🗡️💀 DaSiWa-WAN 2.2 I2V 14B Lightspeed | FP8 Safetensors💀🗡️

Thank you :)

I do not release the fp16 versions. All fp8 from fp16 base.

@darksidewalker Why not release the fp16?

Luck! - color match and images with same composition :)

I give 2 start images, 2 frames from the previous video segment. Doesn't have to be the last frames, works even better if you truncate the previous video mid-action, to continue motion

@catsama Thanks for the reply! I'll give it a shot

@Kazuki_kun just found this, did you try? https://civitai.com/models/2024299/wan-vace-clip-joiner-smooth-ai-video-transitions-for-wan-ltx-2-hunyuan-and-any-other-video-source

@catsama I will try soon, thanks!

Nice to be back! I wish you the best awesome results! And fun!

Can you share your settings? I'm trying Boundbite as well, and there seems to be a lot less motion compared to Synth.

@BenkyouDekinAI Hello, i'm just using higher cfg 1.5 instead of 1 and using loras to help him giving motions. His power is actually the fact that he's "softer" which can make it easier to do soft motion, for something mor lewd such as the clips i shared, u might need loras and play with the strengths. i shared loras/prompts on the clips. for the sampler im using euler/normal 8steps and cfg 1-2

I'm counting the days as well xD

Don't know, never had this happen, are you sure this is not caused by a Lora? Besides that BoundBite v10 is more consistent to details.

@darksidewalker Thanks for letting me know it was user error. I got more specific in my prompt and it fixed it! Amateur move :(

When you mean 7 steps are you doing something like 4 high 3 low or some other combination?

@serpentine18 exacty!

Well that can not be true, since after 3 seconds the video is almost over XD I assume you doing something crazy or trying to do things it can not do O.o

Maybe you mixed the checkpoints up.

Share your prompt.

I was facing the exact same problem. I can't wait for the day when access is unlocked.

You can read up inside the announcement, it can produce whatever you insert as i2v.

Is this person here to stir things up, or did you just not read the other comments?

I have no clue what this should be about, maybe I don't get the reference

@darksidewalker It means that using your models is addictive, whether it's the Wan series or the Illustrious series. The generated pornographic videos/images are so stunning that it makes you want to masturbate,if you do it too frequently, you'll suffer kidney failure (of course, this is an exaggeration). So, people with poor self-control should use it with caution (I will strongly recommend your series of models to my good buddies). I'm not trying to stir up trouble, hahaha.

@stom72569752 Ah okay, thank you for the explanation! Maybe that's why I'm so drained lately :P

https://civarchive.com/models/2217653?modelVersionId=2496754 | https://civarchive.com/models/2217653?modelVersionId=2496721 might help

@g1263495582 i've tried using this before but it fucks up my gens for some reason, especially anime ones, i've been using dreamlay as a workaround but genitals always look half-baked, almost like forced, idfk what im doing so wrong, everyone's gens look amazing, with every bend and fold and texture so clearly defined, they all look so nicely detailed. beats me man.

@redlucario1735 Weird, mine works fine with no issues. What LoRAs are you using when you gen? Both high/low weight ones. Also if you're comfortable, send over your results + prompt (upload to Jumpshare/etc). And what text encoder are you using?

@g1263495582 https://www.redgifs.com/watch/obviouspinktarsier - https://www.redgifs.com/watch/understatedpropertigershark

These are two examples of this issue happening, although it seems to happen also when im using no loras whatsoever

First one on the left - ONLY smoothXXXanims low noise @ 1.0

second one - smoothXXXanims high+low noise @ 1.0 AND doggy high anal cut high+low @ 0.3

