HuggingFace backup - BF16 versions.
INT8 versions - INT8 workflows Use 1.4_DMD_convrot with the int8 V5 workflow.
Ver 1.4
IMPORTANT: VERY DIFFERENT MODEL
Any prompting tricks and image-prompt combos from 1.0 and 1.2 will transfer poorly to 1.4.
1.4 Is more closely aligned with the original 0.3 beta model than previous versions. There's a lot of reasons but mainly based on the issues with the model's negative tendency in 1.0 and 1.2 was only solvable by fully reshaping and remaking it. I have other creators piggybacking off of my DMD and merge work for their own gain, so I won't go into details, but the 1.0 heritage is still fully inside this version's soul, but not in it's actual behaviors. It is also fully designed for use with the DMD lora I attached.
FP8 certainly has some accuracy loss and will blend limbs and details. BF16 is recommended if you have the setup or the int8 models on huggingface.
This model has near-0 anatomy depiction currently.
It fully understands the concepts, but my new merge method is a 3 stage comparison shape-level reform lever that mostly aligns to the base 22b model instead of the usual merging methods. This is essentially an abliterated base LTX 2.3 model. Most prompts will output near 1 for 1 with the base 22b model, but as you can see in the previews, it still knows some tricks. Going forward I will personally attend to it's anatomy issues with my own training, bypassing the issues with letting sulphur control that part. For now if you want to have anatomy in generations, use Loras. All Loras will work, with base model loras being even more effective. This is also a much better training base and may be the recommended training base for the rest of the 2.3 versions.
#1 usage with this is prompting. It is very difficult and will be confusing at first. Scrutinize the prompts in the previews I used and notice how almost everything is plainly instructed and the prompt is formatted like a scene script. All were refined with multiple seed rerolls and refinement, it has the same approach as the base model. That specific prompting strategy can be done by using this system instruct on Grok or and uncensored LLM:
You are an expert at creating short cinematic video prompts from a single attached reference image.
When the user attaches an image and asks for "one for this image" or similar, generate a response using this exact format and style:
Use the provided start image exactly as the first frame. [One-sentence description of the overall cinematic style and camera movement]. [Short description of the environment, lighting, and atmosphere].
Performance: [Detailed but concise performance notes for the main character(s), including body language, facial expressions, emotions, and specific actions].
[If dialogue is appropriate, add a Dialogue section with character names and delivery style.]
Keep the acting natural and cinematic. Small pauses, micro-expressions, realistic movement, and subtle environmental details. No exaggerated motion, no slapstick, no extra characters unless clearly visible in the image.
Keep the entire response to 4-8 sentences maximum. Focus on a simple, logical, and interesting motion evolution that flows naturally from the starting pose in the image. Use strong verbs and cinematic language.
It knows nothing new v.s. the older versions, but what it can do is approach prompts much more intelligently than them and is much less inclined towards explicit generations. This is mainly the one for solo usage and clothed or unique situations v.s. just animating porn clips, although it can still do that with more complex prompting and camera work.
Ver. 1.2
Additional adjustments. Basically a stronger version of v1. This may be the best LTX2.3 can do tbh. It has a bit more tensors. This FP8 is larger, I left all connectors on exclude because that was one small part of identity drift. Full bf16 for quants and training is on HuggingFace backup. It can do anything v1 can with more 'stability' -- less tendency to do errant or odd moves and reaching more stable outputs at lower steps (8-10). Still best used to just animate sex images. Complex storytelling and prompts still requires additional guided frames and high effort, everyone knows LTX 2.3 is very clunky by now. Lighttricks seems to be 4 years behind everything, they just now decided to implement experts on the next version. I guess the totally true-proven method of autoregressive is too futuristic for them (literally the best method for video models - Seedance2, Grok, Kling).
This has slightly better initial character and likeness keep, but prompts can remove it unless care is taken. Expressions, dialogue, movement, and camera/pose change will all attack the guided character's look. Best to always train the likeness as a lora or have a guided endframe to give the model a future target.
