),

Full HD = 1920 x 1080 = 16:9,

2k = 2048 x 1152 = 16:9

These resolutions should work without any issues, cause they are all 16:9 and divisible by 32.

But in your previous comment you wrote: "2560 x 1408 takes 10 min". My question was more about this deviating aspect ratio and if you saw any quality differences to the standard 16:9 resolutions.

@Gavr728 Uhh - slowly: Wich option did you used?

You edited the input slider to be able to set 2560 as longer edge - right? And you got out 2560 x 1408??? That`s strange, cause 2560 is just multiplied with 0.5625 (= 1440) as you can see in the subgraph.

So, please let me know the option you used and I will check if there is something else wrong in the workflow.

@Cybernix Hi, thank you for the hint. That`s interesting. Do you have practical experience with it? Is there a github page or anything else for more information? I just had a quick look, but it is hard to find anything related.

@Cybernix Thank you. Sorry, if my question was not really understandable.

I had allready seen the node and I got your settings and I allready had a look at the LTXVideo github page. Maybe I`m blind, but I found no serious information about the "LTXV Spatio Temporal Tiled VAE Decode" node at all. So my question is, from where do you know "there will be more stability"? Ore simply asked: did you run own tests? Ore did you found some information somewhere?

@Cybernix Thank you so much 👍 That`s a good point. I noticed RAM crashes at VAE decode in the past sometimes too. I am not really sure, but I would say this was all with older comfyui versions. During all my tests for workflow version v3.0 I got only OOM crasehes during the second ksampler pass (the last 3 steps) when I set highest resolution and too long clip length (>14/15 seconds).

Ok, another point coud be: my swap file size is set at 64 - 128 GB wich is disproportionately high. Maybe your suggested node + settings needs less RAM. This would be an advantage of course. I will try this. Did you eventually noticed different RAM/swap file usage?

@Cybernix Ah ok, a few Linux users reported simillar issues, as I remember right. Yes I see this only from my side of a Windows machine 🙄 Best way for a quick side-by-side test would be a simple workflow: Load Image -> VAE encode -> VAE decode (node 1 or node 2) -> Preview Image. When I find some time I will have a look at it.

In my workflows I allways use common nodes wich are installable via the Manager. To get it running, you have to solve the node conflicts in your comfyui system first - see my short Troubleshooting Guide for help.

There are lots of audio engines in comfyui and even cloud services. So of course, you can generate your desired audio first. Then you simply use my Audio input option and you might get whatever you want. But belive me or not - to create professional audio, you need knowlage, capable hardware and lot of time 😉

@paperpunch245 Hi - thank you so much for your feedback 🙂 About the aspet ratio limitiation I got severel questions of course. Please have a look at the discussion here.

@AIden_AIzawa Hi - thank you. I assume you mean FLF to Video? This should work out of the box. Ok - you have enabled "Image to Video" + "Last Frame input". Please set "Cut Off End Frames" = 0 for testing (and have a look at my Input Help note).

@AIden_AIzawa Mmh, that`s strange. I did not run much tests for myself with FLF and mostly used simillar images or the same image for first and last frame. With very different images I got some problems too (random jumping between firt/last frame etc.).

I just checked the template workflow and they use other LTXVAddGuide nodes (2 instead of one) and they use strength of 0.7. Also they use only a one pass image generation. I will see, if I can implement the template nodes - but this will take some time.

As a quick test, please just edit strength_1 and strength_2 in the LTXVAddGuide to 0.7 (open subgraph and go to the bottom left). Let me know your results.

@AIden_AIzawa I tested the template FLF2V workflow yesterday (just changed the models to the one we use here, so I can run it on my machine). Unfortunately I got simillar issues (abruptly jumping from one frame to the other, no lipsyncing, jumping forth and back, etc.. So I would say it is no node or workflow issue. Maybe the full models will work better with FLF, but I can`t test it.

@AIden_AIzawa I just read your first comment again. You wrote v2.0 worked perfectly for you. If I remember right, even with v2.0 I had simillar issues. Sometimes I believe it depends on a variety of conditions, like used images, prompt, aspect ratios, seed, etc....

