๐ธ๐ DaSiWa LTX 2.3 ๐๐ธ
My new LTX 2.3 model for I2V, T2V, V2V generation.
Version overview: https://civarchive.com/articles/23495/dasiwa-model-versions-and-timeline
A comparative study:
https://civarchive.com/articles/29961/dasiwa-or-major-ltx23-model-comparison-part-1
https://civarchive.com/articles/32224/comparison-between-major-ltx23-models-part-2
Expect that not everything is perfect and mind LTX2.3 is not as stable as WAN 2.2 finetunes.
โ ๏ธ Make sure to open the DOWNLOAD dropdown to see all quants possible.
๐ฎ Key Features:
๐ฅ Best With I2V and V2V
๐งช Optimized mixture
๐ Better Sound
๐ฃ๏ธ Better Voices
๐ Enhanced Quality and Reasoning
๐ Unrestricted
๐ช Better Prompt Responsiveness
๐ฅบ๐๐Better understanding of anime/manga style composition
๐ชก Finetuned mixed precision's
๐ตโ๐ซ Reduced some hallucinations
๐ Strengthened visual consistency/understanding for anime
Different versions may have additional customization (read the version notes)!
โ Included VAE and not included VAE
๐งชDistillation and non-distillation
๐Workflow
Make sure to checkout my easy to use Workflows!
๐LoRA'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
โ ๏ธ Read the corresponding announcements.
๐ข Make sure to check it out for in-depth information and a complex comparison!
๐ ๏ธ Recommended Settings
CFG 1
Euler_CFG_PP/linear_quadratic
8-10 Steps (distilled)
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
๐ฉป Known issues
Tell me ๐ซต๐ซข
LTX2.3 be LTX2.3 ๐ซฃ
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
Like all LTX23 checkpoints at the moment this can have ๐ป ghosting dependent on the scene, motion and used Loras and workflow's settings. After all LTX23 is still unstable.
LTX23 is very dependent on the used workflow + settings!
๐ฉบ 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 โ
๐ค Why I Made This
Pushing LTX2.3 to its limits!
This checkpoint is also my personal playground.
Closing words
๐คฉ 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!
โ ๏ธ 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 (๐) checkpoint or willing to share resources I'll gladly give credit! Just contact me!
โ 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
๐ฅ Deeply optimized for I2V and V2V
๐ Enhanced T2V (compared to v3)
๐งช Superior SOTA Distillation (DMD + Lightspeed)
๐ Experimental motion and action enhancement (Highly dynamic)
๐๏ธ Enhanced visual consistency over full video
๐ Enhanced Sound + ๐ฃ๏ธ Voices
๐ Enhanced Quality and Reasoning
๐ Unrestricted
๐ช Better Prompt Responsiveness
๐ฅบ๐๐Better understanding of anime/manga style composition
๐ตโ๐ซ Greatly reduced hallucinations
๐ป Overall reduced ghosting
๐ Strengthened visual consistency/understanding for anime
๐ฏ Included VAE (works with extra VAE; Safetensors)
๐ Up to 4 minutes (multi-shot) video length
FAQ
Comments (67)
Hello ! So, to be clear, V4 is already distilled/DMD and we don't need any turbo lora right ?
You don't
Hi Dasiwa. Your checkpoints have been the best choice since Wan2.2 especially on flfv flow.
Is your newest model available by supporting in different platform?
Not for now, later HF will be available
@Darksidewalkerย Hope this method is activated asap, i'm not friendly with crypto..
@willshawn2519ย After EA, obviously
I've been using your Wan models for a while and they've been great. Can I ask if this Ltx version of yours can be run on my old GTX 1070? Thank you.
If you can run WAN22, you can run LTX23
Its crazy to see how far your LTX builds have come since Treasure Chest. v4 is a massive step up, especially with audio and voice prompting - congrats again!
Thank you! โค๏ธโ๐ฅ
Ive been having a lot of issues with deep penetration with LTX 2.3, your model seems to be the best one in that regard but it still happens. Any idea how to fix the half-inch thrusts issue? xD
ive tried compact prompts, full descriptive paragraphs, motion loras, no loras and i sometimes get one lucky result among 50 failures. Its either that or the rubber dick that just gets compressed without the sliding in motion.
