You can try the smoothMix as highnoise and the regular wan2.2 checkpoint as lownoise to get better results, btw the movement will be impacted

Also try SmoothMixRealism in the positive prompt. The smoothmix animations lora was merged into this one and it will respond to it.

No black output

@qek share your setup then, rather than commenting like some entitled manlet

@kliz71 I have the latest ComfyUI + Kitchen, I have a nVidia GPU, and I do not have black outputs :)

Where?

also have this issue

me too

@lightandmild594
@jshnr006
What? What's the errors? Are you sure you do not use outdated ComfyUI. Also, I recommend to install Comfy Kitchen

I don't know what happened, but it worked fine when I used someone else's workflow.

same issue, did you solve it?

Any lightx2v/lighting loras you recommend?

try to update comfyUI, its fix same problem for me

Yes, I updated to the latest version and it worked, thanks.

How many steps and what CFG did you use without loras?

@kliz71 10 steps, 5-6 CFG for SVI at least - 6 steps works too but less fine detail. Low model I use lightx2v rank 256 at strength 1, with 4 steps (start step 4, end step 8), 1 CFG. Simple scheduler, shift 7 all around.

I think it's fp8 mixed

Yes, as you said, Comfyui ≥ 0.4.0 seems required; my output was also black. Since I only use the portable version, I always reinstall new versions from scratch.

@jorihalgo Yep. Not worth bricking already working installs. ;)

我也想要，还是内置加速的更好用。

I use a 50 series card. While there is definitely less motion overall than Smoothmix I2V v1, there should not be close to none. Try changing shift to 8, and using only 4+4 sampling steps.

If that does not work, try increasing high noise CFG to 2.0 and reducing the accelerated LoRA for high noise to 2.

Same here, maybe we need to make the weight a lot higher than before.

At least to my understanding reading the description, haven't had the time to try high weight Lora's yet though.

I've tested numerous cfg and least configurations. Still can't achieve satisfactory results. The resulting visuals are blurry and chaotic.

yep, all results are pure blur after 2 frames of animation. i use 1.5 cfg and 6 steps for each pass, high and low, for a total of 12 steps.

Same, there are no issues with other mods

same for me. came here to ask that too. maybe we need to use a ggfu loader instead of the diff model loader

no, gguf loader only loads gguf files. doesnt help

@bluenightlagoon I did some testing, adding a dreamer nsfw loras helped, plus, a prompt has to be really descriptive, so try asking an uncensored AI to formulate you one based on the image. But its still not as clean as illd like.

@KoujiAI Adding dr34ml4y introduces lots of transformation on anime styles with the new smoothmix version even at lower strengths, really not ideal. I have yet to find a way to not make v2 look like a blurry mess that doesn't transform the character :(

Edit: Using 2.1 lightning LoRAs at 3high and 1.5low seems to be the only way, too bad.

I'm asking for the version integrated with lightx2v_I2V_14B_480p_cfg_step_distill! Or is the original version more user-friendly?

You're probably missing the "Lightx2v Lora", it isn't merged so u must add it manually. Check out the model description.

You are using NAG so those negatives work right? Also could try changing the fps

If you're using CFG 1, negative prompts have practically zero effect. So as the post before said, either use NAG with CFG one, or raise the CFG.

@Yc3k what is NAG and raise the CFG to what?

@wannakm It depends on your setup but most of the mixes will have recommendations included in the notes for what CFG is required. Sometimes I'll bump CFG up from 1.0 to like 1.3 just to see if it helps it get past something.
Google says "NAG (Normalized Attention Guidance) is a technique used in ComfyUI to enable negative prompting for video models like Wan2.1, especially when running at high speeds."
Find a workflow that includes it to see how it is used with lightning mixes.