4 steps, euler/simple Dasiwa v9 Q6

nsfw_wan_umt5-xxl_fp8_scaled

Wan2_1_VAE_bf16

Prompt: "A high quality anime 2d video, SmoothMixAnime. A woman with long dark blue hair and cat ears sits on a red bed, wearing only a thin red thong and a black lace mask covering her face | Jumpcut to a new scene seen from the side diagonally where the same woman is on her knees on a bed, her face and chest pressed flat against the bed, head turned sideways facing the camera with a happy aroused look, Her bare feet are spread wide apart across the bed flat with soles looking up, thighs vertical with hips lifted high into an exaggerated arch that elevates her buttocks prominently upwards. A man with his back turned to the camera stands on the bed behind her fucking her asshole in a deep squat position, his legs bent and spread very wide out of frame; he leans forward slightly while gripping her glutes as he thrusts vertically downward from above with full length of his large penis inserted into her anus. His large erect penis is fully embedded within her stretched anal opening — balls-deep at the bottom of each powerful motion, which causes visible skin stretching and glistening fluids around both bodies. The camera angle is low, capturing the scene from the side at a 45 degrees angle. her face and breasts remain visible without obstruction."

@redlucario1735 Add the pussy asshole LoRA at a low model strength of around 0.5–0.8.

[Edit] Also, the other thing — you can ignore it if you want. Have you tried changing the word 'asshole' to 'anus' yet?

I have some questions, how many distillations lora did you mix, and could you share the names? If not, that's totally fine.

I use the LightX2Vid distillation loras in a mixture that I invented myself, but I would like to keep the mix and weights a secret. There are so many people who just steal the work of others.

@darksidewalker I feel that V10 has 2 issues I've encountered:

The model tends to be somewhat like slow motion. If prompt with a good description of the motion, it can be fixed to some degree.

The second issue — honestly, you can ignore this one — is that the Low model V10 feels like the nipple detail is not as good as V9 (in cases where the nipple is not visible in the image)."

What aspects of performance do you think should be compared against LTX-2.3?

@g1263495582 
It has sound. I tried it and it's fantastic. You can see my work in my profile. I just need a good LTX 2.3 model (NSFW).

@Renessance OK

Thank you! Enjoy :)

No clue, I don't use WanGP, maybe you should ask the WanGP creators?
I normally put it under "unet" not "checkpoints".

@darksidewalker Thank you. I'll check it again

Did you create a json file to put into the finetunes folder?

https://github.com/deepbeepmeep/Wan2GP/blob/main/docs/FINETUNES.md

@technocratnumber 不，我没有。我该如何创建或者哪里能够下载json文件？

Yeah gotta have a JSON file in the finetunes so it can find the models and open it

Put the models files in Wan2GP\ckpts
Create a .txt file in Wan2GP\finetunes and paste the code below code into the .txt file save and rename to DaSiWa-WAN_22_I2V_14B_BoundBite_v10__Lightspeed__GGUF.json (the name does not matter but make sure to rename the .txt extension to .json.

Code:

https://pastebin.com/R9rXKdKs

Thank you, boss, for allowing me to try this. I don't know if I can do it.

@ernestb83414 Thank you, boss, for allowing me to try this. I don't know if I can do it.

@ernestb83414 Thank you, boss. I have found it in Wan GP and it successfully turned my picture into a video. But I don't know why, the video is very blurry, as if it were covered with a layer of snowflakes. However, I can clearly see the movements of the characters. I set CFG to 1 and Steps to 4, without using any Lora

@simarkatarezzy193475 https://pastebin.com/R9rXKdKs I have uploaded the code to pastebin as the formatting gets lost in this chat

You’re right. I compared V9 and V10 using the same prompts and parameters, and I found that without any Lora, V10’s motion effects are very weak—in NSFW scenes, it’s practically static—whereas V9 doesn’t have this issue...

@Tieasc actually im getting some really interesting results by using speed loras on the v9 model as well, i dont mean to say that it doesnt work without it, its just that adding speed loras suprisingly makes these models even better. I'm trying jumpcut types of scenes without using any loras and just by adding speed loras im getting much better "transformative" generations that actually change the scene significantly instead of just morphing the character into position according to the prompt. this is definitely worth looking into i think.

Just to clarify:
Adding distillation on distillation always add more motion, but also will introduce artifacts, camera distortions and detail shifts. Be aware of this.
The model is made to preserve as many details as possible, so the amount of distillation is already finetuned on that.

