@vrgamedevgirl Btw... this is Dakhari from your Discord :)

所以，我可以抛弃那个十几个G的基础模型了吗，解放我的硬盘空间

lora和ckpt的效果无差异吗，还是ckpt更好

@vrgamedevgirl  yeah switched to the regular version and it works fine. Now I'm messing with your lora trying to find the sweet spot. :)

@vrgamedevgirl I made multiple try and it seems to mainly be on lower quality image while most of the time hq images generated with sd seems to be fine, so that should be good for me for most stuff.

@vrgamedevgirl or you can use the wan base model inpaint on 2nd ksampler with 0.35 denoise strength to enhance the face so the face back to normal.

@vrgamedevgirl Can you share it without the mps reward lora? I can add the mps lora myself and play around with it.

I've taken to using VACE and pulling a random face from a folder downloading a database from something like fairface. Then you have a largely unlimited number of authentic faces. Not celebs, not supermodels, regular people.

@vrgamedevgirl Ok, I'm trying. It seems to work when I reduce the weight of Fusionx and add causvid etc. Thank you. I didn't know that making a new lora was difficult, I wasn't aware of this when I wanted it.

Thank you for fusionx too.

@vrgamedevgirl Can you post a link to MovviGen lora I can't find it anywhere. Also please just post your ratio recipe here, I don't have discord.

@vrgamedevgirl I have tested MPS rewards lora before. When its weight is 0.25 or below, most of the time it does not change the appearance of the character. So the fusion model cannot adjust the weights of MPS rewards lora alone, which is indeed a minor issue. Besides, thank you for your work. It's great!

@vrgamedevgirl I said that I had tested MPS rewards lora before, which would change the appearance of characters and could only reduce the weight, but it was not completely accurate. Previously, I only tested anime characters. Later, when testing MPS rewards lora with a weight of 0.8-1, it worked well for real people, but not well for 2D characters. Other than that, the effects were all very good. Thank you again for your model.

@vrgamedevgirl Cau you share Causvid, Accvid loras for WAN-I2V-14B model? Thanks

@vrgamedevgirl Sorry, judging by your nickname you are a girl :)). So the usual Wan 2.1 14B model in gguf format reduces the time (3 seconds of video generates in +-17 minutes), also not fast... The most optimal option for me is to use Wan 2.1 models with 1.3B parameters... but unfortunately I did not find a Wan 2.1 model for i2v. And if the 1.3B models are converted to gguf, will it be even faster?

Hi I have 100kB VRAM. y wont ur workflows generate 8k HD awesomesauce in 2.5 seconds? Please condense to 100kB thank you.

@vrgamedevgirl Thanks for the quick response! I had been to that page and saw the full models, but missed the lora until I searched the page for MoviiGen. I'm going to emulate yours as much as I can, and mess with MPS strengths to see what they do. Are you still using:

AccVid 0.50

MoviiGen 0.50 (in T2V)

MPS Rewards 0.70

CausVidV2 1.00

Your Realism Boots 0.40

your unpublished detailer 0.4

@vrgamedevgirl Even at 0.2 MPS still saturates the color on skin. Unfortunately I see what you mean about the improved images otherwise with MPS. Not a good tradeoff.

Try loading comfy with the options --low-vram --reserve-vram 0.5 or --reserve-vram 1.0. Don't use block swap or that multigpu node, both are rubbish and slow. Just use native workflow and reserve some of your vram so that you don't get oom when it tries to create the preview or you open another browser tab. Oh and if you also lack system ram, then increase your swapfile size to be very big :)

use regular ram. use the UnetLoaderGGUFDisTorchMultiGPU node and set "use other vram" to true, set "virtual_ram_gb" to as much regular ram as you actually have. if you still can't do it you're probably cooked, but you could still try turning off sage attention and that might help.

Shift

@AIBeauty4Seen with a cfg of 1, did negative have really an impact ?

For me it's the reverse it's way too fast... like a big length video being played 4 times faster. The lower the steps the faster it gets. Am i missing something there ?

Edit: after some tests i can help you:
More CFG = More fast pace (adhere better to lora and prompts)
Less CFG = Less motion (adhere less to prompts and loras which needs to be tweaked to fit)

Also you will need to reduce the Start_Latent_Strength and/or End_Latent_Strength to fit.
This lora is very interesting :)

@Meandre66 I would give it a try. if your using ComfyUi...I am stilling learning. I am using wan 2.1 on pinokio. If it works...let me know. I am curious to see if it helps

@akak7952 shift what? reduce the shift? or increase the shift?mplease let me know

@Meandre66 did you fix the problem????

@vrgamedevgirl @vrgamedevgirl  can you provide the link, cant find it

@vrgamedevgirl  look awesome! can i suggest adding reseize image node {DEPRECATD) and any power lora loader... and prompt generator from Florence , lol!!! sorry for asking that much!! sorry sorry

@vrgamedevgirl I really appreciate your work

what MPS ?

@vrgamedevgirl Nice work! May I know which lora is changing the overall colour of the output? I couldn't figure out which one is giving the orangish tint effect...

@vrgamedevgirl wait for another merge,appreciate:)

Thanks. I tried that - spent 8 hours on it yesterday. Overall, if you generate in 480P+ resolution, there aren't major issues with faces or anything else. Soon I’ll be testing the LoRA on a huge generated video dataset (10,000+), which should make it easier to identify pros and cons. If you're interested, I can post the results here later.

I’m not training. I allow users to generate videos for free in my project. What I meant is that project users will generate 10,000+ videos using your LoRA combined with various other LoRAs, including different values provided by Remade. A couple of hours ago, an update went live in my project, and video generation has already started. I'll wait for a large sample of videos and then analyze it. Overall, initial results look very promising.

Offtop. I don't understand why on windows a can get inference time [00:49<00:00, 6.19s/it] (81 frame 832*480), but on linux i get only 65 frames for this time... 8/8 [00:48<00:00, 6.03s/it]

I’ve downloaded all your workflows—since only T2V and I2V are available right now, I already understand how they’re combined. I’m just not sure if Phantom LoRA uses a different setup.

@vrgamedevgirl What are those lightning workflow ? I'm interested if i can speed up without loosing quality. (At least 5-10% quality) would be great ! Actually i can render 101 frames in about 3 minutes with a 5090 at 624x624 resolution.

"In da pussa!!!"

ProvenFlawless This made me laugh out loud. Family Guy reference?

with my preliminary tests, wan2.1 loras seem to work well with 2.2

nathanhalko Well, console says something like that

lora key not loaded: diffusion_model.blocks.7.cross_attn.k_img.lora_up.weight

Usally i see that message when lora not lading propperly.

I try I2V

nathanhalko Hm. Using T2V version doesnt reproduce error.

vrgamedevgirl Thnx. No rush. Just good to know you will update =)

vrgamedevgirl thanks, the FusionX version of Wan 2.1 is so great that I might not even update to Wan 2.2 but I'd like to try it out anyway just to compare :)

vrgamedevgirl Extremely helpful as always, joining your discord was definitely a good idea. Looking forward to see what you manage to cook up for Wan 2.2. 🙏🏻

@nathanhalko where do you place it

I don't think so. The version with high noise reduction is completely unusable

Can u share what did you do excatly?

do you mean you cant find the lora file to add it?

I'm having the same issue, it's in the correct folder, just not showing up in the Load LoRA node