Comparing WAN 2.1 on COMFY vs. FramePack using an RTX 3060

[LIQUIDMIND111]

BOTH apps i used TEA CACHE and the SAME:

PROMPT
an alien organism with protuberant eyes is swimming underwater.

Size: 512x768 set on comfy same as Framepack outputs

steps 25

Length 5 sec

COMFY = 34:21<00:00, 82.46s/it
FRAMEPACK = 28 minutes and 20 seconds.

WAN =

WanVideo_00067.mp4

Frame pack =

250419_100632_723_5975_37.mp4

[iamrohitanshu]

Instead of no. of seconds, can you tell us the no. of frames, please?
As unless you also tell us the fps, and whether you used frame interpolation or not, we can't determine the no. of frames.
Thank you!

[LIQUIDMIND111]

Instead of no. of seconds, can you tell us the no. of frames, please? As unless you also tell us the fps, and whether you used frame interpolation or not, we can't determine the no. of frames. Thank you!

AHH true, thanks, good point...

well on COMFY WAN at 15FPS i did 77 frames, which is 5 seconds almost exactly.
on frame pack, they use 30FPS, so if it's 5 seconds also, then it's like 150 frames???

NO INTERPOLATION ON EITHER.

[133tech]

Nice comparison

Could you please share some resource on how to run WAN 2.1 on 3060? I thought it needs way more than 12 GB of VRAM

[LIQUIDMIND111]

Nice comparison

Could you please share some resource on how to run WAN 2.1 on 3060? I thought it needs way more than 12 GB of VRAM

i started with 2060 6gb, so it works well with 3060 12gb,,,, comfy has its own worlflow already... a trick is, dont do the OG size, if yo want to have faster generatios, do this..... use 1.3b model for FUN and the 1.3 FUN MODEL for MORE FUN, lol, now.... when using IMAGE TO VIDEO, i use 14b 720p Q6 with tea cache and up to 93 frames of more, BUT i do 480x720 instead of 480x832 and i get AMAZING results.... and 30 steps.......... its like 20 minutes...... but you can reduce STEPS and SIZE and FRAMES for faster generation.,

[133tech]

i started with 2060 6gb, so it works well with 3060 12gb,,,, comfy has its own worlflow already... a trick is, dont do the OG size, if yo want to have faster generatios, do this..... use 1.3b model for FUN and the 1.3 FUN MODEL for MORE FUN, lol, now.... when using IMAGE TO VIDEO, i use 14b 720p Q6 with tea cache and up to 93 frames of more, BUT i do 480x720 instead of 480x832 and i get AMAZING results.... and 30 steps.......... its like 20 minutes...... but you can reduce STEPS and SIZE and FRAMES for faster generation.,

Oh yea, so I should get some quant model of the WAN like eg.: https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf
and lookup some comfy workflow, got it thanks!

[LIQUIDMIND111]

i started with 2060 6gb, so it works well with 3060 12gb,,,, comfy has its own worlflow already... a trick is, dont do the OG size, if yo want to have faster generatios, do this..... use 1.3b model for FUN and the 1.3 FUN MODEL for MORE FUN, lol, now.... when using IMAGE TO VIDEO, i use 14b 720p Q6 with tea cache and up to 93 frames of more, BUT i do 480x720 instead of 480x832 and i get AMAZING results.... and 30 steps.......... its like 20 minutes...... but you can reduce STEPS and SIZE and FRAMES for faster generation.,

Oh yea, so I should get some quant model of the WAN like eg.: https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf and lookup some comfy workflow, got it thanks!

i have used the fp8 models that are 16GB and they are awsome, BUT GGUF seems to work well, depending on the version....

[133tech]

i started with 2060 6gb, so it works well with 3060 12gb,,,, comfy has its own worlflow already... a trick is, dont do the OG size, if yo want to have faster generatios, do this..... use 1.3b model for FUN and the 1.3 FUN MODEL for MORE FUN, lol, now.... when using IMAGE TO VIDEO, i use 14b 720p Q6 with tea cache and up to 93 frames of more, BUT i do 480x720 instead of 480x832 and i get AMAZING results.... and 30 steps.......... its like 20 minutes...... but you can reduce STEPS and SIZE and FRAMES for faster generation.,

Oh yea, so I should get some quant model of the WAN like eg.: https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf and lookup some comfy workflow, got it thanks!

i have used the fp8 models that are 16GB and they are awsome, BUT GGUF seems to work well, depending on the version....

stupid question but - using the fp8 16GB on a 12 GB VRAM card wouldn't work right? Maybe with a slowdown swapping to RAM?

[LIQUIDMIND111]

i started with 2060 6gb, so it works well with 3060 12gb,,,, comfy has its own worlflow already... a trick is, dont do the OG size, if yo want to have faster generatios, do this..... use 1.3b model for FUN and the 1.3 FUN MODEL for MORE FUN, lol, now.... when using IMAGE TO VIDEO, i use 14b 720p Q6 with tea cache and up to 93 frames of more, BUT i do 480x720 instead of 480x832 and i get AMAZING results.... and 30 steps.......... its like 20 minutes...... but you can reduce STEPS and SIZE and FRAMES for faster generation.,

Oh yea, so I should get some quant model of the WAN like eg.: https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf and lookup some comfy workflow, got it thanks!

i have used the fp8 models that are 16GB and they are awsome, BUT GGUF seems to work well, depending on the version....

stupid question but - using the fp8 16GB on a 12 GB VRAM card wouldn't work right? Maybe with a slowdown swapping to RAM?

ok well, you have some issues here: first you have 2 text encoders one is 6gb another is 11GB - then - you have the issue WHICH MODEL ARE YOU USING? a GGUF with 6 to 8 GB or a GGUF with 12GB.... ok?

then do the math..... IF you use a model that is 6gb and a text encoder that is 6, you might have some GOOD SPEED, but, if you use any other GGUF model that is already 12GB, plus the 6gb text encoder, then YES you send some memory to RAM...... so it becomes SLOWER. but still ok to render, its not HOURS....... but 20 minutes for a 6 second clip, could be average waiting..... NOW, if you use the 11gb text encoder, and a GGUF or regular FP8 that is 16gb, then SLOOOOOOOOOOOOOOOOOOOOOOOOWWWWW AF.......

[LIQUIDMIND111]

FRAMEPACK UPDATE!!!! 18 minutes with triton and sage, 10 minutes saved!!

250419_195435_205_3357_37.mp4

[Emieanderson]

WHICH IS WINER?

[LIQUIDMIND111]

WHICH IS WINER?

Quality, WAN, prompt adherence, WAN,
easy way to extend long videos, FRAMEPACK.