@Tieasc V10 is more sensitive to prompts, the same prompts will not make the same results, but adding prompting like "fast, slow, rapid, swift, strong ..." will change the outcome significantly.
For V10 you may have to be more clear how the motion should be, instead of just writing the motion.
V9 added overall significant amounts of motion, that would also lead to unwanted motions and camera movent. What was fixed for v10.
I do not found the images to be static in V10 if the prompt is clear what to do. Static videos may occur if a concept can not be done without lora, but that's normal for all checkpoints.

@darksidewalker can you give a rundown on what prompts the checkpoint will adhere to (especially camera movements) and prompts that work as we don't know what exactly was merged with the checkpoint? I used your guidelines as broke down the prompts with the scene type structure you described, but it provides very little change.

@unknowngreatone I can not provide that, because this is not a lora, trained on specific type of wording. You have to figure out what works best for your input and scene, a full checkpoint has no fixed trigger words.
I tested many camera movements and all worked: "pan, zoom, strife, hover, orbit, ..." - It may depend on the complete sentence and initial image, how good one or the other worked.

I would suggest using a other sampler+scheduler, I never faced that

Should not appear with BoundBite, as far as I tested

No clue

@darksidewalker Could you let me know if this model is compatible with Mac devices at all? I’d really appreciate any insight you might have.

@darksidewalker I’m running this model on a Mac Studio (M4 Max, 64GB RAM), and I found that Mac’s MPS backend doesn’t support FP8 data types, which causes errors. Could you please provide the FP16 version of the model? I’d really appreciate it!

@s_songjiafeng01 sorry, won't provide it. I also have no clue what models are supported on Mac.

Did you open up a previous generation made on an older workflow and just change the model? I had the same experience, rebuilding it in the most recent workflow didn't show any signs of slowdown.

use lower resolution and change fps, my resolution is THD and 16 fps (interpolated) and get slowmo/static movement too

When interpolating, ensure that you're using a workflow that multiplies the framerate when saving the video. If you generate at 16fps and then interpolate to double the frames but don't change the fps, you're gonna have half-speed videos (IE: it's showing 32 frames at 16fps). Current DaSiWa workflows will multiply it automatically if you've got interpolation selected - if you're using a workflow where you added interpolation yourself, you need to up the rates yourself.

I also am getting slow motion videos with v10. I didn't have this issue with previous versions. I just use the default Comfy Wan 2.2 I2V workflow and I would prefer to not need to download a more complex one.

same issue, also any one getting "get image" issue, every thing is loaded works for a few runs then that bug comes up, and have to restart comfy

Yeah that was the number 1 goal for me :)

I also want to know how to get rid of pause at the start

Happens to me as well, I've been using the creator's models since LureNoir with a very simple workflow that I practically never had to edit and I never had an issue every time I switched to any of the creator's newest models, always worked seamlessly, but then I switched from SynthSeduction to this and suddenly my animations barely have any movement now, I tried to tweak everything and simply nothing works, the weird pause at the start along with the static robotic movements is almost every result I get. I had to switch back to SynthSeduction after messing for a couple of hours with no results

I thought I was going crazy or doing something wrong after updating from v9 to v10 and getting this.

应该是VBVR lora混进了模型导致的，可以搜搜PainterI2V这个动作加速节点

I've found that using smoothmix v2 on high with lightx2v=3 and boundbite on low works very well. Sigma/shift 6 or 7. There were certain i2v generations that had the pause no matter the seed, sampler, prompt... One thing I noticed is that if wan thought the subject should be looking somewhere else, or that something else should be moved in a specific place, it would do that before anything else. Even prompting to not move where they are looking didn't help...

How do you do F2LF with out flickering on the last frames?

@granz33 enable perfect loop in DaSiWa's workflow.

Depending on the light on the scene its sometimes better to keep the color correction off

I never get flickering in my results, but don't expect a 100% looping result. The chances of those are rare depending on how fast and detailed the video is.

I have a looping video in my showcase, its X rated so beware (I believe that one is 8s).

