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
Initial release, it can go as low as 8 steps.
I primarily use it for text2image but obviously can make video as well.
FAQ
Comments (17)
Beautiful !
Are there keywords for integrated loras ?
Do you plan to make a smaller version ?
I don't really "believe" in the need for Keywords, I just merge loras at a strengh that doesn't require them. but words like "grainy" "amateur iPhone photo" etc should activate them more.
The fp8 version is uploaded now: https://civitai.com/api/download/models/2052733?type=Model&format=SafeTensor&size=pruned&fp=fp8
J1B Thx !
Thanks very much for this, does it include the lightx2v distill lora already?
Yes, at a low weight of 0.25 I think it is punching up the brightness of the images a bit, I might tweak it down in a future version.
Could you make an fp8 version?
Yeah I think so, I was getting some errors while trying to save as fp8 but I should be able to work something out.
J1B Looking forward to your FP8 version
The fp8 version is uploaded now: https://civitai.com/api/download/models/2052733?type=Model&format=SafeTensor&size=pruned&fp=fp8
It seems to need a few more steps than the fp16 version to look good for some reason (about 10-12), but I could not seem to fix that.
J1B is it possible to also upload the fp8 version of the wan 2.2 merge
I only have 12 gb of vram, so it takes some time to generate an image which I did with with 10 steps using the fp8. I compared it to the base model, and your model is very good. I think it's the best I have seen yet.
You mention that you mainly use it for text2image. Do you have a favorite workflow for image generation with this wan model that you can share or refer to?
You can use mine if you want https://pastebin.com/m9F7qZsE
Hi, the workflow is attached in the sample images, (it's not one of mine so I didn't want to post it separately) you can basically use and video workflow and set the output frames to 1.
J1B In case this is of interest to you, very recently, an interesting FLUX + Wan workflow for realistic image generation was published:
https://civitai.com/models/1808412/wan-refine-fluxmania?modelVersionId=2046518
dispose608 I will test it out, but I think I still prefer the look of my Flux model with an upscale: https://civitai.com/images/90063918
I might try swapping it around and using WAN with my Jib Mix Flux model as a 2nd stage refiner to get the skin better than WAN can do.
J1B btw, to this date, in terms of realism, I get the best results with your Flux model ; if there is any room for improvement at this point, it possibly is a workflow combining your checkpoint with wan