Same bad tendencies in certain scenarios, same unrefined concepts.
Ver 1.0
This is not the official Sulphur2 model, that is also here. They are the team that trained and manage the official model. I am only affiliated as an i2v tester and consultant. This is my personal merge of the actual data and it uses different training steps and layer scaling to create a consistent flexible I2V focused version. Future versions (v2+) will be actually fine tuned by me personally aimed at filling in gaps and weak concepts.
Quality is entirely prompt, image, and workflow dependent. Aside for systemic issues with LTX: motion smear, artifacts, and identity loss, it still behaves like 2.3 requiring extensive prompt refinement, extra sampling, and resolution targeting for best results.
This probably means nothing legally, but I just want to make it clear that you waive any right to blame or use the host site, me, or sulphur in your defense if you download and misuse this. It's ethical artistic intended use is clearly stated:
By downloading, accessing, or using the 10Eros Model (the “Model”), you (“User”) agree to the following terms:
1. Permitted Use Only The Model may only be used with inputs of fictional or non-real characters. Any use involving real persons, celebrities, public figures, or identifiable individuals is strictly prohibited (“Out-of-Scope Use”).
2. Disclaimer of Liability TenStrip (the creator of the Model), together with its affiliates (CivitAI.com / CivitAI.red, Lightricks Ltd., and Sulphur Team), collectively and individually disclaim all liability for any direct, indirect, incidental, consequential, special, or punitive damages arising from your use or misuse of the Model.
This includes, without limitation, any claims related to copyright infringement, right of publicity violations, defamation, privacy violations, deepfakes, illegal activities, or any harm caused by outputs generated by the Model.
3. Out-of-Scope Uses Forbidden All Out-of-Scope Uses are explicitly forbidden. The creator and affiliates bear no responsibility for any outputs generated from prohibited inputs. Your actions with this generative Model are solely your own.
4. Assumption of Risk You assume all risks associated with the use of the Model. Any use outside the permitted fictional/non-real character inputs further releases TenStrip and all affiliates from any and all liability.
5. Indemnification You agree to indemnify, defend, and hold harmless TenStrip and its affiliates from any and all claims, losses, liabilities, damages, and expenses (including reasonable attorneys’ fees) arising from your use or misuse of the Model.
6. No Warranties The Model is provided “AS IS” with no warranties of any kind, express or implied.
This agreement is governed by the laws of the United States. If you do not agree to these terms, do not download or use the Model.
Description
DMD optimized, base model aligned, advanced shape level merge.
FAQ
Comments (151)
v1.4??
Waiting for the non functional site to clear everything.
Hi, I'm using the v5 workflow in stock config, it seems to make my anime girls look a lot like real ones. Like, it kinda changes they're style a little too much. I'm fairly new, what can I do to make it preserve the style a little better.
It does this with or without Loras.
I believe it may have something to do with the upscale process. The preview 1 holds the style better but looks worse in terms of smooth animation and playback.
Be sure to prompt the style fully, anime cartoon line art, etc. Lower preprocessing even all the way to 0 if needed both passes and avoid too many descriptions in the prompt until that goes away. It's also possible the DMD wasn't distilled on a single cartoon or lineart latent and can't do it well.
@tenstrip What setting exactly do you mean when you say to lower all the way to 0 if I have to? Im fairly new to the scene. Thanks again.
@shelbyeldin912 It's set at 35 which is pretty high, so change it to 0 which is basically off. There's no compression which means the model won't attempt much of it's own interpretation, which means there's also less movement freedom. Some of my other workflows have additional constraints from my nodes that keep character and also style from the input.
@tenstrip Oh yes yes, your talking about motion strength, it does help. I'm going to browse your other workflows, wanna see how close I can keep my characters. Oh, and thanks for this amazing product. You're doing the lords work.
I have a problem with eyes on workflow 4. Did anybody face this? I use medium quality
1.3?
Not really worth uploading or using. That was just a dev step.
For people who still just want to use 1.2 and have a grasp on it I'm working on and uploading a DMD baked 1.2 version for that which benefits from the newer improvements.
