🌸💀 DaSiWa LTX 2.3 💀🌸

it's because at a CFG of 1 you prompt doesn't have enough weight to keep certain actions that defy what is "common sense" and someone leaving the ground is against the model's common sense and if you're doing I2V it's also against the reference image, so your fighting against the reference image and the model's "common sense" go do a render with 50 steps and a CFG of 4 without the distillation lora and you will almost certainly get the motion you were looking for, although it's likely that if you are doing I2V that the model will also fairly significantly change the look of the characters, not that I have done a base generation yet.

probably a middle ground of a CFG of 2.5 and ramping down to a CFG of 1 for the final low sigmas stages would fix that though, but you would need to look up or ask a LLM around about's where LTX 2.3's sigmas fall is then match the CFG curve to decay during the low sigmas pass. Again I'm guessing here, the changing the CFG might very well screw up the sigmas being real.

@Etheoma ok,thx.

It might be from moving large models from your SSD and into your System memory all the time. It's not related to this model. A BDOS will occur if some sectors of your SSD are starting to fail and windows loses something critical while doing a task.

Not a checkpoint issue, pretty sure it's related to LTX 2.3 model memory management. Try adding the argument "--disable-pinned-memory" to your run_nvidia_gpu.bat file (assuming you're using comfyui).

I downloaded a new LTX2.3 workflow (using the standard ltx-2.3-22b-dev-fp8 models) a week ago, and it would crash my system repeatedly (4090 24GB VRAM, 64GB RAM). Strange thing was that it didn't crash during the generations. It would crash after I closed the browser or the terminal. Wan workflows with larger models and even other LTX 2.3 workflows worked fine for me though. Claude AI recommended adding the --disable-pinned-memory argument to the bat file which finally fixed the issue for me.

Did you just updated your bios? amd push a new memory feature which may automatically enable but some memory stick won't be stable.

@quadrazz93101 No, I did not do any changes to my system in the past half a year. I also recently ran Memtest86 with flawless result.

@barry9000 I use SwarmUI but it's basically a ComfyUI frontend.

@Darksidewalker I got 48GB RAM and over 100GB of free space on the drive where my pagefile resides. I could try freeing up some more space but not really sure it needs that much...

@Darksidewalker hey i just wanna thank you for all the work, everything i've used work really nice.

Sarcasm?

@Darksidewalker Thanks, will try that.

I am also having the same issue

I'm also getting the same on either fp8 or nvfp4

Interesting, I had horrible ghosting with V2, and V3 (with the latest workflow) mostly solves the issue

Honestly, I blame Civitai's poor layout for this. The author actually put a note under 'About this version,' and the second line clearly says 'True Vision (no distillation).'

It is mentioned in the right side under About this version

@Darksidewalker Thanks

any ltx 2.3 model should work with it

@Darksidewalker I understand, I tried the FP8 model and it's more or less the same (lots of ghosting and blurring). Should I double the steps or sth like that because V3 is truevision ?

Do you mean that this https://civitai.red/models/2498991/dasiwa-ltx23-workflows-or-i2v-or-flf2v-or-t2v-or-v2v-or-audio is not adapted to V3 ? I used the distilled Lora you suggested.

@Darksidewalker where can I find the recommended settings for this?

LTX treats input like guidelines more than literal start and stop points like WAN, you are probably not going to get as clean of a output just running it by itself. It helps if you find a style lora for LTX that matches your input - although v3 does a good job of staying fairly faithful to the original format compared to v1.

You can always combine the two as you can use the V2V option to add dialogue/audio/lipsync to a cleaner clip you generated with WAN first

@FirstPrinciples I'll try thanks a lot for this suggestion!

@FirstPrinciples Exactly this! I have tried 100's of LTX23 I2V generations in the past few months and rarely do they even come close to identity accuracy as WAN22.

I'm basing all my flows on Dasiwa's WAN and LTX workflows. But the difference is night and day with WAN.

Do you think we will ever be able to achieve clips greater than 5 seconds with WAN? For me that is the only defining benefit of LTX, with audio being a bonus of course.

Is it just the nature of WAN22 that you're always going to be limited to 5 second clips, and your only choice is to chain videos using last frame generations (using a 3rd party video editor)?