I do not even know what "half-inch-thrust-issues" could be O.o
Try adding impact language into the prompt like: "hard impact, her body is pushed up with each thrust" focus more on describing the thighs hitting than "the penis goes all the way in" and stuff like that. Try describing what that impact frame should look like more than what it should be doing, if that makes sense.
@Darksidewalkerย It actually looks very very similar to the second video in your model gallery for V4, the one with the standing face-to-face sex catgirl with black hair and pink tips, Lumi i think.
Maybe in your case you simply prompted incomplete/shallow penetration (or maybe the model understands the anatomic difficulty of going balls-deep in that position??) but in my case it doesn't matter if its even like a doggy or cowgirl with ample dick length to spare, it just barely goes in at all.
Either way imma try some prompt workarounds like @FirstPrinciples suggested or maybe just give up on the few images im trying to work with and move on.
what text encoder should be used?
Any LTX23 compatible one.
Iโd like to purchase V4, but on RED, the only way to get Yellow BUZZ seems to be through cryptocurrency, which is difficult for me.
Would it be possible for you to also make the model available on civitai.ai?
Not possible, they won't list it that's nothing I can do. It is NSFW.
@Darksidewalkerย
Thank you for your reply.
Iโll look forward to the day I can use your wonderful model, and Iโll wait until it becomes available.
any chance for an int8/convrot version? if the answer is yes, i'll throw that yellow buzz at you as soon as the int8 is up
Already inside since the beginning the download section. Comfy just has no selector for it, it is the second fp8, mouse-over reveals the true name.
The difference from your first LTX version to this one is night and day. I remember how it had a hard time changing the looks of the characters and now Itโs my go to if I want to keep character consistency. The NSFW side has only gotten better. Just gotta fix the random addition of nipples in random places. I assume you combined the jiggle Lora in to? Other than that the only thing left to be desired is still more mouth movement. They can be very lazy when it comes to speaking sometimes.
Thank you!
The random nipples is the thing I see after everything was done, it is not often, but can happen. A thing my tests did not reveal beforehand.
If I make a next version I'll definitely look into this.
Most audio comes from the model, there is very few things that can be done for audio, so mouth movements maybe not changeable, but overall I find the sound and lipsync very good compared to other possibilities..
@Darksidewalkerย I use your other model (I won't lie I don't remember which one), I generally like them, but it does try to find breasts a lot, and appearance of nipples and the need for breasts to just be somewhere to bounce were a source for me to delete a lot of stuff.
I can't say as a fact that the problem is in the model, but it is a reoccuring problem (At least on wan)
would you put this on a fanvue so its possible for us to buy it. another lora creator does it on there. since crypto is difficult
Great work, friend! I'll always support you with early downloads. I can't wait to test this Int8 on my RTX 3090!
Thank you!๐ค
Love the model, any chance for int4 version in the future?
If any major support and implementation happens, sure.
@Darksidewalkerย thank you so much for the int4
@JSCammieย have fun, but the quality is low
v1 prototype -> v2 usable to some extent -> V3 decent performance -> V4 good performance, looking forward to V5.
You nailed my announcement to that ๐๐
Is this a comfy locked in model like eros10 or a clean open training that actually works with the official LTX-2.3 GENERIC inference code.
I didn't test with LTX inference, but it is not comfy locked
AMAZING!!!!!!!!!!!!!
Simply the best even at lower resolutions. I'll remember to tip when I can at the end of the month.
The sound still has electric current sound, but the recognition of anime characters is significantly more accurate. I set the official distillation lora and voice intensity to 1 to make the electric current sound disappear. I am looking forward to the next version and thank you for your work.
Increasing cfg of the distill lora(only voice attention) will lead to a decrease in dynamic, but the slow insertion scene is also very interesting
I thought the distillation was baked in, you shouldn't need to add one distill lora right ?