"or the int8 models on huggingface." would the int8 model version be found alongside the others? not seeing it atm, but this is a fresh release so I totally get it!
Since this seems to be the most "pure" version we can hope for, are you intending to offer a LoRA for your custom training?
The loras I train that are concept outputs are updates to my oral suite and sexy moves but for this version. The big underlying mini finetune I'd rerank and put out too sure.
@tenstrip To ask my question another way: Are you imagining further 10eros releases (if any) proceeding as a general NSFW checkpoint akin to sulpher with a bunch of additional custom data merged or are you planning to "keep it clean" and rely on concept LoRA's in "the old ways"?
@boobkake22 Yeah this is as much sulphur as you want, I generalized the model enough with it but the sulphur data right now was undertrained and polluted with bad prompts and bad data which ruins the model's prompting and attention layers if you merge it. This fixes that pretty much but you lose anatomy and some finer details are harder to bring out. Any other concepts can be loras and the anatomy is easy to train.
I have a dumb question, sorry. I've been using your other versions (great btw) and using t2v and i2v. They work great. Now, you said this uses a different prompting style; you mentioned in the notes and in the images as well, "Use the provided start image exactly as the first frame." I wanted to clarify if this means when doing t2i, it WILL use the provided frame in a t2v flow or not? It seemed kind of confusing because your 1.2 works great, but I'm not sure if that applies to both t2v and i2v, obviosly with i2v, it does, but if you add that line, does it use that frame as a starting point? I mean, now that I think about it, that's basically i2v. Sheesh, I'm confusing myself. So don't use it for t2v then?
I've never even tested T2V in a single model. I just assume it works. For T2V already prompting is completely different because you waste a ton of tokens having to depict the image first. But that part of the prompt is an I2V only thing and one I recently discovered.
@tenstrip Ah, I see. ok its just because the way you worded it seemed like you weren't even taking t2v into consideration with your model. I use your model 90% of the time as t2v, and it works amazingly, i'm using the full bf16 version you have on git (or hugs, forget which). So that's the only reason it threw me for a loop. It explains why there was no mention of t2v at all in the description. ok thanks!
i would like to ask is loading any lora to this model a bad idea? i have a whole 5090 and it crashes it lol/ exmaple would be NSFW Body Physics Fluid Motion Enhancer V1 and LTX 2 & 2.3 [I2V] NSFW (+Furry) - Multi Purpose. it freezes my program and then i get memory issues
Check your memory page file settings or the comfyui memory handling. I have the same card and merged and ran bf16 with lots of lora stacks.
Eros works quite well with furry and without additional lora
@tenstrip im having issues as well where the program is downlaoding a distilled-lora every generation i take, im unsure why any of this isant working
@mrsanders1313840was working for me 2 nights ago now all of sudden i have memroy overloads running wan2gp
@mrsanders1313840 loading the furry lora gives me this "The generation of the video has encountered an error: it is likely that you have unsufficient RAM and / or Reserved RAM allocation should be reduced using 'perc_reserved_mem_max' or using a different Profile".
wait does this have a lightning lora already baked in?
@EURLO Not this version. You can use it distilled or DMD whichever, baking it in takes a lot of usability out.
@EURLO I've never even heard of that error. Go to google AI or claude or grok and start copy pasting stuff and give it to them, they're pretty good at general debugging. If it's only for that lora and no other ones, delete it and redownload it could be corrupt.
@EURLO You have a workflow with a loader that is pointing to get a file off huggingface instead of loading it from the device. I think with all your problems you may consider trying another workflow.
@tenstrip ok so i was able to just now run V1 of this model however im not sure if it needs a lightning dependent. Im getting distortions but it runs when i use the euler sampler on wan2gp
so i checked the links i see there are other versions of the same model, i treid out 10Eros_v1.4_fp8mixed_learned.safetensors. so it runs very smooth on my gpu, my thing is how were there instructions how to not get blurry watered out renders that have sloppy pixels?
@EURLO You need to run the DMD lora, its in optional files, or with the LTX distilled lora.
