Merge of my favourite Wan image enhancing loras, makes it much faster to use than when adding the lora separately.
Samplers
In my testing/research, I have found the best Sampler is Res_3s with Bong_Tangent Scheduler both from the RES4LYF custom nodes pack: https://github.com/ClownsharkBatwing/RES4LYF
I am using this great free txt2img workflow: https://pastebin.com/GPYQjUrx from AItrepreneur check out his Patreon https://www.patreon.com/c/aitrepreneur/posts
V2 is WAN 2.2. Merge that can do incredibly low steps down to 2 steps.
4 steps look better and 8 steps is great quality.
If you want a bit more realism, you can use the LightX lora with a negative weight, but you might have to then increase steps.
Can make decent images/videos in as low as 6 steps as it has speed loras mixed in.
Description
A merge of Wan 2.2 and My V1 mix
FAQ
Comments (30)
please v2 Pruned version upload..thanks
the fp8 model for V2 is uploaded now: https://civitai.com/api/download/models/2059794?type=Model&format=SafeTensor&size=pruned&fp=fp8
GGUF would be nice :P
Yes please! For us who are poor in VRAM ;)
Only Lora possible?
(WAN2.2) i'm not signing up to someone's patreon. Is this 5B or 14B?
It says right in the description it's 14B
The quality using your merge and the workflow you shared on Reddit are quite good. The only issue I'm finding is it is a bit slow on my mobile 5090, which I can live with, but unfortunately every subsequent generation takes an incredibly long time because too much data stays in my system memory. After the generation completes it is using 72% of my 64GB. :O
I should add that I have 24GB VRAM.
I haven't seen the workflow, but have you tried incorporating a purgevram node?
TeeKay VRAM is fine. It's my system RAM that's the problem. Python is using over 30GB by the time the first image is created.
TeeKay Yes Clean VRAM Used and Clear Cache All are already in the workflow after generation.
poondoggle sorry, purgevram was just the node name, i think you can clean other things but Jib already answered that part of the question too, adding the clear cache.
Now that I think about it, I had this same exact problem when trying to run sage attention. If I restarted my system, I could get one gen out of it. If I tried to do another, it would crash my laptop. So I'm interested in if you find a solution.
TeeKay I used the workflow that @J1B used here: https://civitai.com/images/91371946 and it seems to have resolved the problem. The only issue is that the height and width values will be reversed if you are not making square images. I went from 65s/it on subsequent images to 7.19s/it now. HTH
How did you do this merge? I'm hoping to do the same for I2V into 1 model (or can you do a I2V merge)? Thank you!
Is this Phr00t software? I love 3089.
I use comfyUI with load model ,block merge, lora block loaders and save model nodes.
plastic of flux.
looks like you added flux plastic face problem to wan2.2? am i missing soemthing
I can make an image in 38 seconds (on my 3090) with this model instead of 253 second with the default WAN 2.2 model, that is the difference.
J1B lcm simple works best I have tested the new 1 u recommended and a bunch of others but no combo has surpassed lcm simple--test was with 4 step generation
sikasolutionsworldwide709 Thanks, I will have to test that.
I have just been head to head testing Res_3s Sampler with Bong_Tangent scheduler vs Euler/simple and Res_3s/Bong_Tangent gives way better details than that at least:
https://civitai.com/posts/20269313
J1B Also I have discovered in the last few hrs that the samplers working totally different on ggufs lcm simple gives a horrible result with Q8 gguf.
Update on that: lcm simple works fine for gguf too the issue was caused by the speedup loras I used before, mainly causVid
sikasolutionsworldwide709 oh that's interesting, so what's the best sampler for gguf?
J1B Isn't Res3 just RESTART - 3 TIMES, so if you're using "2 steps", res3 is really 6 and should take about the same amount of time?
brnlittokhoes311
Basically, This is what the github says
"Samplers that end in "s" use substeps between each step. One ending with "2s" has two stages per step, therefore costs two model calls per step (Euler costs one - model calls are what determine inference time). "3s" would take three model calls per step, and therefore take three times as long to run as Euler. However, the increase in accuracy can be very dramatic, especially when using noise (SDE sampling). The "res" family of samplers are particularly notable (they are effectively refinements of the dpmpp family, with new, higher order, much more accurate versions implemented here).
Samplers that end in "m" are "multistep" samplers, which instead of issuing new model calls for substeps, recycle previous steps as estimations for these substeps. They're less accurate, but all run at Euler speed (one model call per step). Sometimes this can be an advantage, as multistep samplers tend to converge more linearly toward a target image. This can be useful for img2img transformations, unsampling, or when using latent image guides."
They are slower but better.
J1B FUlly agree. And with some models they also make faces more unique as well. Using linear quad with res6s with some noobAI models can make some very realistic image and sometimes... interesting images. You have to play with it a bit but once you lock it in then it's amazing.
Will there be a WAN 2.2 - i2v?
hmm, I don't seem much point in adding a lot of style loras to a merge if you are going to start off with an image, but if you just wanted the speed loras merged in I could do some testing.
I wonder if we'll ever see this trainable as a txt to image model? Does fantastical subjects with incredible realism in a way nothing else really does. If NSFW stuff without too much studio lights could be added, could be incredible.
Yeah I have another release nearly ready, that tones down some of the AI look and adds better NSFW, should be uploading it later today or tomorrow.