@wannakm Even tho the Ghost already replied let me try to explain it a bit more detailed. CFG is in essence a way to tell the model how much should it follow your prompts vs how much "creativity" is allowed. Higher the CFG number is the more the model will follow your prompts and less will it tend to be creative with the video. It works basically the same in picture generations as well as videos. Normally the CFG is set up to 3.5 or higher. Now, the problem arise when you are using the Lightx2v Lora. That lora enables WAN to make videos in very few steps, like 4-6 steps making the video generation much faster than it normally would be, but if you use that Lora the recommended setting for the CFG is 1.0 as it works the best with that CFG. Now, if you put your CFG setting at 1.0 "negative" prompts will have no effect on the generation of the video. They are basically ignored by the model completely. To overcome this problem, since negative prompts can be very important sometimes, there are 2 ways you can "include" negative prompts into your generation.

1. Raise your CFG number to at least 1.5. This will make the negative prompts work again, but in return the vide will take considerably more time to generate due to the high CFG.
2. Use the NAG nodes explained by the Ghost, which will enable you to keep the CFG at 1.0, making the generations fast but sill take the negative prompts into consideration and let them impact the end result.

You're probably missing the "Lightx2v Lora", it isn't merged so u must add it manually. Check out the model description.

@boulbi78 you are correct, Lightx2v fixed blurry output :-)

@boulbi78 Man, I can't believe I missed this step! Thank you so much—the videos come out so clear now.

thanks, this fixed my issue

how to use that? high pass? low pass? cfg1? steps 1?

Didn't make a difference for some of us unfortunately. Did you just use the lora model and weights DigitalPastel recommends?

@bluenightlagoon sorry it took some time. lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors into both high and low lora, strength 3 for high and 1.5 for low.

@crocusflowerparadigm lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors into both high and low lora, strength 3 for high and 1.5 for low.

it gets updated all the time so he'd have to update his too. Sometimes you don't want to run it on the high noise model, and there are all sorts of variants like 1130 and 1022 combination that I thought was the standard but now there is 1217 too. Best to leave it off, or use v1 if too lazy

yeah u need the light lora he recommends, otherwise it will always look like shit

high lightx str=3, low=1.5 i missed that first too

When using the V2 model, you have to add the lightx2v acceleration LoRA model yourself; otherwise, the result will be blurry. This is because this acceleration model was not merged into this version.

how exactly do you deal with svi2 pro slowing down? I'm having slowdown only on smoothmix I2V V2, while V1 works fine

@vamorand according to the official SVI page, reducing the lighting LORA strength can help. Apparently removing the Inspector Blend node from the official SVI workflow helps too. I am currently using AIstudynow's SVI2 Pro workflow with Smoothmix and it is currently fine (google AIstudynow SVI2 Pro workflow), works with other models as well.

Lastly for the male genitalia issue from my first post; I downloaded and used the Smooth's 'futanari and male genital' LORA and the issue is now not as extreme. There is some occasional sticking issues such as when the tip of the penis blends with the tongue in some i2v generations, but I've seen this in other models as well so I won't really count it against this model.

@McClippy , thanks for reply. I currently tweaked my workflow a bit, and now it seems fine for my purposes (removed shift from SD3 node, downloaded exact lightx2v model with 128 rank instead of 256)

Update your comfyUI then it should work.

0x368 where? I've tried different workflows and all of them the output is totally brurred =/

@rodszera Same here. I have the blurry problem

Reading better the instructions, it's indicated that you need in the high/low loras fields the corresponding light2xV. The link is provided in the instructions. That solve the blurry for me.

None, in fact, GGUF is mainly suitable for graphics cards of 30x or earlier generations.

@AI_Master_Workflow works fine with a 4080 super. No version of WAN gives problems as long as it fits.

这次训练过程中把Lightx2v Lora给去除了，所以生成的都是模糊的。如果要解决模糊问题只能自己手动加这个Lora，但是我测试下来也没有1代的效果好。

Yes it's good idea

@wealthyturkeyvvkj533 Thanks, if you can create one, that would be great. Because the ones I found here on civitai take 40 minutes to make a 5-second clip :(

@wealthyturkeyvvkj533 @ApexStorm_Ai Did you guys tried to open DigitalPastel profile? The owner of this smooth mix checkpoint? Hahaha come on guys, you're not even trying!