@tenstrip using distilled lora is exactly whats causing my pc to freeze up not sure why. i will try the other option see what happens
@tenstrip with wan2gp its all plug and play, i dont exactly have a workflow
@tenstrip can i ask which text encoder was used for this model?
We need quantization! GGUF!! GGUF!! GGUF!!
Your work is amazing, it works as it is out of the box :)
I agree, the poor people (like me) need to goon too!!!
With in4 support now in comfyui that might be one to use. I linked the lora extractions, they're a light way to use it offloaded on top of a quantized LTX checkpoint, but not really the same are running the model itself.
@mrsanders1313840 @tenstrip I created a Q4_K_M version for my needs and since I'm trying to be a nice dude I'm sharing it here (all credits go to tenstrip, I haven't done anything special honestly):10Eros_v1.4-Q4_K_M-cfg.gguf
And here are more quant versions from another user with more options: https://huggingface.co/vantagewithai/LTX2.3-10Eros-1.4-GGUF/tree/main
@SpazAI thanks, I already found it :)
Hi, firstly thank you for you work :) There is something i don't understand.I'm using the convrot version and i'm unable to do any nsfw content, should i need to use a separate lora even for basic nsfw ? Which one ? Thank you
Thank you for all of your hard work. I wanted to ask why the DMD workflow doesn't use RTX Super Resolution while the non-DMD workflow does
I think the DMD outputs were much more clear so I removed it, a lot of people also can't run and install that node. I still use it personally and did on the previews. Just wire it back in at the end.
@tenstrip Thanks for the response, just wanted to be sure there wasn't a technical reason or something before I tried adding it back
I love this update. I get much less hallucination and weird limbs than with 1.2. Thanks for sharing your work!
EDIT: Oh and if you do not mind me asking... I saw JoyAI lora in your workflow and their hugging looks interesting/promising. I wonder what your thoughts are on this and if you recommend using it all the time and if so on wich value.
So far my experiments did not yield clear results if it is better with or without (mostly kept it on 0.5), but I bet you had much more experiments than me and I would like to hear your thoughts.
Others who tried it are welcome to answer too of course.
Mate, hats off to you. First time I'm actually happy with my LTX gens.
Lots of playing around with settings and refining my prompts, but V1.4 really made it sing (running the INT8 variant). Godspeed
Hey, I hope you dont mide me asking. How do you run the INT8 version? I downloaded the 10Eros_v1.4_Bob_DMD_INT8_convrot and the v5 workflow but get a VAE loading error. Should the VAE be included in the checkpoint or not then?
Iam new to ltx.
@Tanha These are my workflows with int8 loaders https://huggingface.co/tech77/Workflows/tree/main
Dude.
The first tests from my part are insane. This could very well be a big leap forward for LTX. I was almost out of hope, when you stated in V1.2 that you almost got all out of LTX. And now you come around the corner with this one <3.
FYI: Make sure to download 10ero/tenstrips DMD lora. Took me a while to debug that.
Good model btw, good upgrade.
Getting good now, great work man!
Most people have forgotten about this one or moved on from it, but on 1.4 with DMD the newest one is actually a highly functional clean NSFW lora.
I'd agree his Lora is really good and a OG
Yet another excellent update. v1.4 feels like a big step forward, the model now follows prompts much more naturally and accurately compared to previous versions.
Anatomy, movements, and skin details are better handled, though still not perfect (typical LTX limitations). Overall it's significantly more enjoyable and easier to work with. Solid upgrade!
can you please try this for me?
a woman never changes her facial expression nor blinks nor moans, her face stays the same through the entire scene.
see if it works, i know i'm asking this out of nowhere but i've been struggling to make eros do this prompt, please let me know if it works or even better post it if possible.
you'll be doing me a huge favor, thanks.
@necroryona I uploaded some videos on my profile, the last 2 are generated with v1.4
The workflow metadata are also available if you want more information in how i did it.
@DriftPixel thank you for providing the metadata and videos, please give the prompt i shared with you a try, if it works, it solves a huge problem with LTX2.3 in general.
excellent work btw.
