minimax_h3_video_vae_int8_convrot.safetensors
When using this Vae instead of the fp16 version, it produces a black video with sound. Your preview mode will show the video but the saved media file produces no image. I've tried many different media players to make sure it wasn't a plugin issue but nope, nothing but blackness. Is this because I am not using the w4a8 model and still using your int8 convrot with this new smaller VAE? I haven't been able to try this w4a8 yet seeing as Comfy isn't compatible with it at the moment. Also, is there really any benefit to using this vae over your fp16 version. Speed? Vram difference? Or is its only purpose is to save on disk space? Greatly appreciate any info.
Try updating comfy ui, because it's working for me well. I use int 8 pruned ref2va model, output is producing correct video.
Try updating your comfy ui
thanks for the tip. that did it. No notice in quality difference. It uses the same amount of vram and memory but I save 20 to 30 seconds on a 1280x 5sec video compared to the bf16. Every sec counts. Thanks Kijai for making it. Now I'm off to test the w4a8 since it apparently works now with the latest comfy. Thanks again shadow for the response.
thanks for the tip. that did it. No notice in quality difference. It uses the same amount of vram and memory but I save 20 to 30 seconds on a 1280x 5sec video compared to the bf16. Every bit counts. Thanks Kijai if making it. Now I'm off to test the w4a8 since it apparently works now with the latest comfy. Thanks again shadow for the response.
Glad it works. The w4a8 still has some issues with LoRAs that I'm working on to fix.
hmm, seems the w4a8 still doesn't work with the latest comfy version. I still get UNETLoader 'asym_w4a8_int8' error. I thought it was updated for it on account of the post from shadowmonarchforreal requesting for the other model. Oh well, perhaps when it finally does get added support, those issues you speak of will hopefully be resolved. Thanks a great deal for all you do here Kijai. This scene just wouldn't be the same without you.
Hi guys, all updated, black screen. Using Sage and Spectrum. Yesterday I had funny blinking blocks, now it's all black. Thanks for any tip. PS. with int8
Did you use 'update_comfyui.bat' or 'update_comfyui_stable.bat'?
I know in the past when I ran the stable bat file there was still certain things missing that I needed that was not installed when using that bat file. When I swtiched to just update_comfyui.bat it would give me the updated files I needed. This may not fix it for you. Just thought I would mention it. I ran that bat file earlier and it fixed it for me.
I used the button inside comfy, from the manager, but I remember having that effect in the past once, what you said. I try the batch. Thanks for making me remember. I will post the result of my undertaking here.
It worked. Thanks. This method changes the build to nightly, and......it kills all symbolic links. Thank you again, also KJ for all his beloved contributions! Freed up 8GB System Memory also.
Sweet, glad it worked. After more testing I literally am saving 20 to 40 seconds on my 5 seconds vid generations thanks to this vae. Kijai is awesome. I am curious though if the same results can happen with the audio vae.
Is the audio VAE not already instant? I never even noticed audio decode taking any time.
Forgive my ignorance. You are correct. I never paid any attention to the vae files until today after switching to your smaller video vae. I also forgot how small the audio file is already.
updated to comfyui nightly version, bypassed minimaxh3-cache node, but still "vae_int8_convrot" produces a black video with sound. Actually Sampler Preview shows video, but after decoding , on VHS Combine node black screen with audio as final result.
It works, but I'm not seeing any performance difference on a 5090 vs fp16 vae. I just load it in the default "Load VAE" node on nightly Comfy. Is that the right thing to do? Any gotchas I should be aware of?.. Thanks!
When it does its decode at the end of the generation is when you will notice some time saved. For me it fluctuates between 20 to 40 seconds. It doesn't speed up the generation time or save on memory. 20 to 40 seconds may not seem like much, but it adds up big time for me seeing as I do a lot of generating.
Also, Im using a 5070 so maybe with your 5090 it won't change anything seeing as its in a whole other class compared to my slow poke card.
The VAE is worth switching to. Unfortunately I can't say the same for the new w4a8 model. Unless your just looking to save on hard disk space, this one isn't worth using. The generation times are exactly the same as the pruned_int8_convrot model and even though w4a8 is half in size, it actually uses more vram then the 20 gig int8.
I was hoping I could increase the duration by 1 or 2 seconds before it maxed out my vram but it actually did the opposite and I had to lower the seconds by 1 or 2. Strange. Oh well, at least my curiosity is now satisfied. Smaller size doesn't equal less vram.
The VAE is worth switching to. Unfortunately I can't say the same for the new w4a8 model. Unless your just looking to save on hard disk space, this one isn't worth using. The generation times are exactly the same as the pruned_int8_convrot model and even though w4a8 is half in size, it actually uses more vram then the 20 gig int8.
I was hoping I could increase the duration by 1 or 2 seconds before it maxed out my vram but it actually did the opposite and I had to lower the seconds by 1 or 2. Strange. Oh well, at least my curiosity is now satisfied. Smaller size doesn't equal less vram.
This is because the offloading in ComfyUI is asynchronous, and the compute on this model is so slow that the offload transfer overlaps and is pretty much fully hidden, so you don't gain anything from smaller model. Smaller model is only useful when the model is larger than what your combined VRAM/RAM can handle.
Thank you for the explanation. I learned something new. Useful info to have for future models.