LTX2.3 10Eros - CivArchive (CivitAI Archive)

LTX2.3 10Eros - v1

NSFW

kofi - Training fund and Tips (also straight into training fund), the yellow buzz doesn't do much for me.

HuggingFace backup - BF16 versions.

Lora Extractions.

DMD full rank Lora.

INT8 versions - INT8 workflows Use 1.4_DMD_convrot with the int8 V5 workflow.

PlagueKind Workflows.

My Workflows. V5 DMD recommended with preprocessing strength adjustments to control drift and motion.

10S Nodes (also on manager)

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

Quantized versions.

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

FAQ

Comments (6)

DurolithApr 28, 2026

CivitAI

is distilled dynamic rank111 lora needed?

tenstrip

Author

Apr 29, 2026

The reranked distilled loras are much better for i2v yeah.

ddng533Apr 29, 2026

CivitAI

Could you recommend a workflow that would yield the best results using this checkpoint based on the H100 or RTX 5090? Since updating Comfy UI, the EROS workflow I was previously using is not working due to node issues.

tenstrip

Author

Apr 29, 2026

Yeah hold that thought for the new version and workflows. I don't have much experience with the beta version actually. With loras the Mr.Xin ones work well with that compute you can get rid of the optimizations. The full bf16 version of the new one is what you want on 5090.

premiumfp2026Apr 29, 2026

CivitAI

Did you uploaded the strip lora yet?

tenstrip

Author

Apr 30, 2026

Wasn't mine. PlaytimeAI has his version on his huggingface though. I have a "clothing physics" helper on my own train list and collected a lot of good data probably but it won't be on here.

Checkpoint

LTXV 2.3

by tenstrip

Download (Beta) View on CivitAI

base model

Details

Downloads

22,520

Platform

CivitAI

Platform Status

Available

Created

4/25/2026

Updated

7/11/2026

Deleted

Files

LTX23_audio_vae_bf16.safetensors

Size:

347.95 MB

SHA256:

5bc10fa4adecf99dda132d916e23048cbd56797702c5fa50eb5d2079048a38c3

Mirrors

HuggingFace (68 mirrors)

LTX23_audio_vae_bf16.safetensors

ltx_23_audio_vae_bf16.safetensors

LTX23_audio_vae_bf16.safetensors

CivitAI (2 mirrors)

LTX23_audio_vae_bf16.safetensors

ltx23FP4_ltx23AudioVae.safetensors

ModelScope CN (2 mirrors)

LTX23_audio_vae_bf16.safetensors

ltx2310eros_v1_txt.safetensors

Size:

2.15 GB

SHA256:

911d59bb4cb7708179c9a0045ea0fe41212ecfb77aed3a02702b7c0a8274911f