@necroryona I assume that you're using "text to video" feature of LTX 2.3 and i can already tell that the outcome will be really bad just from the prompt alone. I don't need to try it out because i already know the outcome...
With LTX you need to be really specific in what is happening in the video if it's a "text to video" format. The more detailed the prompt is, the better the outcome is basically.
Here's a "better" prompt that will result in a way way better video:
"
A woman whose face remains completely frozen and unchanging throughout the entire scene and duration of the video: her facial expression is locked in place with zero variation, her eyes stay wide open without any blinking or eye movement, her eyelids never close or flutter, her mouth stays perfectly still with no lip movement, no parting, no smiling, no frowning, and no moaning or vocalizations of any kind. The exact same facial features, muscle positions, and expression hold rigid and consistent from the first frame to the last, with absolute temporal consistency and no micro-movements or morphing on the face.
"
And i mainly do "image to video" because it's way easier to animate since the base line already exists.
@necroryona Also you can utilise an uncensored LLM like Grok 4.5 to create an "better" prompt that the model can understand better.
Just feed Grok the instruction set found on the description of this page and it can create very usable prompt, i'm using it myself. (but be careful, it can hallucinate)
@DriftPixel thank you for taking the time to reply, i was using 10eros 1.4 (this one) and tried Winagguns too, both gave me similar results, and no i am using image to video, with the first frame of a woman with her eye open, all i wanted was to keep the same expression + eyes
and it won't let me for the life of me, i tried prompting really focusing on what i mentioned above and it never worked, the funny thing is any other prompt i tried worked except for this one, also i tried different text encoders just in case and got the same results.
for something so simple, i am surprised how it can't handle (not doing anything) lmao.
i'll keep trying although i've lost hope for now, it's a shame since the body movements of Eros10 are superb, it makes me kinda sad that i cannot enjoy it fully due to something so simple not coming up in my videos.
@DriftPixel also i promise you if i prompted the prompt you gave me it'll 100% make her close her eyes and sleep instead, i've tried countless times.
i'll give it another try tonight, i just need a simple workflow to make it happen rather than a complicated one, also i am using convrot int8 versions just in case, i've only got 12vrab and 16gb ram on me.
@necroryona This is out of my league... i need to call reinforcements: @tenstrip
If you want to hold or delay any change use [0-7 seconds] timestamps. "[0-5 seconds] She is motionless and asleep."
@tenstrip thanks, i'll try it now and give my feedback.
@tenstrip @DriftPixel
i tried both using 10 Eros v1.2 — LTX 2.3 Explicit Video Finetune https://huggingface.co/Winnougan/10Eros-INT8-Convrot
didn't work, the girl does random facial movements and blinking, it's as if the generated video is prompting itself..
using prompt relay timeline.
@necroryona 如果你想做这种人物不动的，我建议你用wan2.2+timestop lora，效果会更好，也更简单。
@yazhoulove can you link me the lora and the wan model ?
thank you for letting me know !
@necroryona 这是lora链接：https://civitai.red/models/2437696/timefreezes?modelVersionId=2743366。同时用高噪和低噪模型，效果很好。wan2.2模型千万别用smooth版本，它有更高的动态，会让你的人物冻结后依然控制不住的眨眼。使用官方版或者remix版本都可以。事实上我当初想做一个日本AV“时间停止器”系列的剧情。用了这个lora之后成功了。
@yazhoulove thank you so much for providing me with the lora, but i still need a workflow to make it work, i'm sorry i'm new to Wan2.2, can you please give me a good workflow to make this work?
@necroryona 你直接把示例的视频点击一下，然后下载下来，再把下载的mp4文件拖进comfyui就可以看到工作流了。civitai大部分的图片和视频示例都自带工作流。
i've just tested it a bit but so far v4 is so so good
The way the introduction sounded, it made it seem like a downgrade and like it couldn't do the things the previous one did. Am I missing something?
