I didn't bother pixel-perfecting the nodes or lining them up neatly—as long as it’s clear what goes where and why, the rest doesn't really matter. Wasn't aiming for aesthetic perfection, and honestly, didn't have the time either! =)). Recommended page file size: at least 128 GB. Generating 5 seconds of 1920 by 1088 at 48 fps on 5060ti 16 + 64 ram takes about 500 seconds, 490-530 seconds of 1920 by 1088 at 24 fps on 5060ti 16 + 64 ram takes about 290-320 seconds =)) For optimal performance, NVMe speeds of 3000 MB/s to 3500 MB/s + . The best choice in terms of speed and quality | 0.5 | 16:9 | 1920 x 1088 | 9 | 3 |, perhaps it is still possible to optimize the process and finalize it, but there is no free time yet, it is even better to wait for the optimizations of the model and the associated nodes =))
Description
Changes were made to improve handling of the I2V reference.
Node control has been expanded, the interpolation control toggle has been refined, and an additional lightweight mode has been added.
A pre-upscale preview has been added; this is useful for spotting errors in your prompt immediately,
to be able to interrupt generation before it finishes, the prompt plays a major role and can completely break your reference,
they also allow you to save metadata in case you suddenly need to restore a project, etc.
The notes regarding various components have been expanded and supplemented.
Minor issues regarding node signatures have been fixed.
P.S.
For maximum generation I2V quality and to avoid artifacts at tile seams, the input image resolution must be strictly divisible by 32 pixels. Using automatic adjustment nodes (such as Pad/Resize) can shift the composition. For ideal results, manually align and crop the canvas to the exact proportions required.
There are ideas on optimizing T2V to speed up the process, as soon as I have free time, I will assemble the project =))
FAQ
Comments (6)
Just want to say thanks very much for your workflow. I did add an Evict Text Encoder node to free up VRAM after prompt processes and I'm generating 7 second 4k videos in around 4 minutes with no turbo lora required which is pretty good IMO. The LTX Face Identity Reinforcer does an OK job, but when I tried running Euler_CFG_PP on my RTX5090 30 minutes in the sampler was still sitting at 0%, so I ran out of patience with it. As you mentioned, audio desyncing with the LTX refiners is a little bit annoying sometimes. Have you experimented with instead of diffusion using frame by frame ERSGAN or 4xUltraSharp upscalers at all?
Also - I saw one of your examples was T2I. Did you somehow make Minimaxh3 into an image generator?
Thank you, there are already ideas on how to improve the quality and speed, but it requires tests, unfortunately there is no time. I have a rather modest video card, full tests of working capacity, etc. they take a lot of time =)). There are no problems with the sound if you did not change the node connections from the sound output to the save nodes, since they are tied to bypass LTX upscale and immediately go to the save node without causing distortion or lagging the sound 3 main operation modes 24 def 48 lq and 48 hq the correct synchronization is prescribed everywhere, there is a nuance if you set the VideoCombine or alternative ways to save video on node connections, perhaps the number of frames is not correctly specified there, from node connections and modes 24 def 48 lq and 48 hq in this case, you could get sound out of sync., you can see how much sound is correct in the Preview MiniMax H3 node to see if the model generated the voiceover correctly initially. For the sake of interse, I checked ERSGAN or 4xUltraSharp or the alternative version once. If you use this additional upscale, then only between RTX and KJ sharp nodes the process goes out significantly, in some cases it may take twice as long and as for me it makes little sense, I would rather use the FlashVSR ready-made project for the subsequent resolution increase.
@isaacg72211480 Maybe my tag is spelled out incorrectly =)) or you saw it on the video and the image is signed as an example and it says that this image is from T2V but it is generated in Krea2 =)) in general, if you go too far, you can make a MImimax image generator, but it doesn't make sense to anyone just for fun =)) When I updated the workflow version, part of the text disappeared somewhere and for some reason appeared on the side and not in the center on the model page, I would have to correct it later, I didn't figure out how yet =))
@Nikolos747 Thanks for clearing up that mystery about the Text 2 Image I thought you hacked MiniMaxH3. And I get your point KREA 2 is already pretty sufficient for generating the first frame images for now. I didn't explain that audio thing very well - like you said the audio is fine, it is specifically the lip sync that suffers a little bit due to the LTX Refiner pass. Not a bad trade off though for the detail it gives. I suspected as much about the ERSGAN/4xUltraSharp - it's just going to be too inefficient. Anyway, your workflow is awesome I learned a lot from it and hope you'll keep updating it as things change.
@isaacg72211480 I'm sorry, I misunderstood what you meant when you were talking about sound, with lip sync, when working with Upsacle, have problems installing below 0.5 megapixels, or if the character is too far from the camera, I'll conduct a separate test and try to figure out this issue later., it may also be due to the initial generation even before the Minimax upscale that lipsink synchronization errors were made or the character was far away and didn't even move his lips at all and then it was transferred to LTX upscale in this case, upscale can aggravate the problem and in this case it may be solved in future updates, let's see =)) I forgot to mention before that Euler_cfg_pp yes, it is very heavy, I even have a Note in the fork itself about a possible drop in speed up to 250% in some cases.... I plan to finish the T2V version, then return to the I2V revision =))