I found the oldest version on huggingface, for me, that one seems to work better than all of the others. I could just be on crazy pills and the only thing DigitalPastel changed was the name though. Since they're the same size lol

doesn't T2V model work with I2V? I saw a lot of I2V works here done on T2V model

It's trained on a lot of SDXL images. So you'll see that face a lot. And the SDXL face is everywhere even in top end closed sourced models. That's what makes identifying AI generated people easy, most faces are some slight variation on that girl's face. Susan D. XL is what I've named her.

你不会使用人物LORA吗？

@sea5216 我想将这些背景故事用于其他目的，但大量的背景故事会造成损害。

Ever figure this out? my girls butts are all over the place lmao

@sploopin No, unfortunately, I haven't found anything that works.

I tried using the recommended speedup loras with weights 2h - 1.5l. this hepled a ton

you know, "unnaturally huge" at civitai is normal size :) But you can try the sliders that were released some month ago, they can make them smaller

I said the same thing but deleted my comment 😅. It's actually worse than v2. They're HUGE. If you try to go smaller, it looks like a man's chest and just weird. It's obvious the creator is a big breast man and has trained the model on such. I like big breast like everyone else but not ALL THE TIME. No love for us guys that like petite girls it seems. Got to mix it up every once in a while. I wish he would train on some other type of women than what's specifically to his taste since he's charging us money for it. (buzz.) I spent like 10 or 20 dollars on this model (forget what the conversion is) and I don't even use it. I'm still using v2. I just don't think the women v3 creates are worth the money. It kind has me wanting to learn to train my own models. Maybe he'll read our comments and train the next model with the taste of other men in mind 😅

@dft78750707 Just tried this after reading your comment just to test and it doesn't really work sadly. That lora is 6 months old and doesn't really hold up anymore. The chest are still very broad/ manly. The type of body frame meant for a girl with a huge rack. It just doesn't look right. Oh well. Thanks for the suggestion though

I have quite the same issue, but for me it's related to T2V. With smoothmixV2 no issue. With smoothmixV3 on the exact same workflow, black screen

Downgrade your comfyUI to stable version using the 'update_comfyui_stable' scripts

I'm not messing with my comfy installation, lol. There's no telling what'll break if I do, so I'm going to pass. I don't need to try every new model that comes around. Chances are this gets fixed with a v3.1 anyways.

@Tompte same here, I won't make my config un risk for a model. Let's wait next one

I've had that same issue on the stand alone version using pytoch (nightly) after using git pull. Doesn't happen in the app version. Try updating (or downgrading) pytorch.

having same issue, but for me V1 works, V2 gives black screen

Downgrade your comfyUI to stable version using the 'update_comfyui_stable' scripts

I have the black screen with the V3 T2V. No issues running any other Wan22 model.

@sorrelservices 

I’m on a Linux cloud sever not a Windows local install so I don't have the .bat scripts. Every time
I boot it pulls the latest custom nodes from GitHub making a manual rollback unstable.
I already matched the frontend to 1.41.21 and the backend to the March 19th commit but I'm still hitting c10::BFloat16 runtime errors and black frames.
Is there a specific pip version or dependency lock for Wan 2.2 that your stable script uses to fix the NaN math?

This happened to me after doing git pull. The update seemed to mess up comfy somehow. Hopefully they fix it soon. I've tried everything. Changing pytorch from nightly to stable. Downgrading python from .14 to .13. Reinstalling comfyui. No dice. The app seems to still work though just not the python local version.

test

Yes , optimizations and speed. Triton/ Sage Attention can speeds up the generation. Runtimes can offload model weights to CPU, RAM or NVMe so low‑VRAM GPUs (like 10 GB) plus lots of system RAM (64 GB or above) can run larger FP16/BF16 models, but slower than staying on GPU.
Model Patch Settings (fp16_accumulation):
FP16 is more prone to NaNs, so enable it if you use FP16 model and disable it if you use BF16, it is more stable and less likely to produce NaNs.

I think the baked-in lightx2v loras are not compatible with the new comfy loaders, If I'm not wrong it has to do with the new stable version of pytorch.

yes，I encountered the same issue yesterday. Only the smooth i2v v1 version produces an entirely black output, while all other models work properly. In the end, I chose to roll back the version of ComfyUI.