@roboticone it listens better to what u want it to do without forcing someone funking all the time and it still can fuck, exactly what i've been looking for
example: he lifts her skirt > he just fucking lifts her skirt
before > he lifts her skirt and some cock appears to fuck her (which is cool but not what i asked for)
@roboticone Not a downgrade but a major change. With 1.2 you can add explicit images and just prompt "futa sex" and you will get a decent output. This one is way more compromised to be all around which means everything require prompting and refinement, explicit or not.
1.4 is superb! In my opinion responds better to prompts.
It should. LTX responds poorly to lora merges because of some internal adjustment it does when the model is trained that has huge effect on attn_2 layers which is mainly prompt. I isolated how to severely dampen those changes it does with a remediation step, which imo are somewhat of an issue with the model itself that should be looked at by Lightricks around high rank lora training. Version 1.3 was a full adjustment, then merge-tuned to 1.4 to do a bit more than that one did as a blueprint demo of what data adjustments I need to train into.
I've tried to plug this into my LTX2.3 Director workflow. All my videos come out fuzzy. Any ideas?
Use the DMD lora or distilled lora.
1.4 officially beats wan 2.2
thank you
Personal choice: v1.4 is lighter and faster compared to the v1.2 and it’s easier to control.
Speed: for 12s video I can create it with v1.4 around 1436s but v1.2 takes 1800s on my 3060 12GB with my advance prompt relay workflow.
Control negativity: The v1.2 sometimes straight up ignores my negative prompt like “cums on body” that’s why it’s harder to make a good “creampie” video before. It needs more trial and error.
I used the turbo version of workflow, 10 seconds in 640x832 resolution are created in about 135 seconds. RTX 4080, on a 4070 card it takes about twice as long, do you use sage attention? In my experience, the acceleration is about one and a half times.
Using the cachedit LTX-2 node gives another gain of about 30%
@mrsanders1313840 Yeah with sage attention. Try to use the workflow that I attached on my latest post “reze” if you want to compare it precisely. It’s from Kenpechi. I use 12 reference images, on 720p, 12s generation, and some Loras.
I have a 3060 12 vram too, did you used the GGUF version? If you did Wich one? Or you used the safe tensor version?
@kibed52314943 fp8 not gguf
@Marau Hi! Sorry if this is a noob question. I've only used Stable Diffusion XL with AUTOMATIC1111 so far. I have an RTX 5060 Ti 16GB. Would you happen to have a workflow you could share? Also, can the FP8 model run on this GPU?
@Dardranight Kenpechi FLF workflow v1.2 try it
The quantized version is here, my choice is Q6.
https://huggingface.co/vantagewithai/LTX2.3-10Eros-1.4-GGUF/tree/main
do you have an a workflow that you can share?
@Dardranight https://cloud.mail.ru/public/LX2Z/GEA9C5L6q here with all the folders
Wow! What a huge difference! Thanks for this magical model. I think the 1.4 version is on par with the Wan 2.2, only it also has sound.
And it's faster!
I agree with others that it's better than WAN2.2 for NSFW now. I'm testing it and it's so much easier to make scenes than WAN2.2, especially for longer scenes (10+ sec). WAN2.2 have to rely on SVI wich is very hard to get constant motions trough the scene with loras activated. I did not get ghost artefact on the 10+ generation I made too, and every one of them were stable without obvious deformation or artefacts.
Testing specs : INT8_convrot version on RTX5090 Inference specs for a 15sec video 1024x1376 : 42 sec for 1st pass and 2m24 for 2sd pass/upscale, total ~240sec for end-to-end workflow.
Congrats ! :)
Are you getting unrequested transitions that end up being distorted mutated people in mid parts of your clip?
@roboticone usually a symptom of bad prompting, misdirecting it with bad camera prompts, using distilled lora. You don't want to every try a transition without a guided frame.
Great work! Which uncensored LLM are you using to generate your prompts?
I am an NSFW expert.
v1.4 is far superior to WAN 2.2.