Seems he have a workflow for this ckpt: https://civitai.com/models/1847730/smooth-workflow-wan-22-i2vt2vfirst2last-framemmaudio , also look for more civitai sugestions, this creator has a lot of content!

@wealthyturkeyvvkj533 @ApexStorm_Ai If you are using it, please let me know if it is good for making expressions of faces in 16;9 portraits (SFW faces :)

Sorry, where did you see this? Could you possible share the workflow?

yo bro could you post the comparison

what about high noise? do you put the same one on both?

@whateverr in high use: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors

on low use: https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors

i have rtx 5060ti 16gb vram and 32gb ram 480 x 720 it took me 120 secs for 5 sec I2V

是的 我也注意到了

I found it changes to low quality from 2th frame

@1589673625848 yes, smoothmix degrades faster over time it seems.

@DigitalPastel so if I'm understanding this correctly, it merges all the loras under https://civitai.com/models/2040641 except for the newest "Futanaris and Males" so we don't need to load those loras anymore?

I mean who doesn't like them big ol' tiddies... :)

variety

@seaborgiumthemad I'm messing. I hadn't noticed this but then I don't think I've tested for it.

@NyxxiNyx To each their own - I like big ones if they are well-formed, but once we get into the 'huge' department they are a huge turn-off for me.
Same with badonkies.

Do you have the right files? The GGUF files are larger than these?

same i could only get the gguf files to work in any respect, wanted to use the safetens, and used the given work flow i only get blank/black out put as well,

@Fleshcrafter ok, you need to use the smooth xxxanimations high and lows in the loaders, or something like that, you can't pipe em through like older versions, i guess,

same still don't know how to fix this

The same

rapid "action", rapidly, swift "action", swiftly, fast "action", her breasts are bouncing on impact, her body shakes on impacts, rough "action", violently, violent "action", the shaft goes in and out on full length, the penis is inserted up to the base, rough sex, hard sex, looping motions

Use it in your prompts and it will improve motions

increases the value of cfg

@fdefreeza109 Well If you increase cfg your gen speed slows down like 1.5-1.7 times with not so big difference in output video. It is better to use WanNag which gives you opportunity to provide negative prompt without time loss and it works well.

well i think i had figured it out, it was wrong lora weights, had to decrease them like 2 time

No matter what I generate, everything is blurry, but it works fine when I switch back to the old version.

mine is blurry as well the first i2v worked well and i use the older workflow with a few things deleted like saving images

It works if you follow exactly what is explained in the guide about adding the lighting loras.

The lightx2v lora introduced here is for WAN2.1, so if the key does not match it will be displayed in the command list, but it actually works fine, so you can ignore the command list.

You need a lot more steps without the lightning loras 20 - 30 steps instead of 4 to 6.

@aifan Yeah, even when I use 30 steps, I'm getting really weird distortion. It might just me my setup (I'm using WanGP instead of Comfy)

Any style will shift towards SmoothMix realism/animation style (only face). If you generating videos where characters are very emotional (screaming, moaning or crying) their faces will change over time and will not be like in ref image coz this checkpoint has backed in style loras.

If your chars are calm and emotionless then the shift is very small and unnoticeable

Can confirm this works perfectly and simply, if you're coming from v1 of this checkpoint just add these 2 lighting loras and you'll be ready to go.

Wait... speed loras are not included in this one by default?

@CrystalVisage Yepp...not included in this version.

@Herastura but he recommends 4-6 steps? 4 for each, he means?

@CrystalVisage 4-6 in total. i recommend 6 steps. 4 is a bit low. I always use 6 and it looks good.

@Herastura bit difficult on 12gb vram 🤣🤣 Thanks for your reply.

@CrystalVisage Then you should try the GGUF version. It need less VRAM
Hash -> 5A4B1B9C62

This version is also without lightning Lora

@Herastura I mean I'm generally fine with 4 steps. Sometimes if I don't get what I want from 4 I do 6, but with 6 it's usually around 12-14 minutes or so. Too long.. 4 steps is 6-8 minutes. I'm now more curious about the best possible lightning lora. I'm trying different combos to find the ideal mix. wan 2.2 lightx2v 1022 seems like a good candidate (for 4 steps).