@Elmer588ย Yes, but I have the same issue, I also use Wan2GP, I am not a expert so sadly and actually I don't understand where the problem is.
It works really fast, but for some reason, it makes everyone's makeup look bright (especially their eyelashes). Am I doing something wrong?
Hi, Dasiwa.
Since v4 is distillation baked-in,
I'd like to ask for recommended manual sigma values.
Honestly, for two step workflows, use 8 steps of linear_quadratic, denoise 1 for 1st and 4 steps of beta, denoise 0.42 for second. There's no magic in the sigma lists, and those will give you outstanding results, while letting you add an extra step here or there to experiement without recalculating sigmas.
Are we sure the distillation is baked in? Because I'm getting terrible results unless I add a distillation lora in my current workflow...
@Hookflashย I would guess not, which is why he released his own DMD lora.
@MikoGoblinย i initially tried with traditional sigma and the result was awful, which is why i suspected if more than 8 steps are needed.
@Hookflash which is why i posted this inquiry. enabling distller with low denoise (0.5 on single frame, 0.8 on first-last frame) helped. might try using his DMD rank72 later.
V4 is distilled, but I'm experimenting with the needed step count. Sometimes 8 works perfect, sometimes up to 12 is needed. Adding extra distillation may help in some situations or destroy the videos, it depends.
I'm getting consistent crash with the gemma 3 12B you recommend. Anyone else has this problem?
I'm running a 4090 and it was fine until this particular workflow :/
I using Gemma 3-12B-it-heretic-v2-fp8 and don't have problems. So can you to try too?
@herkus_baronas631ย I tried both versions (heretic and non heretic). It seems that SageAttention is causing the issue! I did not encounter the problem by deactivating it. It just destroys my render speed :')
@Fellaitioย ask any agentic harness to fix it for you like grok/chatgpt/claude, they can do a way better job than I ever could of fixing this stuff and its extremely fast, chatgpt in codex is free btw
@torikokoย Copilot, Grok and Google told me to use Quantized or GGUF gemma. I tried, it didn't work. I'm using LTX workflow from Dasiwa and a fresh portable installation from Dasiwa as well.
@Fellaitioย im talking about a agentic harness, you know like claude code or codex, they can act inside your pc directly and make the changes necessary to make it work
@torikokoย oh, my bad. Misread! No i don't have that installed and will probably not, but thank you for the idea.
As far as I remember I recommend Gemma 3 12b fp8, and it should work as intended. Or is there a wrong link?
Use Gemma API text encode, you will save on resources.
Nobody has a good P pulling out of V for LTX. Make a name for yourself, and bake something. This looks great.
https://civitai.red/images/136797189 is the best I have gotten by setting the creampie as a prompt block end frame in LTX Director and letting it bridge the two
DaSiWa just doesn't miss. As an experiment I tried setting up an I2V pipeline only using tools he has uploaded. Used his comfyui installer, LTX workflow, models, and loras, and within 30 minutes I was generating videos just as good as anything you've seen on this website. Long Live The King.
Yep his suite is cradle to grave and it is great
Thank you both for this super kind words ๐
Wait is finally over! Thanks for this, I became a fan of your work after your latest WAN2.2 model... character consistency is already great in it, so I'm hoping your LTX version does the job as well.
Compared to the previous generation, Version 4 brings significant improvements in image quality, consistency, stability, motion effects, sound, and audio effects. This is a fantastic release โ many thanks to the creator!
v4็็ธ่พไบๅไปฃๅจ็ป่ดจใไธ่ดๆงใ็จณๅฎๆงใๅจๆๆๆใๅฃฐ้ณๅ้ณๆๆน้ข้ฝๆๆๆพๆๅ๏ผ่ฟๆฏไธไธช้ๅธธๆฃ็็ๆฌ๏ผๆ่ฐขๅคงไฝฌ
is V4 already distilled? when i gen without the DMD i notice the usual noise from forgetting using a distil lora
i tried int8convrot and nvfp4
with DMD they work perfectly
in wan2gp, maybe that's the problem
Best sound for a video compare to 10 Eros