@The_Cook great, currently testing with my SVI wf and try to use your promt to see if it works in batches.

btw it works great with distilled 1022 lightx2v lora on low...

@The_Cook I CANT SEE THE WORKFLOW! :(

@CharlieBrown0115 Link works fine for me after providing buzz to the legend who made this

@The_Cook spectacular results, the higher u can go with high res the more stunning it looks as long as u pick the right photos and properly crop as explained in wf. Some issues with how genitals looks, I think I could try wan2.1 female genitals helper, it would be perfect if the innies of that lora could be reproduced in these videos, will try later, it took around 30 minutes 1536 x 1024 to finish with sage attention and fp accumulation on 5060 ti 16gb. trying higher res I get out of vram with Q4 KM model.

I wasn't getting RES4LYF generating anything on the 2nd stage and just hanging until I defined these WxH values in the Upscale nodes, thanks for mentioning this! It was either this or lowering the High Rez Pass WxH +(AND) connecting to those undescribed WxH values that I got past the hangup and ClownSharK sampler started actually working haha

@The_Cook I guest I got it now. Additionally to the starting frame, I provide the guidance for her face and full body

High + Low = Poor quality
https://civitai.com/posts/25959349

Only High = Good quality
https://civitai.com/posts/25959403

@AishaAI Would you post your workflow? The original poster's link to his workflow isn't available for me and your second result looks amazing!

I'm pretty sure the low noise contains information about what vaginas and nipples look like, etc. But if you're using with other genital loras or a custom merge of Wan 2.2 I guess you don't need it.

If i only use the high noise, the face completely changes.

I use the ultimate anus and pussy lora for the low one and face consistency is acceptable and the genitals are WAY better then with the original low noise lora.

I've added fp16 accu and sageattention to wf and it works just fine no blurriness...

@erdelman73267 I don't run into this issue in other native wan2.2 workflows, but it just doesn't work with this one created by the author. I'm using an RTX 5090, with torch 2.7-cp310-cp310 and CUDA 12.8.ComfyUI 0.9.2. It might be an environment-related problem.

@white1147 I am on cuda 13 on freshly installed UmeAirt comfy https://huggingface.co/UmeAiRT/ComfyUI-Auto_installer

I hope he has. -_-

@The_Cook Thanks for the tips !
Maybe I should just try your recommended workflow. You ve warned us... Sorry If my comment looked like a complain, it wasn t ;-)
Was reading a comment here about multiple images input, seems very interesting too...

@The_Cook just finished a 3rd render... I tried to mix with a slop more natural bouncing lora, but it s still very "NSFW all in one" amusing bouncing...
my prompt about it was : "her large breasts sway naturally. she s posing nude, teasing the viewer." > nope... did not work as expected...
Still don t know If too much "bouncing" is that bad XD

LTX is terrible. Doesn't adhere to prompts at all.

@LetTheBassDrop neither dose wan withought the right loras

what's the time difference ratio and what your vram?

@olivereads38255 omg dont even get me started. i can do a 10 second 1920x1088 video on ltx in 3 min. wan is a good 7-8 min for a 720x512 video. i got 24 gb of vram

@Ragamuffin20 damnn. how's the prompt adherence? do you need about 2x the attempts vs wan or less??

@olivereads38255 I mean honestly everyone's experience is different and i haven't run any professional tests with it but from what I've used it for I've come to discover to fairly accurate. as much as wan or more provided you give it an appropriate amount of time for your prompt. like you don't want to prompt something that will take more time play out that how much time you set your generation for. you can also use your own audio which is a HUGE help in that aspect. the REAL problem is though that its heavily censored. people are making loras and text encoders that bypass them perfectly but withought the loras you cant really generate anything NSFW and my god do they take long to train. i have on right now thats been in runpod for the past 2 days.

@Ragamuffin20 ahh, i see. some decent tradeoffs. how long did a wan lora take for you in comparison?

@olivereads38255 20 min

@Ragamuffin20 sorry, i meant for training

@olivereads38255 i actually never train for wan but i did train for z-image and that only took 3 hours for 5000 steps. this one was like 3000 steps

@Ragamuffin20 i see, whelp looks like i have a lot to experiment with this weekend! thank muffin

@olivereads38255 gl man and it if helps AI toolkit https://github.com/ostris/ai-toolkit has wan2.2 and ltx-2 which makes training 1000x easier and Aitrepreneur https://www.patreon.com/c/aitrepreneur/posts has a batch file for an easy runpod set up. its only like $5 to get in i think

@Ragamuffin20 thanks man!

how to turn on combine videos? I enable nodes but they dont start

@The_Cook It doesn't always delete the first frame, even with the 1.1 WF.

Having the same issue here. The preview image is sometimes flashing in the beginning of the other phases. seems a bit random.

Yeah, it's still happening to me for every generation. Not sure why.

I fixed my issue. If you use the FL Image Batch Node to combine the frames, you need to have the Select Every Nth Image node strip the first frame before sending it to the FL Image Batch Node. So for instance, its: Select Every Nth Image Node > FL Image Batch Node (input image_1) then do the same for (input image 2), etc.

@TheAIAddict I've actually always had this bypassed at the end because I saw too many say it caused OOM errors, so I just combine the 4 separate outputs with ffmpeg later. But I do have Select Every Nth Image node active between the Video Combine node and the FL RIFE Interpolation nodes. These came this way by default and haven't been moved or turned off.

yeah, that really sucks.