@ArcleinSK Wow - thank you so much for your hint 👍 and I believe you are right 🙂 I just checked the template workflow again - as you said: first frame = frame_idx_1 = 0 and last frame = frame_idx_2 = -1. This makes sense. I did not worried about the values before 🙄

The only thing that surprises me is that I haven’t achieved satisfactory results with the template workflow either (see above).

I will test this in my workflow now. So let`see.

@ArcleinSK I did a first quick test and it seems to work now. Pictures are in the right order and I did not got any artefacts. Just a quick flip between first and last frame image - but I used slightly different images, like different clothes. Please let me know, if you get good results now or if you see any other issues.

@ArcleinSK Thank you so much for your feedback 😋

I ran my tests yesterday without cutting the last frames and as mentioned I got good "quality" till to the last frame. The strange thing is just the sudden flipp between the first and the last frame image somewhere in the final third of the video - instead of a smooth transition. I did not run much tests with FLF at all, but I would say I saw this behaviour in all workflow versions. One point coud be to play with strength_1/2 too (I only used 1.0 yet). I will give this a try, if I find some "free" computing time 😅

Workflow speed/stabillity: As you can see in the subgraph, I mostly used the template workflow and the models wich urabewe provided originally for 12 gb vram, so there is nothing special, except the more complex option switching and the calculations for width and hight. I just did a lot of test runs to keep the selectable resolutions in a range wich works well with 12 gb vram (mostly 😉).

@ArcleinSK I ran some quick tests with lower strength (down to 0.4) allready, but it did not change much and the issue with the sudden flipping still consists.

@ArcleinSK I just saw your linked video. Amazing! Apart from it keeps me laughing 😂 it is unbelievable that you got the whole scene in one run.

Oh, oh - this stuff is more complicated, then I suspected. To be honest, I just copied parts of the workflows, without understanding what specific nodes are really doing....

Your investigations are very interesting and gave me the motivation to try to dive in here a little bit deeper.

Frame index "-1": To understand it as a placeholder value makes sense (I was allways wondering about it before). But I did not really got your explanations how you think it works.

In my opinion (or more what I actually guess) and your idea of the placeholder I think it should work like this:

frame index = 0 inserts the given image to the latent image queue at position 0 (= first image) and

frame index = -1 inserts the given image at the last position (with your example of total frames = 121 -> it means position 120 = last frame). Everything should be fine - so far 🙂

I do not understand the video generation in detail, but simply spoken it should work this way: the generation process allways interpolates across several frames (in the future) to generate smoth movements following the given prompt. So it should start exactly with the first frame and should narrow to the last frame at the end - frame by frame. But maybe I am completely wrong here, cause it should not give a "hard" flip.

At first quick "researches" I found only some simple descriptions about the LTXVAddGuide (Multi) nodes here for example and something interesting too here wich gives a hint about the LTXVCropGuides nodes.

I did not know that it is even possible to insert videos instead of single images (described in the first link above). It seems the LTXVAddGuide (Multi) nodes have much more capabilities then simple FLF or multi frame generation.

I will try to create a most simplyfied workflow for faster testing the basics of this node.

If you get more informations meanwhile, please let me know.

@ArcleinSK If you like to get completely confused download my testworkflow from here: https://civitai.com/articles/23042 (choose LTXVAddGuide.json from the Attechments). It makes simply visible what the LTXVAddGuide nodes do.

Just enter a short video (it gets automatically croped to 25 frames) and a first- and last frame image. The workflow vae encodes the video to latent images -> runs the LTXAddGuide node -> vae decodes back to images. Finally you can compare the original images with the processed ones.

The node for i2v works well. But what the LTXVAddGuide node do is completely strange:

It sets the first frame image at position 0 (that`s ok), but then it sets allways 8 first frame and 8 last frame images at the end 🙄 (I ran tests with all available nodes).

You can even test it with the LTXVCropGuide node. Most strange thing is, it does not matter, wich values you set for frame_idx and strength.....

@ArcleinSK Just found this github documentation wich explains the frame_idx values.

@ArcleinSK Thank you so much for your comments at the article page here.

Just a quick reply for today: I`m pretty sure you hit the nail at the head 🙂 I allways wondered why the latent output of the LTXVCropGuides was not connected. Yes, with loosing the necessary information this will generating some nonsens for sure 😅.

