Clean X2 E26 video VAE is a 2× spatial video VAE for MiniMax H3, providing cleaner high-resolution decoding, improved detail preservation, and reduced grid artifacts.
Compatible with the X2 VAE decoding workflow.
This node package is required: https://github.com/TripleHeadedMonkey/ComfyUI-MiniMaxH3_LatentUpscaler
A little more information about CLEAN X2 E26 VIDEO VAE H3
No external upscaler.
The frames you're looking at are decoded directly from CLEAN X2 E26 VIDEO VAE H3 at 2× spatial resolution.
This project started because I wasn't satisfied with simply generating an H3 video and throwing an AI upscaler on top of it.
I wanted to know:
How much more can we get directly from H3's own decoder?
What followed was... a lot more work than I expected. 😅
Why is it called E26? 😀
E26 stands for the 26th major experimental iteration — but there were countless smaller experiments, training runs, failed ideas, A/B tests and intermediate versions inside those iterations, sleepless nights and the desire to give up 😅.
But I didn't give up, and here we are 😀👋🎉🥳🥹
What it does
-Native 2× VAE decoding
-No external AI upscaler
-No post-generation super-resolution model
-Clean output without the old repeating grid
-Works directly with real MiniMax H3 generations
-More spatial information than the normal decode
Description
Clean X2 E26 video VAE is a 2× spatial video VAE for MiniMax H3, providing cleaner high-resolution decoding, improved detail preservation, and reduced grid artifacts.
Compatible with the X2 VAE decoding workflow.
This node package is required: https://github.com/TripleHeadedMonkey/ComfyUI-MiniMaxH3_LatentUpscaler
A little more information about CLEAN X2 E26 VIDEO VAE H3
No external upscaler.
The frames you're looking at are decoded directly from CLEAN X2 E26 VIDEO VAE H3 at 2× spatial resolution.
This project started because I wasn't satisfied with simply generating an H3 video and throwing an AI upscaler on top of it.
I wanted to know:
How much more can we get directly from H3's own decoder?
What followed was... a lot more work than I expected. 😅
Why is it called E26? 😀
E26 stands for the 26th major experimental iteration — but there were countless smaller experiments, training runs, failed ideas, A/B tests and intermediate versions inside those iterations, sleepless nights and the desire to give up 😅.
But I didn't give up, and here we are 😀👋🎉🥳🥹
What it does
-Native 2× VAE decoding
-No external AI upscaler
-No post-generation super-resolution model
-Clean output without the old repeating grid
-Works directly with real MiniMax H3 generations
-More spatial information than the normal decode
And will try to improve it
FAQ
Comments (14)
Why i required latent upscaler node pack? I think this vae just replace stock video vae. Or am I wrong?
@nafikovtt777 hello. Yes you're wrong 🙂 because this is 2x VAE and without minimax h3 vae decode (fast) node from upscale node pack this VAE will give you only error, it doesn't work as stock base Video VAE with base VAE decode node. Different math. This VAE give 2x size output
I add how look this node in description
I just bought it. Is there a image of the workflow connections for the VAE mentioned below?
@SkycladsquirrelArt just use any WF and replace base node vae decode with node from node pack I mentioned I show in description
@SkycladsquirrelArt I add WF to additional files
@speach1sdef178 I just wanted to say Thank you. I got it working. I have to use regular VAE at the beginning of the workflow at use the X2 at the decode.
Tested at 0.6 MP. Compared to the original VAE, it only makes the image slightly sharper, but nowhere near matching a 2x resolution increase—there are no extra or improved details whatsoever. VAE processing speed is noticeably worse than using the original. The author's workflow also includes 2 unlisted custom extensions (why include them in the example workflow if they aren't required?). Overall, no miracle happened—classic sampler-based upscaling remains the best option.
@asfbx Please pay attention!!!
The main use case of the 2X VAE is very simple: it decodes the same H3 latent at twice the spatial dimensions. For example, a latent that would normally decode to 640×480 with the standard H3 VAE can be decoded to 1280×960 with the 2X VAE. H3 itself is still generating the same latent — it is not being asked to generate at 1280×960. That is the reason the 2X VAE exists. “Preserving detail/quality” means that increasing the decoded resolution should not come at the cost of noticeably degrading the reconstruction compared with the normal VAE. It does not mean that the 2X VAE magically creates twice as much fine detail. The “Detailed” mode was a separate experiment attempting to recover additional detail, and the write-up explicitly explains that the original latent-only approach did not succeed as intended
Yeah I don't think you counted the pixels coming out, because there was 2x of them. Resampling is an optional step that can still be done afterwards.
@speach1sdef178 So tell me, what practical application does your VAE actually have? Even without using it, I can just load a video at original resolution into Topaz Video and get 2x or 4x in a few seconds—and with better quality to boot.
@asfbx Everyone finds their own application, and I've already shared mine. I think you already have a better option. Whether to use it or not is up to you.
@speach1sdef178 I think this one has better face recognition since it decodes at 2x the resolution...
INFORMATION FOR ALL!!!
The main use case of the 2X VAE is very simple: it decodes the same H3 latent at twice the spatial dimensions. For example, a latent that would normally decode to 640×480 with the standard H3 VAE can be decoded to 1280×960 with the 2X VAE. H3 itself is still generating the same latent — it is not being asked to generate at 1280×960. That is the reason the 2X VAE exists. “Preserving detail/quality” means that increasing the decoded resolution should not come at the cost of noticeably degrading the reconstruction compared with the normal VAE. It does not mean that the 2X VAE magically creates twice as much fine detail. The “Detailed” mode was a separate experiment attempting to recover additional detail, and the write-up explicitly explains that the original latent-only approach did not succeed as intended.