🌸💀 DaSiWa LTX 2.3 | Lightspeed 💀🌸
My new LTX 2.3 model for I2V, T2V, V2V generation.
Version overview: https://civarchive.com/articles/23495/dasiwa-model-versions-and-timeline
⚠️ v2 is BETA RELEASE! | v1 is ALPHA RELEASE!
‼️IMPORTENT: I found a bug, making the model v2 worse as it should be!
I uploaded a better version. All who spend buzz can just re-download when the new files.
FP8/NVFP4 are back again!
I apologize for the inconvenience and hassle 😣
Expect that not everything is perfect and mind LTX2.3 is not as stable as WAN 2.2 finetunes.
🔮 Key Features:
🔥 Best With I2V and V2V
☄️ Really Fast generation
🔊 Better Sound
🗣️ Better Voices
🌟 Enhanced Quality and Reasoning
🔞 Unrestricted
🪄 Better Prompt Responsiveness
🥺👉👈Better understanding of anime/manga style composition
🪡 FP8+ mixed precision
😵💫 Reduced some hallucinations
👘 Strengthened visual consistency/understanding for anime
🍒Workflow
Make sure to checkout my easy to use Workflows!
🍄LoRA's
But: This checkpoint is not meant to replace all LoRAs, it is meant to:
Perform better overall at his own
As easy as possible to use
With LoRAs to be more awesome
⚠️ Read the corresponding announcements.
📢 Make sure to check it out for in-depth information and a complex comparison!
🛠️ Recommended Settings
CFG 1
Euler/linear_quadratic
8 Steps
Dependencies
VAE
LTX23_audio_vae_bf16.safetensors
LTX23_video_vae_bf16.safetensors
Dual CLIP (Encoder and Projection)
gemma-3-12b-it-heretic-v2_fp8_e4m3fn.safetensors
ltx-2.3_text_projection_bf16.safetensors
🩻 Known issues
Tell me 🫵🫢
LTX2.3 be LTX2.3 🫣
Hands are sometimes unstable
Shifting of fine details (e.g. eyes) without prompting or really high resolution
Needing way more runs for good results than WAN22
🩺 Fixes & Feedback
If you use LoRAs, try to respect the LoRA training triggers and try some versatile descriptions, most LoRAs will work with 0.3-1.2 (start with 0.3)
Do not mass add LoRAs, just add 1 or 2
Negative prompting do not work with cfg 1, thats a limitation of speed-ups with cfg 1
Before posting any questions I suggest reading my guide.
Update your ComfyUI ❗
🖤 Why I Made This
Pushing LTX2.3 to its limits!
This checkpoint is also my personal playground.
Closing words
🤩 I want to thank all the fantastic other creators who made super nice LoRAs and concepts to play with! Support that awesome creators by using their LoRAs and post to their gallery and share the meta-data!
⚠️ I made all this with permissions or open-source resources (the time it is incorporated).
I share as much insights as I can without compromising my work. I'm doing this for fun as my hobby and just do not want my hobby to be destroyed.
More details can be obtained in the corresponding announcements!
If you would like to contribute in my awesome (😉) checkpoint or willing to share resources I'll gladly give credit! Just contact me!
✅ All credits / resources are mentioned inside the announcements! - Since different versions may have different resources.
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Disclaimer
This models are shared without warranties and with the condition that it is used in a lawful and responsible way. I do not support or take responsibility for illegal, harmful, or harassing uses. By downloading or using it, you accept that you are solely responsible for how it is used.
LTX-2.3 Custom Addendum: Fine-Tune Integrity & Attribution
Base License: LTX-2.3 Community License Agreement
1. Verification & Integrity Requirement
This model is a fine-tuned or merged derivative of LTX-2.3. To ensure users receive the correct weights, safety metadata, and version updates, the Official Source is maintained at: https://civarchive.com.
Notice of Non-Support: Any versions hosted on third-party platforms (mirrors) are considered "Unverified." The creator provides zero warranty, support, or safety guarantees for unverified files.
2. Trademark & Branding Restriction (Pursuant to LTX-2.3 Section 8)
While the underlying weights are subject to the LTX-2.3 distribution rights, the name "DaSiWa [Model Name]" and any associated logos or promotional imagery are the intellectual property of the creator.
Renaming Rule: Any Entity or individual redistributing or mirroring this model on a third-party platform (including but not limited to Hugging Face, Tensor.art, or SeaArt) MUST remove the original model name and branding unless explicit written permission is granted.
Source Attribution: Redistributors must provide a prominent link back to the Official Source as the primary point of origin.
3. Commercial Platform Restriction (Pursuant to LTX-2.3 Section 2)
Commercial Entities (as defined in the base license) that generate revenue through the provision of "Generation-as-a-Service" or ad-supported hosting are prohibited from using the official branding of this model to market their services without a separate agreement.
If your platform charges "credits" or subscriptions to access this specific fine-tune, you are required to contact the creator to ensure compliance with my project.
Description
Initial ALPHA release
FAQ
Comments (94)
Oh my.