I will have a look at your workflow soon as possible and trying to prepare a new update to fix this stuff. Your example video looks very well 👍 smooth at it should 🙂

Thanks to this amazing community we get better and better 🙂

@ArcleinSK ltx-2.3 combi 3.0 - Arclein edit: Very cool! Thank you 🙂 I will test this now.

What is your opinion - is there a greater demand for 3 frames (FMLF) for a couple of users? Cause switching between FLF and FMLF will need some more logic (at least 2 seperate LTXVAddGuideMulti nodes). For myself I was mostly too lazy to create just 2 frames (FLF) 🙄

Interpolator + Upscaler: I will test this soon. If this runs really fast, this would be a great job. I tested the "normal" image upscaler and the Rife framerate multiplyer in the beginning, like I used it for my Wan workflows, but this was much too slow with the higher LTX resolutions.

Qwen-Rapid-AIO: Yes, I allready heard from boinobin730 about a Qwen workflow to generate the frames. I really have to test it now 🙂Looks all like some sleepless nights 😅

I'll be back as soon as I've got a bit deeper into all this stuff here....

@ArcleinSK Thank you for your explanations. I think you are right.

Meanwhile I have the new workflow running. I had some ideas over the last days - so did a complete redesign of the "aspect ratio AND width/hight divisible by 32" calculations. Fortunately I found a pretty cool node wich combines this more "complex" inputs and calulations in just one node. This allowed me to significantly simplify the input and preliminary calculations in the subgraph, and also to add the option for <any aspect ratio> with width/hight divisible by 32 too. With these changes creating the swich logic for both options (FLF and FMLF) was much easier as expected.

I will run a few more tests and hopfully publish it soon. First quick pre tests ran without errors so far 🙂

@ArcleinSK Btw. I tested your Qwen workflow and started with "all kinds" of Qwen experiments. Amazing stuff. I tested Qwen a couple of month ago but gave up, as it didn't run properly with 12 GB of VRAM. But now it runs like a charme. I really did not know about the pretty things possible with Qwen. Thanks again.

@ArcleinSK Oh my - Photoshop 🙄 the good old times 😅

It is really unbelievable how well and easy Qwen can edid images - and I just did some first quick tests.

Do you know any Lora trainer for Qwen wich will work with 12 gb vram? During my last search some month ago there was nothing available for low hardware.

Btw. ltx workflow v4.0 is out now 🙂 hopfully bugfree.

@bobbobster Hi - and thank you 🙂 Short anwer: No.

Creating the previeous LTX workflows I tested both: frame interpolation and final upscaling, like I do in my Wan workflows. But with LTX it makes no sense. The slight improvement in image quality is out of all proportion to the significantly longer processing time. LTX-2.3 is made for 2k and 4k generation. Unfortunatelly with 12 gb vram we have to use much lower resolutions. So, if you have better hardware, first increase resolution and framrate and use beter models. This will increase quality significantly.

@bobbobster I forgot - even with low hardware, with ltx try allways to generate in the highest resolution possible. Happy generating 🙂

@VT95 Total frames are automatically calculated by the clip length you set via the slider node. And just in case you can not see or set any value in the sliders, please look her in the last comments. This is a known issue/node conflict.

@zorojim Hi - thank you. Wich option do you use?

With workflow v3.0 and the strict 16:9/9:16 aspect ratios I have not seen much artefacts anymore for myself, even with clip length up to 14 seconds.

Only FLF to Video has still some more serious problems, that’s why I have added the option here for automatic cut off.

@WigglyDad148 Do the basics: check twice you selected the right models. Destilled Lora is active and wight is set min. 0.5???

@WigglyDad148 Thank you. T2V works out of the box, no need for a negative prompt.

Ok. I assume you have cleared all node conflicts on your system and all components are updated (see my Troubleshooting guide).

Just to be sure, we are talking about workflow version v3.0 and you use my workflow without any modifications. That being said your blurry outputs are most likely a model mismatch. You have to check every blue loader node, if you have really selected YOUR local model. Please understand: downloading the right models to the right folder is one part, but you have to select these models in the loader nodes too of course.

@WigglyDad148 Yes, a lot of user forgot to "select" the models. I`m glad you got it running. Happy generating 🙂

@sexgod1979 Thanky you mate 😋

@sexgod1979 Yes, that was the intention, cause even with v2.0 it was a lot of stupid work to maintaine 5 or more nearly identical subgraphs just for some different inputs.