I noticed that this sometimes is happening, it's also on some examples, but on others not. I don't really know why this could be, maybe just a prompting thing. Describing the initial situation seem to prevent this for me, maybe it's an "analysis" thing till it understands the situation. But it is only sometimes, so maybe just bad luck.

@darksidewalker good suggestion, I'll try to prompt in a way where it has kinda whats going on to start

@CubeyAI Any luck? I tried a lot of prompting techniques, then also some loras and even some speed loras. The most success I was able to get was maybe around 20-25% decrease in pause with the speed loras. With the length of the pause, it's not enough.

I really like this release, it looks good, but this is a killer for me.

I'm facing the same issue.

I was able to achieve satisfactory results by trimming the first few frames using FFmpeg.

Good 🤞

For me it was 31.03.😄

There is an obvious reason mentioned inside the announcement 😉

@darksidewalker Ah! interesting, thanks for the response. What would you say the trade-offs are? Also I should have specified even when using a lora like DREAMLY the controls and detail of this aspect is much worse than I could get before. Would v8 or v9 be a better option for me maybe?

@SolidBold Tradeoff is quality degeneration and detail shift. What is not happen with BoundBite v10.
If another checkpoint suites you more, well, that's on your side to decide ;)

That's an inherit thing, because v9 had so much motion values that it introduced camera shifts sometimes. So v10 got this tunes down a bit to stabilize details and camera shift.
Therefore v10 has to be less motion than v9, exaggerated motion values lead to unwanted jiggling, drifts, distorted details.

What helps (for me) is using motion descriptors like fast, rapidly, swift, ... depending on the scene.

Like this, on other checkpoints were exaggerated motion is baked: https://civitai.com/articles/27986/
This can be good in some situations, it is not in many others. Choose your poison 😁

This is true. Workflows running on V9 produce far more satisfactory results in camera movement and motion than V10 running the same workflow almost to debilitating limits, but the facial consistency of V10 is unparalleled. So I tried the hybrid model, using V9 for High Noise and V10 for Low noise, and that output gave better results.

@vredbat02664 I assume you just refer to overextended fast hyper thrusting motions... Since the motions are fine with v10🙄 camera movement is worlds better with v10, v9 is not even close there.

You can always add more distillation with lighx2v loras and extend the speed, sacrificing details, understanding and prompt adherence for the repetitive motion 😏
The model is not made for 1 simple task, it is made to overall perform (also on SFW!) and many styles.

@darksidewalker To be honest, ComfyUI changed its calculation methods for certain parts of the model. Because of this change from the older version, the outputs might look different or have minor bugs—something some people may or may not encounter.

@g1263495582 Any source? I'm not aware that comfyui is doing the calculations of the tensors itself and alter them...

@darksidewalker https://github.com/Comfy-Org/ComfyUI/issues/13106

@g1263495582 Thank you, very interesting O.o
But there is a mention that the issue exists if your dependencies are outdated and the problem get solved with:
2.10.0+cu130 and xformers to 0.0.35

Also it was said v9 vs v10, so if it would be comfy it would effect both.

@darksidewalker As for v9 vs v10, hmm... was v1022 mixed into the v10 mix distillation?

@g1263495582 it is partly included in both versions

@darksidewalker for example a workflow that worked well with V9 said (excerpt from the prompt) "The girl has a green emerald pendant around her neck, the girl lifts the red-wine glass from the table and walks towards the fridge. The camera zooms out to reveal full body shot of the girl walking towards the fridge with the wine glass in her hand". -- The output of this is exactly as expected, all items and parts visible and well produced.

Now to V10, the output is crystal clear with great facial consistency, but there is absolutely no zoom action, the girl's pendant is present but the camera angle is such that its not visible, the wine glass is present but is only partially visible and there is no camera motion for the full body shot to show the girl walk. The view is fixed while the girl walks to the fridge.

This output happens with no change to the prompt, only removing V9 and loading V10 and running the workflow.

Again, I will iterate V10 is an insanely good model, but perhaps for a different use case such as multiple short subjective situational prompts stitched together for long videos.