@Herastura what weight are you using those loras with? the combination op wrote works for me, but I tried other combinations and nothing works, always blurry

@Derpxten You replying to me? If so, I did try that. Still worse visuals than the model's previous version. I tried both at 1, and then high at 3 and low at 1.5 (what model's author had recommended but with other loras, not these) both had artifacts.

@Derpxten For the lightning lora i use strength/weight 1
For low- and high noise.

@Herastura Thanks! I just tested it and got pretty good results. There's also another comment that recommended using a low T2V lora that also worked well, not sure yet which of them was better but both are awesome.

update your comfyui

@Frank22 thank you friend

exactly, no matter what prompt you type the breasts are always the same

The same

把comfyui版本更新到最新版试试。

4.0 steps, euler a, 1.0 cfg, 8.00 shift. aside from that it could be anything loras too high, too many?

me too

Update comfy

"Lightx2v Lora is NOT merged this time so be sure to use pick any LoRA you prefer to accelerate generation as well as how much weight you use on them - all videos on the showcases used "lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16" set with weight 3.0 on High and 1.5 on Low."

@discreet001 Actually, not using the Lightning LoRA. It doesn't really make result in a black square video; it just makes the output blurry.

@g1263495582 I finally had a chance to test it, and I found better results with both high and low loras set to 1.0 strength.
If you are not using lightx2v, how many steps and what cfg are you using on your ksamplers?

when you scroll up. it's named "I2V v2.0 LOW"

See Herastura's comment from January 25th, it's working pretty well for me

I can confirm - recommended settings are absurd and generate noise blobs.

@yeahrightiwill confirmation seconded. i can't get anything but smeary noise blobs.

@yeahrightiwill You need to use lightning Loras. They are not integrated in this version.

@zombycowvonzombe7433 You need to use lightning Loras. They are not integrated in this version.

"Lightx2v Lora is NOT merged this time so be sure to use pick any LoRA you prefer to accelerate generation as well as how much weight you use on them - all videos on the showcases used "lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16" set with weight 3.0 on High and 1.5 on Low."

@Herastura ah, thank you.

@zombycowvonzombe7433 I finally had a chance to test it, and I found better results with both high and low loras set to 1.0 strength.

"Lightx2v Lora is NOT merged this time so be sure to use pick any LoRA you prefer to accelerate generation as well as how much weight you use on them - all videos on the showcases used "lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16" set with weight 3.0 on High and 1.5 on Low."

update comfy

@g1263495582 This is what fixed it for me. I had my ComfyUI version locked because updates were blowing up my workflows. Latest build is working, thanks!

@g1263495582 updating it worked! tysm.

all depends on what workflow you are using, what time of wanImagetoVideo sampler you are using.for remixWan2.2 i use painterI2V node. for smoothMix i am using regular wanImagetovideo node.

i have tested smoothMix with REMIXwan2.2
i think smooth mix had been able to keep the face consistent (if you are using SVI)
RemixWan2.2 is one which loses face identity if expressions weren't neutral throughout.

@NOTORIOUS_EDDY 뭐라는거냐 이새낀 다른 얘기를 하고있네.

preview image/video node.

@NOTORIOUS_EDDY you mean the node where i load the image?

if you're using the videohelpersuite nodes to save vids with vhscombine, you can connect the "filenames" output from that to a "prune outputs" node.

@bpbp2 I don't know what any of that is, I just want to eliminate clutter in the output folder by not having the input image unnecessarily saved. I only want the last frame and the mp4 saved

a better practice for testing any new model is
set following settings.
res: 400x600 OR 450x450
frames 25-29.

if the output is there, keep increasing resolution and frames. and see where you find it'd peak point. i have 16GB vram too i do this always.
make sure you also have a VAE tiled decoder. normal vae often is root cause for OOM for me.
good luck.