I started my AI journey with WAN2GP and it had a great tool where you could continue your video, using a starting video. I can't find that feature in Dasiwa's WAN workflow, only in his LTX workflow can I see V2V.

Seems better, but large motion still causes some changes. The image strength settings needs to be towards 1

@Darksidewalker  I see thanks ! it tends to give bigger boobs to character's that doesn't have.. but need to test further !

@Darksidewalker That nudge actually help. I turned on LightSpeed/Distillation and got a coherent image out. I'll cut if of and try a higher (than 1.0) CFG with 20+ steps and see what I get. It may just be that my copy of SolsticeCoin was distilled and I didn't realize it. Thanks for taking the time to reply.

@Darksidewalker dunno how but people did it for the sulphur and eros models, again of course bf16 is needed which you would have.

@Darksidewalker Thank you for your reply! I was also confused. Normally, it shouldn't throw an error; there shouldn't be any operation that could cause it to malfunction, it's just selecting an option.

I downloaded all the VAE files through the Model links in the workflow, and I also installed KJNodes, and downloaded the plugins in the Requirements.

@Darksidewalker If the downloaded file is fine, could the problem be with the KJNodes version?

@Darksidewalker The problem has been found; it's caused by the version.

you underestimate the fact ltx2.3 makes videos 2x longer, 24fps native and is 50% faster than wan 2.2, wan is better in the smaller actions.

@Darksidewalker Thanks. I ll try them out. You're making LTX reach WAN 2.2 levels. Appreciate all your work with both models.

I use ZiT MoodyPro model v10 (v13 is ok but less natural penises overall)

@TheLastRemain Thankyou for the suggestion. I ll look into it. I m using Klein ovaNS W , for realism its quite good & fast. Might have to work with the prompt a little for accurate generation.

this model isnt great at it, but then again ltx 2.3 base thinks cum is pop rock candy for some reason lol

there's a cum lora, use that, its realistic, but lacks the ''shot'' aspect of it as far as im aware

Sulphur is quite overfitted.

I would not recommend sulphur2 over 10Eros.

@Darksidewalker Thanks for clearing that up!

@Darksidewalker Ah the abuse fears alas ... lots of normal use cases like taking off the characters coat, shot at the laundro-mat taking off the sweater to add it to the wash etc... Ok, got to look for some proper prompting then and see what we can cook up.

is it the model or the lora, or is it your prompting? as i have no problem inpainting nude characters through prompting alone on the base model/no lora for it.

if you prompt like ''the woman takes her shirt off, the woman becomes shirtless'' etc, then ltx often doesnt take the hint, its the same when prompting for a blowjob, wont work when you prompt ''the woman performs oral sex on the penis'' 9 out of 10 times.

you have to guide them in steps, like - the input woman with brown hair wearing a green jacket raises her hands towards the the silver zipper of her green jacket, she uses her fingers to pinch the zipper - she pulls the zipper the zipper down that she has inbetween her fingers to unzip her coat, when the coat is open, she grabs the left side of her open coat with her left fingers and the right side with her right fingers, she takes the coat off and lets it fall to the ground.

stuff like that works fine for me, i keep it superficial like she uses her hands to unzip her coat, if that doesnt work, i go more in depth and so on.

but there's other factors aswell. if you use prompt enhance and it censors your prompt it can botch the generation.

aswell as when you use distilled cfg 1 your prompting gets impacted by it

and what also matters is sigma/steps, scheduler, etc, its a complicated mess as im a novice but ive become aware that scheduler/sampler really matters on what your trying to do, it can be totally different if you want to maintain your input image or alter it allot since with one you dont want allot of noise at the start compared to the other

Bro, I've hardly seen any behavior LoRAs related to taking off clothes — maybe the creators also think this action can be left to chance

@anth0nx Thanks for the tips, will try them out and see how it goes.

highly recommend this lora https://huggingface.co/Kijai/LTX2.3_comfy/blob/main/loras/LTX-2.3-OmniNFT-RL-Lora_bf16.safetensors

it can help a lot with sounds.

@Darksidewalker fp4 only if you have 50 series rtx