@darksidewalker Hmm, regarding Motion V10, I think the 1022 that was mixed in might have been overshadowed by other distilled models. Personally, I feel like 1022 is the distilled LoRA that makes videos look the most 'active' compared to the others. That’s probably why V9, which used 1022 directly, performed better in terms of motion.

@vredbat02664 Regarding the camera, I feel like using 'camera pulls back' yields better results. However, you need to format it something like: 'camera pulls back, [describe the background/what is visible in the scene], [followed by what you want to see, e.g., full body, etc.]'.

Regarding the camera, it depends on the input image plus the camera commands. From what I've shown in this official guide under the camera section, I feel like there's a specific pattern to it. If the format is wrong or incorrect, there's a chance that Wan will either follow it or just ignore it.

V10 needs promoting for the motions you want. The reasoning needs details. The benefit is it will not add unwanted motions out of the blue.

There is no movement problem, the reasoning needs a bit more detailed descriptions.

@darksidewalker gotcha thanks

Were You be able to get rid of initial freeze and robotic movement and lack of subtle body motions? If so I'd very much like to hear how. So far v10 is collecting dust, while v9 bangers, although half is scrapped because of camera drift.

@Kolt93: You have to describe the initial state of the image, or you'll have a bad time, like I first did. Without that description, you're animating an image instead of generating the rest of the video, if that makes sense.

@SamuraiJack109 Thanks, I'll look into that!

Thank you for your support and tha kind words, I hope everyone can crate awesome art with the model🖤

@darksidewalker You've got the best model for stylized stuff (I use v9 for nearly all my videos :D). I'm still figuring out V10, and it's great to have variety in how the models perform; different tools for different jobs!

I would say you are using it wrong. Considering the hundred examples proofing otherwise. 🤷
The reasoning will need more description, if you do not describe anything, it will not do anything ... it is more prompt sensitive. What is a good thing.

@darksidewalker I will try working with it more, I just dont understand how to describe every subtle motion that V9 did on it's own.

What about freezing? What causes it?

@Kolt93 Nobody said you need to describe every single bit, but description matters with the reasoning build in.
The good thing is it will not add unwanted motion, clam faces, you got the power, slight smirk, go for it. ~
I never noticed any freezing, so I do not know what that should be.

@darksidewalker Literally every video from v10 starts from around 1 second freeze. I didnt notice it right away too, but after looking through other's comments about it, I rewatched my generations - and there it is, every 5 second clip is basically 4 second with ~1 sec freeze at the start.

@Kolt93 Looking at my latest vid here (https://civitai.com/images/126424971) there is no freeze.

@darksidewalker I double checked - first 7 pixel frames are static initial image (I2V). Thank god it doesn't affect SVI workflow.

@darksidewalker I can only add that I'm not alone with this problem.

@Kolt93 7 frames would not even a second. Even if this is happening to you and as I said it is more likely a prompting problem. There is no proof to tell there4 are others that prompt your way and get that same behavior. I could also say there are 100 examples with no problem, so this must be a setting or usage problem.

Even IF this problem occurs and would exists, there is nothing I could change now, since the checkpoint is released this way. I do not get what you want from me.

If the problem is real and reproduce able constantly I could only try to fix it with the next checkpoint.

So if the v10 is not your cup of tea and do not produce the wanted results for you, there is no reason to use it, you can just use any other out there.

@darksidewalker I wanted to report a problem (freezing), that you would either confirm (and maybe patch) or not, thats all. 7 frames is half a second at 16 fps and less than a third at 24, so may be not that noticeable at 24, but either way - if it exists, you will know about it and remove it from V11. Peace, man <3

@Kolt93 Okay, now I understand. I can confirm this happen occasionally. But I can also say, that this is not in general, since I made a lot of video.
My clue is this is the deeper reasoning from vbvr.
I can just give you a heads up, that this is somehow connected to the input and prompt used.

It is a I2V model, so this is not intended for T2V and will not reliably work as T2V.

Thank you! Yes, I tuned down the motion a slight bit, to eliminate unwanted motions and camera shifts.

You have to run both high+low

BTW I am a loyal fan of this work. Best Wan2.2 ckpt ever!