On the other hand, comfyui is very limited in "programming" options - no true runtime logic, no real subroutines to handle more complex workflows. Lots of buggy or very limited swich nodes (I tested nearly everything available).

Subgraphs are a good way to organize workflows visually, but they are not comparable with subroutines wich organize processing too. And subgraphs nested in subgraphs getting buggy as hell.... So I had to "redesign" my ideas several times.

And the most worse thing: at some point of complexity, with every node you add - error debugging in comfyui gets exponentiallly more painfull 🙄

But finally I learned a lot about comfyui itself and I believe the handling of the workflow is mostly usefull and easy as possible. Even if comfyui "newbies" have a lot of problems to get the nodes running and there was a "rebellion" against the new aspect ratio restrictions in the beginning 😂

Btw. I just started to test some of your ltx Loras. First quick results looking amazing 🙂

@Delavestra "any audio I have plus text prompt and initial image goes to video?" Short answer: yes.

Just a stupid example: You have 10 seconds of a pretty cool rock song with clear understandable female vocal and you have a nice image of a woman playing a guitar. Choose 10 second clip length and create a prompt like this: "The woman in the blue dress sitting in front of the camera plays the guitar and and passionately sings a rock song while the light dims and the camera zoomes in". With some luck this should generate a lip-synct video of the woman singing the song and playing the guitar.

And of course: Your audio + your image + your prompt (discribing what you want to see) should not conflict each other - unless you want to carry out some completely crazy experiments 🤣

@Setraether Run highest resolutions possible first, to get much better quality (edit the slider nodes, but keep in mind: aspect ratios of 16:9 AND width/hight have devisible by 32).

Use higer models like Q6 or Q8 and/or try higher frame rates.

@Setraether Uhh - with higher resolutions I meant at least 2k -> 2048 x 1080 of course.

Sorry, but without a link to your other workflow it is not possible to compare anything.

@Setraether Btw. the workflow has 7 sliders at all.

But I guess the problem is more, you asked a question for very skilled comfyui users at a pretty advanced level, but it seems you are struggling with the comfyui basics in practice.

Let me know if I`m wrong, otherwise start simple as possible with t2v and try to understand step by step what is possible and what not.

Even this workflow is made for advanced users, a lot of comfyui "beginners" trying to get it running - and with the right questions you mostly will get some helpfull hints here.

@Delavestra Interesting. I had a quick view at the simplest ID-Lora workflow. But why integrating it here? The whole workflow are just a handfull of special ID-Lora nodes. So this is a complete different "architecture".

Just use it at it is. But I guess the main problem would be to get it running with 12 gb vram.

Yes, there was a "rebellion" against the restricted aspect ratios after publishing version v3.0 😂 and of course I know, other aspect ratios will work too generally. The problem is to arrange this in a "user friendly" "interface".

The simplest way would be to offer sliders for width and hight (restricted to divisibility by 32 only). This would also reduce the amount of calulations in the subgraph substantially. But the con is: You always have to "measure" your input images/videos manually and to set manually two input values near your measured values wich are hard to remember each time - not to mention you want to have a certain aspect ratio too (like 16:9 or 4:3 or what ever you want), cause keep in mind: your manual values have to represent your aspect ratio AND divisibillity by 32. You really should try this "simple" task with your pocket computer for some aspect ratios 😉

So in my opinion, the most comfortable way is at it is: just input the longer edge and the algorithm takes care of everything else in the background.

If you have a good idea to solve this, please let me know.

@haymaker No, with any earlier versions, like v2.0, there were a lot of reports of issues with artefacts, bad prompt following, bad image qualitty, etc. So I definitely will not go back without a solution of the above mentioned math.

@biodizain Hi, sorry - this is actually not on my priority list, cause I allways try to get out much quality as possible and as I can see from the comments, that seems to be the common use case.

To implement a real one pass option just to reduce generation time for maybe 20 - 30 % and probably much lower video quality would be a lot of work and would completely run against the actual used swich logic. For myself I just cancel a generation, if I see it goes in the wrong direction.

So, short anser is: No. Btw. New version v4.0 is out 🙂

@raylar529 OOM errors: ram or vram??? I guess ram. Increase your swap file as specified. You should run up to 14 seconds with high resolution (longer edge = 1536) without issues. Btw. use my new version v4.0 now.