@ACancd i think it fine now i just try the v1 and it work like before, i guess they fix it

Black output on v2 for me, using Q6 2.2 normal wan files works normal though.. but I’m on forge. Maybe it doesn’t like safetensors

use f4c3spl4sh, so that they come out from the rigth place. No guarantees.

@CyberCream Is that a lora or prompt?

They also keep spitting and sweating and dripping fluids that I never asked them to and I'd like it to stop >.>

@Umbra It is a LoRa F4C3SPL4SH Cumshot - I2V Wan 2.2 Video LoRa - K3NK - Low noise - v1.0 | Wan Video LoRA | Civitai

@Umbra The first move would be putting words like "spit" or "wet" in the negative prompt, and randomizing generation seeds to find a seed that's less drippy. Maybe also adding some detail to the positive prompt so that "fluids" don't have a chance to fill in.

You could also try using a cum/fluids lora but give it negative weight if the prompt/seed can't keep it out, but negative weighting is more experimental for concept loras compared to sliders.

My third and least sure suggestion is changing the shift value. You can test it on one seed and see whether a flow shift of 2 vs 8 changes it. Sometimes the higher motion on a high shift 'creates' fluid. But lower shift favors detail, and fluid is a detail... The first two options are more straightforward.

@SIDK Thank you, I'll see if that helps. I'm so used to A111's SD randomising the seed every time I generate, I forget I have to manually do it on Comfy.

@Umbra It is a LoRa : F4C3SPL4SH Cumshot - I2V Wan 2.2 Video LoRa - K3NK - Low noise - v1.0 | Wan Video LoRA | Civitai

This seems to be a common problem across the board. The work around to it is to prompt out the peoples faces like saying "only visible from the chest down" . It shouldn't have to be like that but it's what we got.

I had to add 4-steps loras

On the Power Lora Loader (rgthree) you need to load two loras of,

lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16 (rank 32 also works for me),

one at 3.0 strength and another at 1.5

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.

I have the same issue. Have you fixed it?

@ws1gbg No, for now I give up and using another merge. The model worked well without accelerator loras, but I haven't tried using them since I prioritize quality with more steps.

@hectorium It works now.. U must add an speed lora "lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16" set with weight 3.0 on High and 1.5 on Low."

if u use lightx2v u need to adjust the ksamplers https://imgur.com/vCaxOOv

Because it limits you. lightx2v kills details and motion. for speed and fast movement, lowering lightx2v is a must. for example in the KJ workflow you can control CFG values more precisely: like in the first 2 seconds the model should absolutely follow your prompt, and in the last 3 seconds it can get creative, using lightx2v prevents that. In short, it’s bad idea to limit a powerful model like wan2.2 for the sake of a LoRA.

Just use comfyui bro.. so much better

play with the shift value, set it to 8. set lora strength to 0.65 if u use any. change the low noise to SmoothMix I2V V1, connect another pair of lightx2v for the high noise and low noise with the strength of 0.33 (for i2v v1 0.23) avoid using prompts more than 120 words. repeat words; (woman with closed eyes walking and the woman with closed eyes suddenly stop, at the end the woman with closed eyes, opens her eyes)

it is effective range. after 5 second model start hallucinating due to its token window limit (80 - 120 words). when u evaluate a model u should look at the text encoder ( UMT5_XXL ) not the model itself, the model can generate images based on what it receives form the text encoder.. you can extract last frame and use it as input image for the next video. alternatively u can use (WanFunInpaintToVideo) it gives the control similar to first frame, last frame without need to change base models, the difference is last frame will be in a better quality than extracting the regular last frame. SmoothMix I2V V1 (by far it's absolute beast compare to all other merges.) can generate up to 133 frames without hallucination it get even better if u use UMT5_XXL FP32, it let u use more complex prompt which allows u nest several actions in one or two sentences (25-50 words).

You can use SVI ,or using VACE to stitch FFLF clips.

it's not lightning issue. wan is 3 frame per step, last 3 frames shifts from decoded latent samples to your last frame image which is a photo. it's color profile is different form your video color profile. the last frame used for guide the model not to actually use it. u have 2 optiona: 1. color correct the whole video based on color and brightness of the last frame photo, 2. cut the last 3 frames.

change the low noise to v2 to official one or use character lora with small breasts (Sierra wan2.2) , or use I2V model but provide 100% black or white image and connect t2v lightx2v loras instead of i2v lightx2v loras then prompt it as t2v (detailed)

you probably need lightx loras for it

Indeed read the description for more info. All you need is there for you to find.