@lillblues78249 Seem to happen on some inputs and prompt types. Best use a brief description of the setting as first 2-3 sentences.

@darksidewalker Thanks for the tips. How come V9 doesn't have such a problem?

I'm having the exact same issue. tried simplifying the prompt to 2-3 sentences, tried ensuring all the sentences describe motion rather than the image, etc. but seems to be the issue with the model.

@penguin235 as I said, it can happen, but most of the time it is an prompting issue. V9 is different and the motion vectors where too much resulting in camerashift and some detail shift. V10 has build in reasoning and needs more prompting, but responds better.

You can not effectively merge loras into quantized models.

All information I want to share is inside the corresponding article. Anything above this will not be shared to protect my work from stealing :)

@darksidewalker ok but how to know what we can do with no lora ? you don't list prompt word or exemple ? for blowjob exemple, i use sensual_teasing lora or Deepthroat/Face Fuck- Wan2.2 I2V lora, but i see problem with this Lora, then if you already use this one in your check point i use its two time for nothing :/ Do you use your own blowjob lora ? Thank you very much for your answer and your work

@djoedjoe15454845 You are right and also not. Since it is NOT a lora and not trained on triggers or specific sentences and be a full checkpoint, it will listen and understand complex settings and natural language. There is no way to fully provide a list of working sentences, since it always is depended of the context, attentions, encoders, seeds and inputs.
To check if a lora will add any benefit you can just run the same seed and setting and see if this lora will add anything you want or the model already understands that on its own. that's why I recommend to try without first.

To know what you can do without lora is trying out, that's how it works with ALL models out there. Nobody can tell you all capabilities of an checkpoint and there is no way to test every possible setting.

You can find detailed descriptions under 'About this version' on the right side of the model page. Additionally, NVFP4 and GGUF versions are available for v10.

Also there are detailed articles and announcements on my page to explain everything.

Thank you:)

What is your workflow? Using a regular (default) WAN2.2 workflow v10 is broken for me. The result is unpredictable.

@deepz hey sorry late reply but i use darksidewalker's FastFidelity all in one I2V workflow with just the I2V, upscale x2, NAG, and interpolation turned on. it can do lots of other stuff too check out the description. link: https://civitai.red/models/1823089/dasiwa-wan22-workflows-or-i2v-or-svi-20-or-s2v-or-flf2v-or-audio-or-combine

This is my SVI WF: https://civitai.com/models/1823089?modelVersionId=2580650

DaSiWa (the same account that posted this model) has a great all in one video workflow that I highly recommend. Its very modular so you can set it up however you like!

Here: https://civitai.com/models/1823089/dasiwa-wan22-workflows-or-i2v-or-svi-20-or-s2v-or-flf2v-or-audio-or-combine

@Nk0GGry the fastfidelity workflow? That's just a straight resize upscale, not a full final pass upscale.

I am having issues with v10, v9 works fine even without any loras.

@deepz Whatever you issue is, please read the announcement and front-page to the model. There are in depth explanation how the model works and that it is not meant for replacing all loras.

@darksidewalker I mean v10 doesn't work. It is broken for me. It doesn't do simple SFW prompts what v9 used to do well

@deepz I doubt that, v10 has much more reasoning and prompt adherence. I have no clue what you are doing, but v10 is much more capable of doing things, even more for sfw.
You need more prompting as for v9, because of VBVR - Just use a bit more detailed prompts. And a brief start (as always) describing the situation. There are plenty of examples already in the gallery.

@darksidewalker understood, I will double-check with more test cases assuming the parameters stay the same as in v9

@darksidewalker  i like how there this big talk in side my question but my question never got answered lol

@crafted101 Oh, yeah, I assume this is prompting or the situation needs a lora.

I did test it again. v10 is better at preserving identity but movement is stale or non-existent

@deepz Sight... How can you say that? Look on the gallery. It is just not true.
But well, if you do not like it or its not how you want it, there is nothing I can do, you may just use another checkpoint. There is no point in telling me that over and over in the comments and other comments. I can not just rewrite the checkpoint to your needs.