It's seen your checkpoint to like 3d too much, wend come to the video, I try to recreate this scene- https://civitai.com/images/127361754, but the character looks real
Yeah noticed that too, more power to anime is on the list
@darksidewalker I think the anime also needs to work since I tried with one of my Lora, and its 3d. its work best for real people the anime
@Agino I do not understand O.o
Well, at the moment, this is just a test, a test, hands penetrating the body, second and third penises... And this is i2v :D
Well, as I said, expect it not to be stable 🫣
@darksidewalker I'm waiting and hoping that you'll compare ltx to Wan's quality.
@darksidewalker Try using the Eros model posted here as a basis. The author has stopped supporting it and has given everyone the right to improve it. Perhaps it will be easier for you this way than from scratch.
@Renessance I know this model. Not a good base, sorry.
Also I'll not compare WAN 2.2 with LTX2.3, since they are totally different, this makes almost no sense.
@darksidewalker Poor translation, apparently I meant that you will make a model that is not inferior in quality to WAN
Ohh whaaat no you did not.. I must try it :o
😆🙀I hope it will not destroy your expectations. It is ALPHA and for funding my next iteration.
@darksidewalker Haha nah I get that. Ill be sure to test it like crazy.
From testing It does like to lean more on the 3d side from what I noticed as well. It struggles with interesting angles. I tried a from below shot and some reason the mouth didn't really move, and the body reasoning was pretty bad from camera movement. All in all pretty good for alpha. I'll have to try more stuff.
@PopcornMaven Yeah I noticed that too with the "3d". The rest got not tested by me yet. But I noticed resolution plays a major role for ltx2.3, although it can generate much higher resolution for me compared to wan22.
@darksidewalker Yeah, I noticed that it generally performs better with higher resolution images as well. For resolution I've only tried precision up to 1.35mp but it looks pretty clear and definitely has less face tearing. Though it leans toward 3d it does keep the general look right now so that's pretty good.
Thank you for your work and effort. It would be helpful to know which LoRAs you've baked in. Thanks again. ♥️
Hi! Thank you!
All extra shared information are here: https://civitai.red/articles/28671/
@darksidewalker it says 'article does not exist'
@McClippy blame civitai I guess...
@darksidewalker Thanks! Helpful article. It didn't load the first time but now it's working again. ❤️
Hii, thanks for making this :D... I just have one question, it's normal all the videos having a random woman voice laughing or moaning even if there's not a woman in the scene ? haha
Absolutely ... This is a feature 😆, but if you don't like it try to prompt what sound is to be heard, since ltx23 will do random sound if not prompted.
@darksidewalker Ohh.. That's why xD. thanks
Compared to the distilled_transformer_only_fp8 version, the liquid details are much richer; however, maintaining character consistency and controlling the voice output are somewhat difficult to manage. I ran several tests today, and in some instances, it caused the hentai_voice_ltx23 LoRA to stop working. I hope the developer continues to improve this project!
I knew you were going to do it. :P I wish you good luck improving it in the future. :3 ♥
Yeah much luck I'll need 😆
you are magnificent, thankyou so much!!!!
I don't get your workflow running. Even after I install all the custom nodes and downloadedthe models. And I can't figure out why because it is too complex for me to understand. Is there a more simple workflow?
Not that I'm aware of. There are others, but more simple ... 🤷
You could try UmeAiRT's workflow here:
https://civitai.red/models/2329567/ltx2img-to-video
It may be easier to set up, though the quality might not be quite as good.
Ty. I will try the other workflow and also try to find out why yours is not working.
@Pressydent If you find out, please tell me :)
Reporting in: Generating Chinese songs directly works fine. The automated dubbing and background music are amazing.
promising start - i'm getting better detail than the Eros model for less time spent, but having issues with generated audio and anatomical sense for high movement scenes (e.g. people's arms pointing backwards, head rotating 180 degrees etc) with just naive model switch in my existing workflow. Needs some more tweaking on my part but i like where this is going. looking forward to watching this family of models progress as the WAN set did
Yeah I'm totally will update the series and when stable, also add more quants if possible.
For scenes with large-scale movements, increasing the LTX 2.3 FPS settings will yield much better results.
Noted
I'm getting choppy videos if I tune it up to 48fps. Could you tell if any LoRa helps with this ?
Whenever I run into transition issues, I try using this LoRA: https://huggingface.co/joyfox/LTX-2.3-Transition-LORA
Great workflow, it works for me. but...i keep getting the error about missing placeholder.gguf . can not find it with a google search. i prob am doing it completely wrong but how do i get placeholder.gguf ?
Placeholders are placeholders, no real files.
Exactly. Its just there just in case you want to use a gguf version of the file. Errors are a little annoying, just gotta ignore them <3
@MedliKnight Or hit the disable option in comfyui to not show them
i went into the subgraph and delete the two disabled nodes. Seems to have 'solved' it.
I receiving errors for 'tae.safetensor' in Dasiwa workflow...
@herkus_baronas631 Download an place the file
@darksidewalker How to download, if i have it? and where to place it exectly, if i spammed it in many folders?