i use wan2.1 vae fp32.safetensors

@himanshu991123744 
Ty i too use wan2.1 vae fp32.safetensors, but but an error but an error occurs during generation in the console
[ERROR] Exception in callback ProactorBasePipeTransport.call_connection_lost(None)

handle: <Handle ProactorBasePipeTransport.call_connection_lost(None)>

Traceback (most recent call last):

File "asyncio\events.py", line 88, in _run

File "asyncio\proactor_events.py", line 165, in callconnection_lost

ConnectionResetError: [WinError 10054]

@RomeoAI https://www.reddit.com/r/comfyui/comments/1i17jk3/exception_in_callback_proactorbasepipetransport/?tl=hi-latn

(check this out). maybe it will help

@himanshu991123744 
Dude, you're the best. Thanks for the recommendations, they helped. But the generation still turns out to be a mess. Can you tell me what settings to use?

@RomeoAI hmmm i use like 6 steps, euler simple, 8 shift.

maybe try adding something like " a realistic girl" or maybe try the smooth mix animations lora. if blurry then i guess u can just experiment with another sampler or reducing shift or trying both normal vae vs vae fp32.

(these are for t2v smoothmix).

(sorry for the late reply btw)

@himanshu991123744 
It's very interesting, I did 20 and 40 steps and it always turned out to be a mess. It seems to me that I have tried all possible variations, but the result is the same(

@RomeoAI why use so many steps?

maybe try like 4,6,8 steps. check if the model have lightx baked in, if yes then just use like 4-6 steps, otherwise use lightx lora and 4-6 steps

DO U MEAN THE KSAMPLER WITH NAG NODE?

2.1 is what you're supposed to use. 2.2VAE is for the 5B model iirc.

Can't help with pixelation issue. Go over your workflow and make sure you've got the correct loras and cfg settings for the model you're using.

yes you're right

@MrSjeikz thanks for the reply

Yeah. I still does not find better HIGH model. The creator did really good work. Wish he come back with the v3 version.

The same lora must be loaded for both the low noise and the high noise model. So in both instances of Power Lora Loader you must select "lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16". Then you can set the weights according to the instructions.

@MrSjeikz thank you very much for your help!

Isn't it the case that Wan2.1 I2V loras do not work with Wan2.2 I2V checkpoints? (i.e. lora key not loaded errors) Only Wan2.1 T2V loras can work with them. Please correct me if I'm wrong.

Getting black screen for T2V , V3 :(

Same problem for I2V V2 here. All lightx2v loras generate errors. I've tried lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16, lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16 and wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022

you're talking about Bernini-R right? and yes the official comfyui workflow for bernini is kinda wrong, just use official wan2.2 replace image2video node with this node "Bernini Conditioning" for multi image references use "Image Batch Multi" from ComfyUI-kjnodes , if u see ghosting or extra characters, switch to I2V lightx2v lora instead of t2v. u can also completely ignore the lightx2v if u use dpmpp_2m_sde_heun with sgm_uniform , 8 steps, it is faster too (assuming u r using the official wan2.2 workflow)

@nafawex632141 Would you mind to elaborate the dpmpp_2m_sde_heun sgm_uniform combo? (i.e. workflow) Thanks a lot. High sampling of Bernini-R takes forever whenever extra loras are added or there's large movement in the video.

It means to use the same lora file for both high and low acceleration. Cheers.

@cloudreadypc sure here is the workflow https://pastebin.com/t32LhTSq

Good question!, SmoothMix I2V v1 is an absolute beast, no other merge comes close. I’ve tried every single one. It excels at prompt understanding which eliminating the need for other LoRAs, maintains scene consistency because it is not merged with other people's LoRAs (it is only merged with SmoothMix animation and style) , and supports long videos over seven seconds without hallucinations (most of other merges are using wan2.2 t2i loras and motion suffers).

IMHO, DaSiwa has better details, SmoothMix has better dynamics.

1代正常的 但还要配NSFW LORA 2代应该是没集成LIGHTX2V 要自己添加

for hq (720x1280) text-to-image use the same cfg and steps you use for flux or sdxl (Euler/Simple - CFG 8, steps 20), only thing u need to change is start_at_step and end_at_step as follow (disable lightx2v loras if u use any):
High noise:
- start_at_step: 0
- end_at_step: 10
Low noise:
- start_at_step: 10
end_at_step: 10000