I use this checkpoint for I2V. By default, the workflow includes something like "stylized, artwork, painting, illustration" in the negative prompt. Remove that and replace it with something like "realistic, photograph". With the right approach, I've quickly solved these kinds of problems. Perhaps you could also achieve significantly better results with a style Lora like "Anime Style" by user RelativlyObjectv.

Yeah same, I can't find any workflow that keeps style consistent.

@Fuzetsu I configured it according to the author's parameters, but the facial deformation and limb distortion are still extremely absurd. Furthermore, I noticed that most users of V2 don't seem to have any significant advantages over V1. Therefore, I've abandoned V2 and switched my daily model back to V1.

@mmrabati2154 I can only continue using version V1.

i have been having issue with v2 also, i just post a new video showing v1 vs v2 and version one still does a lot better. is there any other creator that can chime in on what v2 actually improve?

@charlenebelmont I also end up with more stable and better outputs using the old version. I wont change it since it actually work properly.

@boulbi78  same here. i always thought the new one might be better since version one was already so good. but i will keep both version and keep testing more.

@charlenebelmont I think it is AI, we can't all use the same workflow, can't all get same results. Such as some softwares updates. Back up the more stable and stay on it. try updates, but don't force urself changing for something which doesn't work just because it is new. Good luck !

@boulbi78  talking about workflow. is there any t2v or i2v you recomend?

I2v v2.0 destroys penises with awfull quality render. What happened ?

true

That's 100% true, i tried other lightx2v Loras and only the one author wrote about works. I tried 1022 lightx2v and lightx2v 1030. 1022 seems to work, but results are still meh.. 1030 doesn't work at all

having same issue, yeap and I use normal Eualer, but the vids are always getting brighter

@evilstormhot109 Do you use the smoothmix animations loras with this model?

@Scringleblip I've started to use SmoothAnimation Loras only recently, like for a week or so. Previously I barely used andy Loras with SmoothMix v1

@evilstormhot109 try genning a video with and without the lora, see if it does the brightening happens with or without it

Same for me. Workflow gives me only blurry results.

Yep, even in higher resolutions output is bad

Yeah, I also noticed it and was wondering if it was only me, or the prompt.. But I started to see the blur effect everywhere on my vids

apparently this is indeeed the case, i wish i knew that as i've been trying EVERYTHING for two weeks now to get it running without blur (mostly noise) (especially around the eyes). what a nightmare and wasted time, downloading v1

Indeed. I'm using the old version. It works perfectly. This is blurry and horrible.

Even when I got it working, I was not satisfied with the results which were either weird or not as nearly as good as what I got with the original Smoothmix.

Yep, v1 is great, v2 is terrible. Not sure what's the point of this new version actually, and it's not very clear from the author's article either

everything should work, but i would alway use the lates stable release, never use the nightly unless you are testing, i had this issue earlier, check the requirement.txt make sure you Venv have the same package version for stability atleast. i also have 2 venv build. one fore a stable environment for a torch plust cuda wheel i use specifically for isolation of comfyui, and the other is for the latest torch and cuda build to test if updates is worth it and would it break anything.

workflow / settings は何ですか？

@DarkEngine2024 Lightning lora は使わないで.すでに含まれています

@DarkEngine2024 日本人でもないのになぜ日本語を話しているのですか？XD

@DarkEngine2024 I was wrong. v2.0 needs lightning for faster generations

you have to lower the cfg of the high noise model a bit

You may be able to prevent it with a small breast lora https://civitai.com/models/1983079?modelVersionId=2244856

Or try using NAG to negative large breasts maybe?

increase the lightx2v lora to 1.23 for high noise, if there isn't any hard core porn , for low noise you can use wan2.1 or full fp16. (official models are not trained on genitalia so everything else works just fine). low noise works as refiner

whiteout loras u need to change the CFG to 3.5 - Steps 25 | ksampler high noise start_at_step to 0 - 10 and for the second ksampler start_at_step 10 - 10000,

@n00b_TGT Thanks!