???

is there a reference

What is happening? please link info

So, after some tests, I think the main problem of OOM error is comfy memory leaks (or, it just offloads everything from vram to ram, cleans vram, but not cleans ram, and finally gets OOM at some VAE decode, like I got recently at 2nd pass at 4th clip - almost at the end -_-).

I divided authors wf into three separate wf's (1st pass, 2nd pass and final concatenating into one video) and it worked better then one wf (at least, now in 2nd pass there's no loaded at ram high noise model). So, final by now resolutions are 768x512 (that's +- max I can afford for 257 frames on my sys) for 1st pass, 1344x896 for 2nd pass (tried 1536 x 1024 on whole wf, but got OOM on that last 4th clip at VAE decode stage, not tried with separate wf's because...)

Because it takes about an hour... ^_^ I did some simple wfs for prompts like "She turns herself straight to 90 degrees and stands still." or "She turns straight to camera and stands still." for preparing side view or front face view pics if there is no such picture (ofcourse most of a time there is no such picture). Some of them run then through SeedVR2 to add some details (anyways the quality will be doomed when concatenate to one 1344x896 picture, but better than nothing). Also, I use for side body shape view authors previous lora "posing v2", because from what I saw, it somewhy do better shapes for "shape" picture, than if it does this lora if "shape" are not properly... indicated.

No, it's still about 40 minutes.
Also, find out that for face consistency sometimes helpful to do just 1pass render, than do faceswap with reactor of whole clip, than run in through SeedVR2 for upscaling to 1080p - faster for me, better face at the end, but... reactor isn't perfect, if you tried it before - you know weak spots (sometimes can't swap face at some angles, and when half of face at the frame), but works mostly good as the face almost every time at the frame and facing at working angles. Also, tried to use 2pass after reactor, but 0.65 denoising ruins all the work, something like .12 might keep produced with reactor face and blend it deep into clip judging on previous experiments, but not tried that here.

in image_concat_multi node set match_input_size to true

@The_Cook Do you split the Frames with Premiere or a video editor or am I missing something ?

I have found it is best to crop the images outside at 1:1 2:3 and 5:12 then they never fail. The only thing i can't figure out is what are the crop markers. Is it to the knees or to the finger tips for 5:12? Is it to the waist or to the gap for 2:3? how close is a closeup face shot, just neck and head, to the shoulders, or cut to the chin with no neck?

26 minutes on a 5090 for me, but you get a preview in just a few minutes, before it does all its refining.

about 1.5 hours for me on 4060ti 16GB

That's at 896 x 1344

Preview in 6mins or so

@dzuodn9i303 possible to let me know how much RAM you having? i have 32gb ram with 5060ti 16gb vram, every time it finish render 1 part of 896x1344 will auto crash my comfy, so i need to restart comfy 4 times to get all the 4 parts :D

@erioca 64GB DDR5 it gets pretty full

@dzuodn9i303 ohhh....thats double of mine hahahah, no wonder mine died after 1 video

Just over 9 minutes for me, but thats because I'm only doing 2nd phase at 896x1334 and portrait mode.
Using H200 SXM and 234gb of ram

did you find any fix ?

It's happening to me too. Does anyone have any ideas?

You can turn it into hardcore.

@LetTheBassDrop Do you have any tips? I tried using some Blink LORAS, but I haven't had any success. Could you share the prompt and Loras you used?

@MaxtorCivitai using the supplied workflow would be trivial, simply provide it the hardcore source image and only add the loras the that particular sample step. Doing hardcore and the smooth transitions of this lora won't work as well or at all -- as you claim. I'm sure some significant prompt-work could help, but sampling more than once is the real solution

That's if you're using the supplied workflow for max quality. If you don't want to do all that then just use it as a normal lora and prompt for scene changes. Personally, I already use a hard cut lora but still use the low noise of this lora as it maintains facial consistency a LOT better between cuts.

My vae encode and decode start to slow down periodically with no reason, even though I still have enough memory. I replaced them with the tiled versions, and it seems to have helped.

yes, my suggestion : ( this is what i did ) 1. use GGUF Q4 models, 2. use block swap ( i did 40 ) 3. use context options , 4. use tiled vae decode - 4.a use tiled encode for high rez pass, 5. i use 2 GPUs - a rtx4060ti 16gb as primary and a 12gb 3060 as secondary - i used multi gpu nodes to load the clip into secondary gpu ( this frees up 6GB ov vram on primary ) 56. finally use a purge memory node between high noise aand low noise pass - i usually use 3 steps for high noise and 5 steps for low noise - --- the work flow gave me excellent results - but with my 2 GPUs and a 64GB RAM - it took exactly 1 hour 21 minutes to generate a 257 frame video in 1536 x 1024 size

I personally wouldn't try to repeat such a feat a second time. Almost an hour and a half is too much for one video.

SVI has been a disappointment for me until now(3090 24gb 64gb ram), so this one suits me much better. Can you share your workflow so we can try it? Thanks!

i get IndexError: index 3 is out of bounds for dimension 0 with size 3 on batch 2 batch one is working any help ? i will try the model above ( 9070xt 16 giga vram / 32 giga ram )

Wake up, brother--for he has returned!