To be fair the Wan comparisons are flawed. Wan has received thousands of well trained loras. Base Wan2.2 is incapable of anything people associate with the model nowadays usually, all of it's modern quality comes from massive amounts of lora data merged together. LTX it still incredibly underserved with Loras.
@tenstrip Your honest assessment makes me trust it even more. Actually, there have been attempts to merge many LORAs into LTX 2.3. The results have been quite positive. Here’s the link below.
https://civitai.red/models/2543443/dasiwa-ltx-23?modelVersionId=2967331
i havent done a single video generation locally can someone plz tell me can i run a single one of these models on a 12gb rtx 3060 with 32gb ram
If you use comfyui, download one of my videos and import it, I use ltx director, in my profile there is the guide and links to the videos to use it.
LTX2.3 is a rough model to start on. Very nonsensical and with a lot of limitations, you only get better use out of it when you get the feel for what it can't do more than what it can.
@tenstrip you just need to switch the checkpoint
when download completes it just resets and never downloads? anyone know why?
Google chrome has had this issue recently.
Happens time to time, the # verification probably failed due to a corrupted file, so it starts again. Eventually it will get it right, it's common when downloading big files like this.
try use Edge, it will download full file
1.4 is crazy good.
i use this model,V5 workflow, DMD lora, and distill lora.
it's worked, but quality looks like xhit, i don't know why.
any ideas?forks?
disable DMD lora and use the 1.1 distill lora at 0.6 or 0.5 strength dmd lora didnt work for me in wan2gp but with normal distill lora it works great
@MisticRain69 fix my broken workflow, and my broken heart. You are amazing
Any suggestions on getting good porn t2v from 1.4?
Shouldn't technically be possible without loras, stack a bunch of loras and use the Sulphur lora with them.
Better Female Nudity 2.0 LoRA
Omni RL NFT LoRA to bring out some additional 'shine' to the output.
I find Sulphur to be more restrictive compared to 10Eros.
That's pretty much my stack, when using the 1.1 distilled mxfp8 model from Kijai.
Is it just my imagination? or this version has improved character consistency? great work.
No it's been a driving factor since LTX2.0 came out. Prompt/likeness were paramount with this revision and also cutting down on the errant sulphur issues. I sacrificed all the detailed explicit data for just a loose generalization, but it's gonna end up better since it can be target trained and wrapped in additively with my new merge method or with loras. Also 1.2 was too far gone the embedded issues in it prevented using it to continue versions.
In wan2gp at least not sure about comfy, use the 1.1 distill lora. I have tried the DMD lora and it fries the output quality. Not a issue with his DMD lora just wan2gp is not setup to properly use the lora. Also using 30 steps, 1.1 distill lora 0.5 seems to give very good quality results. Just dont try using a CFG above 1. I have tried The standard dev model settings and it fries the outputs so likely the DMD merge is conditioned on a 1.0 CFG. Not a problem, 10eros 1.4 is a very good model and trains well in musubi tuner. 5 epochs in on a 1270 NSFW video run and it looks pretty good so far.
If you can dm me I'd be interested in any musubi templates you find for it I have a ton of data to run through it and am going to try rank 4 and 8 LoKr on anatomy concepts. DMD is temperamental with sampler and sampling steps, if wan2gp only does the default LTX 8 step sigmas with no way to change it it will not sample fully first pass and it might be forcing another distilled on top of it. You might be better running the DMD merged into the model first and increasing steps to 9 or 10.
Thank you.You were right‼ The low-quality, garbled text overlays produced with DMD-LoRA disappeared completely once I switched to ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors.
1.4 + DMD_reshaped_r256 using workflow from huggingface and always getting weird hands. Am I getting only bad seeds or what? Ltx with 1.1 distill lora improved hands a lot and it was a couple months ago. This is not a criticism btw just a question
Some things I have learned.
1: never use a distilled 1.1 lora. The audio distortion is terrible. Use a 1.0
2: prompting she says "mmm, mmm, mmm," while sucking a dick works super well.
3: prompting she moans in orgasm or she breathes heavily is very helpful.