Read the description carefully! There is a workflow, it is basically plug and play. Just unzip, drag and drop the json file into your comfyui. Install any missing nodes with the comfyui manager. download all the models (high and low), vae, and clip files and put them in the relevant folders in your comfyui/models. You will also need the lightx2v lora (read the description carefully and follow all recommendations.) You should be able to generate videos like in the examples if you followed everything TO THE LETTER (don't deviate until you know what you're doing.) After that, start looking for loras to finetune your output. Smooth Collection (to the right of the example videos above) is a great place to start. Hope this helps, happy gooning!

Step1: Download the files (2 models, High and Low)
Step:2: Save them in "where is your ComfyUI? /models/diffusion_models/{here}

Step3: Download VAE and save it in 'vae' folder (https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/vae/wan_2.1_vae.safetensors )

Step4: download text encoder and save it in "text_encoders" folder ( https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors )

Step5: download lightX2V LoRAs and save them in "loras" folder (2 files:
- High noise: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors
- Low noise: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/resolve/main/split_files/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors )
Step 6: Load a wan2.2 workflow. Open CmonfyUI templates panel and search for "Wan2.2 14B Image to Video" and click (https://imgur.com/a/WHFcbqb )
Step 7: hit the "R" key in your keyboard and select the downloaded files for each field. (second image: https://imgur.com/a/WHFcbqb )

Experiment,
Good luck!

Here's the tutorial series I followed to get started ;) This goes above and beyond basic video generation, but I believe he walks you through downloading files to the appropriate folders etc. in a straightforward way. The whole series is incredibly useful. Just replace downloading the wan_2.2 diffusor models with the smooth_mix versions. The rest is the same; VAE, clip models etc.

ComfyUI Video Models: InfiniteTalk + Wan 2.2 + SCAIL + LTX-2 (Ep06)

Thank you, I looked at the tutorial, finally until I no longer understand anything, I drop it, no need to answer me, I deleted the model, thank you for your attention.

some people like to use whatever LoRA they want for this purpose. And they were using a wan 2.1 lightx2v at 3.0 strength for the merge. I dont use wan 2.1 lightx2v in wan 2.2 because it reintroduces issues from wan 2.1 back into wan 2.2 that got fixed, and how wan 2.2 converges is different fundamentally from wan 2.1. It works fine with Lightning LoRA in my experience. Leaving it up to personal choice instead of just baking it in and saying 'deal with it' is the right choice in my opinion.

1. There are many different lightx2v loras and this gives you control over which you want to use
2. You CAN generate without lightx2v, you just need a LOT more steps and time, but it actually should generate better results...

Params for most of my generations:

Shift: 5.0

High noise steps: 3

Low noise steps: 4

cfg: 1.0

Sampler: lcm

Scheduler: beta

Agreed. Smooth is better for motion but DaSiWa is better for details and detail preservation

Your link is already dead.

What are the weights for the two Light LoRAs?

@jesper123160 I don't think so? I only included 1 link and it is going to the DaSiWa Lightspeed model page on my end.

@KuKu12345 DaSiWa Lightspeed already has baked in LoRA so you do not need to apply any lightning LoRA when using this as the Low Noise model. SmoothMix v2 does not have baked in LoRA so you just use lightning as described in the model page ("lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors" at strength 3.0 for high noise)

@altoiddealer I assume then you are taking whatever your standard workflow is and you using smooth for high noise then and dasiwa for low noise? Any chance you can share the workflow?

@CallMeMaybe I would have preferred to just send a quick pastebin link but dear lord is it finicky about NSFW language. I've uploaded my workflow just now, here. https://civitai.com/models/2453238

@altoiddealer I didn't realize it scanned the content for nsfw language! Thanks for posting <3

Thanks a lot man... It indeed works very well & very fast !
Had not even dared to mix my gguf high noise with a not distilled low noise model.... (I m low on RAM - 12Go - & 16Go of Vram)
Strangely, it works well at 512x768 but it gives me very strange-out-of-control outputs for 640x960...

