darksidewalker What's the "Lightspeed" version compared to the regular Vision H version?

darksidewalker Please don't take this the wrong way. I just curious the film vfi node I set it`s wrong , I think it can reduce a bit the execution time if I change the right parameter or add some node.

darksidewalker im assuming he cant mount the model, probably running out of Vram

"The point is that Wan 2.1 Lightspeed is a pre-optimized model with FP8 quantization and built-in attention mechanisms, offering faster video generation and lower hardware requirements without needing LoRA setup, while maintaining similar quality to Lightx2v LoRA with less complexity and consistent style."

This answer is not mine, it is the response I received from Grok 3 when I asked him your question. I'm not an expert, so I don't know if this answer is accurate or not.

grrrr It makes no sense! "without needing lora setup", yeah.. same description would be better applied to the lora "without needing 15gb file and clips"..

and what does it mean "with less complexity and consistent style"? is less complexity good or bad? is the style less consistent or are we meant to interpret it as meaning more consistent?

sibdwqwawmndbdolrf674 I discovered this model while searching for ways to generate Wan videos faster. After creating my first video, I learned about other options like the FusionX model, the LightX2v you mentioned, and CausVid. As you pointed out, adding a single LoRA to an existing workflow isn't complicated, so I can understand why you'd question the point of this model's existence.

But consider another perspective. As I mentioned, I initially found this model when looking for faster ways to generate Wan videos. I'm not a native English speaker, and while I use translation tools, acquiring technical knowledge isn't as easy for me as it would be for native speakers. So for people like me who are first encountering this technology, something more accessible and straightforward is needed. I actually found this model helpful when I used it, and that alone is sufficient justification for its existence. You might feel this model is redundant and unnecessary, but that could be because you're already an experienced user.

darksidewalker the brightest day in my life .. followed by the saddest one. Well, let's hope somebody else will. I mean who needs T2V, but FLFV would be really nice. Thanks anyway !

darksidewalker And you did it anyway! Outstanding job on this

same. Swarmui, can'Ät get it to work :-/

If you have a sageattention node in your workflow, turn it off, and if it's in the startup options of comfyui, try deleting it.

@DoroArmy Thanks for the tip. That was my fault. After turning off the sageattention the workflow runs.

Works fine on my 4070 Ti.

save the checkpoint into diffusion models, not stable checkpoints, change load checkpoint node to load diffusion model, set dtype fp8_e4fmn(or whatever that is, the one with 4 not 5) and it should work.

fp8_e4m3fn

darksidewalker You already helped to save us a big amount of effort & time :) , I have a machine armed with RTX 4080 as main GPU & also RTX 3090 as second GPU fed by a separate 1000W PSU, your custom LightSpeed model was what I need , But I Only downloaded the 480p version. the 720p version looks tempting but honstly I got tired of the never ending of downloads for everything (checkpoints , Loras , ...etc) so I wanted to just choose & GO.

darksidewalker would you be willing to share your process to do this single checkpoint conversion with all speedup tools included so I can try to do the same for T2V?

Awesome, you da man

Tested another image used in a previous run with "regular" Wan 2.1 I2V 480p + CausVid and it worked fine with this model. Hmm. Same image has rendered with no changes with other models though. Is there anything particular about image resolution, dpi, etc. I should be aware of?

Tried a bunch of different images, almost all have the same thing. Checked every setting, but I'm VERY new to this, so I'm probably just missing something. I'm using SwarmUI - should I be using a custom workflow (like the one you have here)? As far as I know everything's just default other than the settings I've changed as you recommended. Thanks!

darksidewalker Well I figured it out (I think). I was using init images that were zoomed way too far out from the subject(s), thus making it infer the greater and greater detail as the camera zoomed in. Combine with less than tack sharp photos and more complex camera and actor movements and you get face mush, changing hair styles, etc. Did confirm almost identical issues when using your other GGUF + CausVid, and even the regular fp8 safetensors file + CausVid had issues if maybe SLIGHTLY fewer.

Also: REALLY need to listen to your recommendations on LORA strength. :D

I was just asking too much of our AI overlords. Thanks for the great work man!

This absolutely works in 12GB cards.

just tested on 16GB vram, works perfectly. 2:45-2:55 seconds 16fps 81 frames. i just used the sample workflow. switched load checkpoint to load diffusion model and it loaded.

Oh, sorry. My RAM is 16GB but my VRAM is only 8GB.

newsatrandom22785 16GB of ram has been obsolete for at least 8 years now. 32Gb is the minimum recommended ram for any computer.