Don't use euler as your sampler. That one is bad for audio output. Stick to res_2s. You can clean it up with the Omni RL NFT LoRA as well.
Great work.
Quality has definitely improved compared to the previous version. I’ll have to test whether it also works better as a training base than the LTX-base model.
Might be a rookie error but could use some guidance. I prefer using Rune's workflows - 1.4 fp8 works great, slots right in and works a charm. Trying to use the convrot version though (with or without the associated DMD lora and it works but completely ignores the reference image and bears no resemblance.
Actually I think I get it - it's because the convrot is just a diffusion model not a checkpoint?
@lonlon1990 Long story short there are different methods to create int8 convrot models and: 22Gb version is diffusion model: transformer-only and need seprete vae, audio vae and text files. The 29Gb version is a checkpoint worked alone.
@flo11ok874 Yeah got it working with the checkpoint version. Seems a bit odd compared to the FP8 though - occasionally produces something insanely good but mostly just seems blurrier and less prompt adherence than the FP8. Surely there shouldn't be that big a gap between the two?
Make sure you recache text embeddings and latents if you are using the same cache from a different LTX model like sulphur when you train the new eros. Used a old cache and a lora came out completely fried lol.
why hard to make dick/penis , even using lora
Relax, he draws a disgusting pussy too :)
Hello, why are all the voices female in the LTX2.3 10Eros-v1.4 version? Even when I explicitly specify a male voice and use the DMD full rank LoRA? The workflow I am using is LTX23LTXDirector2.
Hello, Can I use Ltx2.3 1.1Distill + 10Erosv1.4 Lora with standard LTX workflow rather than using 10Erosv1.4 workflow with DMD lora？Is it the same effect?
You can use it, the effect is worse. I am using eros 1.4+ https://civitai.red/models/2781695/ltx-23-dmddistilled-lora dmd quality 0.5 and Distilled Rank 72 energy quality 0.5. In my opinion, the best option
Noob question: I tried using this checkpoint by substituting the fp8 version to ComfyUI's built-in LTX2.3 img2video (fp8) workflow, but it gives me an error with the "Load LTXV Audio VAE" node. (It receives this model as checkpoint parameter.) What am I missing?
Yeah it was a noob question, my downloaded checkpoint was corrupted. It works flawlessly.
Hi Tenstrip.
Can I ask you a few questions regarding your recent update?
1. Is below model DMD baked-in?10Eros_v1.4_DMD_int8_convrot.safetensors
2. I see you loaded the int8 convrot text encoder as well. i read the model card and couldn't find any nsfw tag (abliterated, heretic). Did you already confirm if it works well with your model?
Thanks.
Noob question, but how can you do negative prompts? I dont see a place.
LTX2 NAG by kjnodes. Put it right behind the CFGGuider. You can plug your neg prompt into both slots or create 2 seperat Conditionings for audio & video. The output will be your new model. It's not as reliant as like a SD neg prompt, but it helps.
How do I prevent getting subtitles? Sometimes, the video created contains subtitles of the spoken words and sometimes it does not.
Haven't figured out what triggers it and how to prevent it from happening...
negative prompt maybe?
Line kj node nag right before sampler model input. Negative prompt is option. Also avoidance of those terms. dont use "" just use x whispers: plaaplaaplaa.
[ERROR] !!! Exception during processing !!! Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor
[ERROR] Traceback (most recent call last):
File "D:\ComfyUI\execution.py", line 543, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
How can i fix it ?
Great work! I'm unsure why the second video generation is coming out as like a weird graph video but the first video in vhs combine is working. just no reason to upscale because the output is like a 'mesh' instead of the video itself.
does it support flf2v? FLF2V not working for me.
ya it works for me. Make sure you got the DMD lora (at 1 by default)
my criticism, it is very eager to show nipples and genitals (though I could NAG/neg-prompt it out), and for lesbians, it's near impossible to get breast fondeling. It 99% will generate 1 girl doing terrible groping, and the 2nd girl start grabing her own breasts.
runtimeError: shape '[2048, 2048]' is invalid for input of size 3039426
help me please lol