@BretChampagne I'm getting good results resolutions like you stated, and up to 1280 x 720

@Polkster oh boy... I tried so many mixes since I discoverd I could use this "Dasiwa" (great) model for low noise... Tonight I m playing with this great "enhanced camera prompt" model gguf4 for high noise & it works beautifully too...

@altoiddealer with CivitAi messed up these days, I just stopped posting and took the time to try to improve my workflow... finally managed to have a cool SVI one that suits my needs. I guess I won t go above 640x960. But this Dasiwa model is indeed a good pick for low noise fast refining.
SmoothMix is a very popular one, but I m not so sure I ll keep it as my 1rst choice for High noise...

@BretChampagne Glad to hear you've tested and verified what I've shared here, and a part of me was expecting my comment to be met with criticism, turn out to be some hallucination, or something. If you determine a more superior High Noise option or some other combo, would love to know - I'll be coming back to add to my comment if I find further improvement.

I would actually in many cases recommend using the vanilla Wan 2.2 checkpoint for the low noise. Why? It preserves faces the best and LoRAs actually work and don't destroy the quality.

I took a shot at this and it actually generated really good results, really good idea here. Trying to consider other "combos" of models I can try now :)

@aurelius Thanks for the piece of advice. You re right, it work very well too ;-)

@aurelius I tried using vanilla wan but it added a lot of weird fog and stuff, not really sure why, any ideas?

@Derpxten If you have a nvidia rtx, make sure to use the FP8 version (of course the Light one)... and don t forget the lighting lora

@BretChampagne I tried, weird results. Which LightX2V lora do you use with vanilla wan?

@Derpxten Any of them work, including ones meant for Wan 2.1. The results vary, also it matters how much strength you put. I don't know what's the best one, honestly. I use 1.0, but they also work with other values, maybe better in some cases.

And yes, your problem is almost certainly with lightx2v. The fog is what you get when you gen without it with CFG 1.0.

@aurelius hmm odd... I use a pretty complex workflow with 3 KSamplers + SVI, shift 8, and for some reason, only when I use vanilla wan 2.2 on the low sampler, I get those gooey / water splashes artifacts. If I use in the exact same workflow the Smoothmix low - it works. if I turn off lightx2v + use DaSiWa - it works. I've been messing with it for the past 2-3 hours and nothing seems to work well, some were slightly better but I still get weird artifacts

@Derpxten Try this: https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/I2V/Wan2_2-I2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors

@altoiddealer OMG! I’m really glad I came across this post! I checked out your profile, found the combo workflow, and decided to try it myself—and I can confirm it delivers much better results. You should seriously consider writing an article about it <3

I was getting super frustrated with Smoothmix constantly messing up some of the detials like nipples— they would come out blurry, distorted, or just plain ugly. But Dasiwa fixed it! In about 90–95% of generations, it preserves all the details, making the whole video look much cleaner with far fewer artifacts. Thanks so much for sharing this <3

PS: though IO am using SmoothMix v1 for high (I don't like v2 it has too much bugs) and Dasiwa v9 for low. This combo is perfect if you want to get the best out of two :)

@evilstormhot109 Glad my comment here was so well received :) I agree that SmoothMix v1 has better consistency in the high noise, but I personally preferr using v2 because it tends to be a bit more creative with movement/composition. If result comes out wonky, switching to v1 usually fixes it. Looking forward to see what V3 will be like if I2V model ever releases

HOLY $%#t

this works amazing! Especially when using the 480p lightning I2V meant for Wan2.1 on strength 3 in conjunction with smoothv1 or 2. The action and motion it adds is honestly, wow. But the standard lightning I2V WAN2.2 high on strength 1 does also fantastic work when you want more subtle movements.

The DaSiWa does a excellent job with presving the details. I've even tried with the PainterI2VAdcance node that adds more believable action and the results are great. Gonna have to test more model combos. Getting gold results is totally worth it.

Use something like https://civitai.com/models/1395313?modelVersionId=2235288 with it at like 0.5 weight, or any of the other DJ/deepthroat ones.

@NyxxiNyx Hey thanks! That helped! I'm still learning how all this stuff works.