@FX_FeiHou At around 800px resolution, when the head takes up only a small portion of the frame, is shown in profile, or is partially obscured, the face tends to gradually become photorealistic during animation generation. The original wan model can also show this behavior, but much less frequently. Don’t use 2.5D images — try light-novel–style illustrations and the issue becomes clear.

@FX_FeiHou you could include at least one example in english - especially because you use very specific prompt structures. What am I supposed to do if I don't know chinese? I tried translating with Google, but the model doesn't seem to like it.

Yes, the negatives in examples are in chinese...but negative prompts don't matter with CFG 1. Plus, even WanVideo team gave plenty of examples in english (yes, with that negative in chinese).

@FX_FeiHou I'm not clicking that link, mate, this is weird. If you have examples in english, post them on Civitai or huggingface, not on some sketchy website.

@dobomex761604 you can translate those tag to english, it will work fine

Actually you can use simple English. Simple is beautiful. the problem is not the translation because more words in the prompt is meaning more confusion. Most people use AI to make longer prompt, but Keep in mind it is only video with 5 sec. long prompt help create picture, but for video creation, it need simple and direct instruct.

@iluvlamia nope, tried it, it doesn't work well.

@takomli2013979 Yes, the model works with normal prompts, short and long, but the developer gives good examples with specific prompt structure, and it doesn't work in english. They should include at least one example in english.

Interestingly, they still haven't done it - which means that their prompts don't work in english. Shame!

@dobomex761604 It works perfectly fine with English in the same structure as OP provided, you are just too aggressive mate, no one owes you shit

Nearly all T2V loras will change face, so only use this model to render Anime style or increase your own character lora weight to offset effects.

I've ran a few I2V with some unique faces so far and it seems like it retains the face. I could be wrong but that's what it seems like after 8 jobs ran.

@FX_FeiHou Thank you for your response! While using your workflow from google drive for I2V would you be able to help me with using additional "external" Lora's? How would it be set up? What node do I add and where in the whole chain? Thank you so much in advance. So far it's very fun and yields awesome results!

Also would like to know how to work with

Prompt reverse -Qwen3

as we cant edit the text( written in chinese) in that input field for describing the video using your prompt guide

it generates prompt based on your input image and then creates scenario prompt. I do use node called "Prompt Verify", that I insert between the output of QWEN and Clip Enoder(where you manually write prompt if u were not using QWEN-VL) that alows me to edit text before resuming WF and sending it for process.

If you have enough RAM

No

@Sassyphrassy No. Yes (you can with 6 VRAM)

No, each model is like 13gb and to fully use this you need both low and high, even my 24gb isn't enough

Is it so hard? :/

Try to

If you can't, you can play with ComfyUI's launch arguments or use quantized models

Doubtful unless you're doing really small/short videos. Using ComfyUI's built in T2V 14B workflow, with a 960x720 video size and the default length of 81, Comfy is using 8.5Gb of VRAM (RTX 4070ti) and 50Gb of system RAM.

I'm running the Q3_K_M GGUF on a 3060ti 8GB VRAM with 32GB RAM.

@FX_FeiHou using local, maybe thats what it is then

@FX_FeiHou How to do it? and what does this mean?

You can work around that with more system RAM. I'm using Comfy's default 14B workflow on a 4070ti (12Gb) and at 960x720 with the default length of 81, it's working using 8.5Gb of VRAM and 50Gb of system ram.

@Lonmine disable "block swapping" you are good to go with 3090

i have a 3090, 128gb of ram, (as a 3D designer, you cant have enough RAM!!!) dont need block swap yet with doing vids of 853x480, but I guess when testing 720P will need block swap set to active (turned off right now).

853x480 / 4steps on each sampler, 97 frames = 5 to 6 minutes of render time, and for the quality it gives... WAUW!!! Textures have such realism.

PS. Don't forget to add sageattn (between the "model loader" and the "blockvideoblockswap" it will speed things up drastically.

The latest is just an img2vid MoE model

Maybe WAN may struggle with making such videos with length>81?

I happend me too! (121 frames)

@FX_FeiHou like @qek and @sepesusi said it only happens if you generate a video longer than 81 secs. After typing this I switched back to 81 frames and the generations stopped looping back to its First frame. If it's possible, can you also provide a new workflow with all the loras included in the remix as a seperate box so people can manually turn them on and off? This could allow the user to have more control. I personally use a Power Lora Loader for extra Loras I'd like to use (like x-ray lora etc.). Good work on the remix once again.

Isn't it img2vid?

@qek It is, but still. After a few frames I get a very "pronounced cameltoe" that subsequently ends up making the labia protrude outside the underwear. It might just be the quantized model I'm using. Some prompting tweaking helps a little.

Yes

Can even try larger res, I was able to make 1000p videos without an upscale

Can do a high quality vid at 720p, with a 4080 super, only on 32GB RAM. Takes over 9 minutes by my chosen sampler/scheduler thrives on more steps than the usual workflow, and I don't use the garbage 1.0 CFG that everyone recommends.

@rivdemon1221554 do you have an workflow?

@rivdemon1221554 give yourt WF or it didnt happen :P

@blhll I just use the standard comfyui template. There's literally nothing wrong with it, like how hard is it to believe LMAO. It happens all the time whether you like it or not <3

Is your prompt enough detailed? I know that such models don't always give consistent results

@qek i mean it is Img2Video so the image I uploaded is that of an asian model but it would slowly morph into another character within the 1st 3 seconds. didn't have this issues with other model or if i use a european model image. I tired putting Asian face just now and it at least stay the same race but it didn't look like the reference image

extremely good, much more detailed and smooth in my experience so far, ive been using this model and smoothmixwan equally as much lately

There's already Rapid Mega for that, I like it, no need for separate models like t2v, i2v, Fun Control, Fun Inp and more

@qek Rapid Mega is quite poor at maintaining character consistency.

Can you share your workflow?

@qek im not asking for what there already is, im asking for the combo of this one and fun or vace control....that keeps good character consistency, which Remix does and Rapid, doesnt

@JMC3DX Yes, the creator should try to make it, there is a good reason to have VACE instead of separate models. I'd never make t2v and i2v merges instead of 1 VACE model because it's better than Fun Inpainting

Try the SFW version

I agree, characters like to squat comically instead of doing their stuff, especially often and more strongly in standing poses.

you will have to adjust the prompt. it is very senstive and i recomend using json format to writ out what is happen every second. That is how i got my videos done so far and posted here.

@Kary0819 Unfortunately, it lacks sensitivity to male body types, even though it renders male bodies normally when generating side-view poses. However, in rear-view poses that display the male body from the front, its shortcomings become glaringly obvious: male body types are limited to a single style, resulting in disproportionate figures.

curious are the faces that are changing for you non-European?

@Kary0819 haven't tried many faces, but one them was my friend, u can somewhat, call him asian. but the other face that i've tried, was an image which i've generated before and she wasn't asian. the problem is, the faces, changes so much, that it won't be even close to what the original face was.

Use the lora with negative strength

It's just image2video

@FX_FeiHou  I used i2v 2.0. And the model is wonderful, thank you for your work.

@FX_FeiHou  i use https://civitai.com/models/1847730?modelVersionId=2264611 nothing change just replace smooth checkpoint to remix.

要啥自行车？先用着！

如果参数正确那么这种倾向不会很严重，但是问题确实存在，不过这个模型的动态实在太优秀了，等更新了

用原版low noise