@herkus_baronas631 Read the notes please, all links and where to place is described
The model is good, but the workflow, unfortunately, doesn't work. It's impossible to switch anything (operating modes), everything is unclickable. It's like the switch is always on for all 5 buttons, and that's it, absolutely everywhere. Nothing clicks anywhere, it just says "frontend only" and that's it. As a result, only image-to-image is available, and even that periodically breaks due to TELTX or something related to VAE. I don't know what's wrong, but the workflow is simply unusable.
Seems like you enabled Nodes 2.0 (they will break everything). Also the kijai nodes need manual reinstall since last comfyui update.
@darksidewalker Actually, I just installed the latest version of Comfort for Windows and deleted everything beforehand, so there was no node transfer; everything was installed from scratch. And yes, you were right – these are Nodes 2.0. I'm not sure what's wrong with the second version, but at least it's now possible to switch. It's likely that if the installation was clean and there were no updates, it's unlikely to break anything.
@darksidewalker Regarding the KZ nodes, you were right. This problem actually appeared now that I was able to switch modes. Deleting the nodes and reinstalling them solved the problem (but I didn't know about it back then). The issue with TelTX is still unclear, but maybe the problem will go away now. We'll see. And we'll eagerly await more stable versions. :) In the meantime, since you're reading the comments, I'll give you some feedback (it's unlikely you're not aware, but just in case): the genital detail is poor. And more importantly, the prompt enhancer doesn't work with Unsafe, although it generally doesn't work with everything "unclothed or partially unclothed." Well, we'll be happy to see new versions. :)
@lera2222 prompts with NSFW needs an uncensored Gemma model
Hello Dasiwa, should we use a distill lora with your checkpoint or is it already distilled/fast ?
It's baked in, also stated on the front page
cooms in my pants
please set up a ko fi so we can show our appreciation !
There is, also mentioned on my page ~ Thank you for the support :)
https://ko-fi.com/darksidewalker
@darksidewalker done ! thanks for your hard work!
@gackt2 thank you very much 🙏
can you create a version that doesn't include the distill? what strength is it appied? i get baked to hell results
Edit: adding distill lora at -0.5 seems to help
Not for now, it's an ALPHA and still in development. I'll not make multiple versions on the fly till I got a stable base.
bro added the distill in negative weight 😭🧠
作者辛苦了，画面质量很好，就是感觉动态比原模型慢了不少？是哪里设置的问题么？
when do you plan on dropping the gully trained model because this alpha aint it cu
What?
@darksidewalker was a joke but in all seriousness, atleast for me the version you released for early access was unusable. no prompt adherance whatsoever
@Ragamuffin20 Well, it is ALPHA. Also you could speed up the release by sending me an rtx5090 😆
@darksidewalker ik im just giving feeback and trust and believe me if i could i would but a 5090 is looking more and more like a pipe dream
@Ragamuffin20 I know and time is sparse on my end atm. 😑
@darksidewalker, or an NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition! :3
one step at a time, good luck if you ever keep trying to get a goot base model for ltx
I'll, just need time😺
@darksidewalker for now i'm honestly surprised with the results, really looking forward for whatever you have in mind next
Unfortunately this model completely modifies faces unlike your wan 2.2 model which has amazing identity retention
May happen, based on settings. It's Alpha and still LTX23 and not WAN 22.
@darksidewalker do you have any recommended i2v settings to reduce this effect?
@Plaguekind not really, still investigating myself
@darksidewalker what strength do you have the distilled lora merged in at? 0.4 keeps faces well retained
@darksidewalker thank you for your hard work . we're looking forward to any discoveries. I've been poking around myself to try and tweak LTX2.3.
I made a clip with humanoid dogs, so instead of dog heads, the model stuck masks on people's faces, even though I2V was used.
I made the video with regular wf because the DaSiWa workflow doesn't work for me because it can't find one of the text models, even though they are there.
@herkus_baronas631 Funny one XD
Besides that this might be from overtuning, the next version will be more stable.
Regarding T2V WAN22, I have no T2V WAN22 workflow.
Can you share which distilled lora have you baked in ? 384 v1.1 or dynamic fro09 ?
384 and fro09
@darksidewalker both ? I suggest you only use 384 v1.1 and at 0.5 strength max as it has much better prompt adherence & motion guidance.. whats happening is that while the generation is fast its not making a lot of objects move as they should, and in many cases turning anime images into semi realism. I tested prompts without any loras, also with imagetovidadapter with yours comparing it with the distilled v1.1 gguf q4 that I use otherwise.
@BopStar I mean both on different models.
@darksidewalker the beta version will have which one ?
@BopStar the one release along sulphur
WTF is text projection ?
A CLIP file. Read the notes. 👍
I thought it was the model's ability to render text in the image.
好奇怪，做出来的视频有明显的卡帧。我试了两天了，用了作者的工作流，也依照他的说明把所有的模型都重新下了一遍，结果还是很明显的一卡一卡的。也不知道哪里做错了。