@darksidewalker maybe I have a broken model I will re-download it. But I did test in more than 4-5 repeatable cases where are got excellent results with v9. Could be the prompting or the settings.

@darksidewalker You are right! I might have had a broken model the first time. It seems to be much better when I re-downloaded it! Good Job!

@deepz Glad it now works for you :)

@darksidewalker maybe give some prompts example?

@junyung99_tan302 99% of my uploads are with prompts, also from others... So there are enough examples already.👍

After 5s WAN 2.2 will gradually start looping back.

@darksidewalker Thanks. I suppose it's not that big of a deal, just need to stick to 5 seconds to avoid undesired looping behavior and go longer if I DO want the looping, since it is actually helpful in some instances like creating extended clips via end frames

There is, just look on my profile/collection 👍

I have no clue what your talking about can you specify?

@darksidewalker Sorry about express not clearly.I don't know much about backing models，but I realize that VBVR don’t has a lora format released until a few days ago，so V10‘s VBVR originates from the whole VBVR Diffusers format model as base model？Since V10’s
generation often has a slower motion than V9 when using same input and promots especially with SVI Pro,I wonder if it's VBVR Diffusers model‘s influence？Now VBVR has official WAN2.2 lora released，could backing VBVR lora in V9 provide better results since lora is more concise，conducive to a balance between dynamics and reasoning？

I will be very sorry if I am wrong with how V10 implements VBVR，but anyway you did a great job，your model is consided as one of the best WAN2.2 models in our community.

Thanks for your great work！

@ql_weepingwind036 VBVR also has general Wan 2.2 checkpoints with it built in. One person I read claimed the lora can't work as well as the checkpoints because it's a lora but I don't know if that's the case. One video tester claimed the KJ lora works better than the checkpoint but he only tested the "surgically" compressed version rather than the full-fat VBVR checkpoint. I have found that the full-fat VBVR checkpoint has better quality than the allegedly "surgically" compressed one. It is slower, though. I am not sure VBVR is always an improvement. Sometimes it seems to improve things and sometimes it seems to be a hindrance, whether it's the lora or the checkpoint. I have not tested any lora other than the unofficial KJ one.

@ss9999 I see....So the most important thing is still to use more detailed prompts.Oh.I hate constantly adjusting the prompts and generating them repeatedly....sad....

@ql_weepingwind036 There are a lot of variables (which loras to use and at what strength, how many steps, what sampler/scheduler, how much shift, which checkpoints to use, whether or not to use accelerator loras and which ones, whether or not to use things like NAG and VBVR, etc.) and unfortunately there is a lot of trial and error.

Well BoundBite is more prompt related than SynthSeduction.
I took a lot of time to test it and found BoundBite way more "precise" than SynthSeduction in understanding the prompts and also is more versatile by keeping the input image without too much changing it. I'd say, darksidewalker made it's own VBVR in his models without using the actual LoRa in it.

To be clear, i always test without any LoRa. This is the best method to get the boundaries of the checkpoint, it's capabilities and what it can or cannot do.

I noticed that BoundBite react better with "strong", "repeat", "fast", "wide" trigger words. Also specifying physics helps a lot like: "velocity", "friction", "gravity", "weight"; specifying material "type", "state" and also motion "type", "state", "direction", "placements" (which are not necessarily needed in SynthSeduction).

I found that BoundBite react way better to science than SynthSeduction and is way more stable. Yes, motion is fast almost everytime in SynthSeduction, but it lose a lot of the prompt's detailed explantation and for what i've tested is too much doing whatever he wants.

The workaround for BoundBite: If it's slow at 16fps (81 frames) that means it will be okay at 24fps (121 frames). If it's too fast at 24fps it will be okay at 16fps. If the motion takes too much time to start (like early frames taking time to start the motion), that means your subject's description trigger some motion informations from the model. In that case put the subject's description AFTER the motion explantations and make sure to weight what you really want to see like (motion:1.2).

Also don't push the number of frames (number of seconds) too much expect if you wish for a loop/repeat result. For example: "A man walking from left to the right" > Result: going from right to the left in last frames (like backward) means the time is too long for what you asked according to motion description.