<h1 id="ltx2.3-fp4-models">LTX2.3 FP4 Models</h1><h3 id="because-no-one-uploads-fp4-here.-and-they-should-(fp4-works-on-3xxx-and-above-for-vram-savings.)">Because no one uploads FP4 here. and they should (fp4 works on 3xxx and above for vram savings.)</h3>LTX2.3 Distilled FP4ME - Distilled FP4 Mixed Extreme - 14.1GB** LTX2.3 Full Dev FP4ME - Full Dev FP4 Mixed Extreme - 14.1GB** LTX2.3 Official NVFP4 - From Lightricks - Updated 3/17/2026 - 21GB LTX2.3 Dev NVFP4 - from <a target="_blank" rel="ugc" href="https://huggingface.co/Hippotes">Hippotes / </a><a target="_blank" rel="ugc" href="https://huggingface.co/Hippotes/LTX-2.3-various-formats">LTX-2.3-various-formats </a>- 18.15GB LTX2.3 Dev FP4 STFO - from Kijai's Transformers only Scaled FP8 - 16.63GB** LTX2.3 Distilled FP4 - My first FP4 - Transformers Only - 18.15GB** ** Are transformers only, or I broke vae/text projection :) needs seperate Vae and Text projection downloads.I have a fix for lora issue on FP4ME models. Comfy Bathroom, custom lora loader with presets as well as custom config. <a target="_blank" rel="ugc" href="https://huggingface.co/MrReclusive/LTX-2.3-FP4">https://huggingface.co/MrReclusive/LTX-2.3-FP4</a>Diffusion model loader/unet loader seems to get confused with these, use the checkpoint loader even though it has no clip/vae

LTX2.3 FP4

ltx23FP4_ltx23AudioVae.safetensors

LTX2.3 Audio Vae

LTX2.3 Official Dev NVFP4

LTX2.3 Full Dev FP4ME

LTX2.3 Dist FP4

LTX2.3 Distilled FP4ME

LTX2.3 Video Vae

LTX2.3 Text Projection

LTX2.3 Dev FP4 STFO

LTX2.3 Dev NVFP4

<a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3-10Eros/tree/main">HuggingFace backup</a>.<a target="_blank" rel="ugc" href="https://huggingface.co/vantagewithai/LTX2.3-10Eros-GGUF/tree/main">Quantized versions.</a><a target="_blank" rel="ugc" href="https://github.com/TenStrip/10S-Comfy-nodes">10S Nodes in my Workflow.</a>This is not the official <a target="_blank" rel="ugc" href="https://civarchive.com/models/2594061/sulphur-2-base?modelVersionId=2914022">Sulphur2 model</a>, that is also <a target="_blank" rel="ugc" href="https://huggingface.co/SulphurAI/Sulphur-2-base">here</a>. 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.Both BF16 and Fp8 versions operate the same as beta. Drop them into any Eros beta workflow checkpoint and loaders. For difference, it is subtle. BF16 will always have greater motion accuracy and attention, but given the size the trade-off is questionable. If you want to scale or merge it for yourself you need to start on the BF16 version. If you are going to train a lora, do not train on this version (I think), train on the official <a target="_blank" rel="ugc" href="https://civarchive.com/models/2594061/sulphur-2-base?modelVersionId=2914022">Sulphur2 model</a> BF16.!!!!! The LTX2.3 Distilled LORAs are detrimental to your outputs. <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main">I worked on and fixed this problem before you guys experienced it by severely reranking and gutting the distilled loras</a>. There are many long complex reasons why, but simply put: it was most likely trained on T2V latents produced by the base model. Now, this is a finetune I2V producing model, that distillation has the opposite effect. You've probably noticed it many times and thought it was just the model. The 384 1.1 official version and even the other reranked versions reduce trained anatomy accuracy, reduce sexual motions or any trained motions not in the base model, and they add motion blur/choppy edges by disturbing feed-forward layers (that is a serious issue Lightricks needs to fix). They are also insanely overkill for what they need to do. My reranks are linked there, especially the cond_safe versions which have conditioning and feed-forward zeroed out letting them produce much better with sharper motion and upscales with a 1.0 first pass and 0.5 upscale pass, or 0.65 single pass.. Use those reranks with this model and Sulphur T2V and I2V... or don't read any of that and use the normal distilled lora and then leave a comment about all the issues it causes.<h2></h2><h2 id="workflows:">Workflows:</h2><h3 id="my-workflows-with-ltx-fixing-nodes-and-vertical-aspect-ratio-fix"><a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3-10Eros_Workflows">My Workflows with LTX-fixing nodes and vertical aspect ratio fix</a></h3><h3 id="eros-workflows"><a target="_blank" rel="ugc" href="https://civarchive.com/models/2488266/mrxin-ltx-23-i2v-eros-12gb-vram-and-32gb-ram-workflow?modelVersionId=2835183">Eros Workflows</a></h3><h3 id="easyflow"><a target="_blank" rel="ugc" href="https://civarchive.com/models/2516082/ltxv23-easyflow-gguf-or-safetensor-2-or-3-stage-run?modelVersionId=2864466">EasyFlow</a></h3>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 is not simple AIO lora merge, it uses no on-site loras; any 2.3 loras can be stacked on top of it and if you need them use them. It's a layer-scaled integration of 768-ranked data with a vibe-coded custom script. It fills in the nsfw and explicit motion gaps in the base 2.3 model without fully intruding on the 22b dev model's functionality and audio, at least an attempt to do so. New anatomy is depictable and introducible but subject to quality loss at low resolution and sampling steps and definitely not always fully accurate. All general nsfw and explicit ideas are within reach with proper prompting and sampling, some full form changes or scene changes may require additional conditioned frames for best results.LTX is still a combined T2V/I2V model. For I2V at inference, the model will shift towards drawing it's own learned subject matter through prompt and this will usually be most apparent in facial identity changes. For maximum identity preservation of unique faces unknown to the base model, a subject-specific LoRA trained (30 images, ~15 minutes) provides actual likeness data that any ID Lora or manipulation cannot match. This is currently the only fix for identity drift and something Lightricks needs to address or the model will never be useful as a studio tool.Common issues:<ol><li>distortion, smear during fast motion - <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main">Use these different distilled loras</a>.</li><li>Likeness change, deviation from start image - Try using edit-anything IC Lora for initial conditioning instead + default i2v conditioning on upscale pass to restore some initial image quality and detail. It does not need 1.0 conditioning strength for the first IC guide, try 0.6-0.75.</li><li>Artifacts, extra limbs, body contortion - is usually from conditioning or distilled lora strength which need to be tweaked up or down. I find actually that the model does not usually want to produce anatomy horror; it arises from a bad mix between prompt, distill, and conditioning leading it to do that.</li><li>Vertical upscale artifacts - this was supposedly fixed, it is not. Extreme vertical aspect ratios that upscale to 1280+ will have body morphing. Crop the input more square.</li><li>Bad quality/appearance, just terrible output - Make sure your workflow is working correctly. <a target="_blank" rel="ugc" href="https://huggingface.co/TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments/tree/main">Use these different distilled loras</a>.</li></ol>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 (<a target="_blank" rel="ugc" href="http://CivitAI.com">CivitAI.com</a> / <a target="_blank" rel="ugc" href="http://CivitAI.red">CivitAI.red</a>, 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.

Files with SHA256: 5bc10fa4adecf99dda132d916e23048cbd56797702c5fa50eb5d2079048a38c3