H3 LTX 2.3 I2V T2V Video Reasoning lora VBVR H3 VBVR Pro

This was the first VBVR lora trained for LTX 2.3 and H3 You can make money with my models. Sell what you generate, use it in paid work, run it on paid services, merge it, sell the merge all of it, no permission needed. Credit is appreciated but never required. New releases might have a short early access window behind Buzz, then go free and free is permanent. Once a model is free, I will never re-paywall it or pull it for a paid platform, and you're welcome to mirror, archive, and rehost it so it stays free no matter what happens to any one site. Feel free to mirror any of my loras to HuggingFace The H3 VBVR Pro version was trained on 1k videos from the VBVR Pro SFT dataset and with a rank of 16 instead of 8. You should see a noticeable improvement vs the previous version. Experiment with the strength of the lora since everyone's workflow and lora stack differs. I used a strength of 1.0 for the example video. If you use it with 10eros max beta 5 drop the strength to 0.5 to begin with or else the effect is too strong. The H3 version was trained on around ~400 videos and It seems H3 does not need a high rank for this type of lora so it was trained with a rank of 8. You should notice a nice bump in prompt following. The generated video also used my unreleased MisticNSFW lora which I am still testing. For better NSFW in general with H3 probably use one of the AIO NSFW loras in the meantime while I am testing MisticNSFW H3. Don't want to post it until its ready so you don't get subpar gens. It is looking good so far though. The VBVR lora works on non NSFW as well the same as it does on NSFW. I use the lora at a strength of 1.0 but experiment like usual for the best results. The lora was trained for I2V, T2V I have not tested. V4 has been trained on Sulphur 2 Base and was trained on 7000 videos with rank increased to 128 to give it more capacity to be able to improve prompt following. Using the distilled 1.1 lora at a strength of 0.5 in wan2gp you can crank steps all the way to 50 which can improve quality of gens. Not required though like it was not required in V3 or below but a nice option to maximize quality if you want. I usually do 10-12 steps. Videos are generated using 10eros since its the version of Sulphur optimized for i2v. Wall of text with a bunch of info below V3 has been optimized for motion. Expect noticeably livelier, more dynamic movement compared to previous versions. I generated the example videos using the int8/fp8 distilled model using 8 steps in wan2gp. For on site generation the dev model has been reported to be broken, if using on site gen try the distilled one instead! I can highly recommend LTX 2.3 Sulphur 2 base if you are using comfyui and if your using wan2gp use the Sulphur 2 base rank 768 lora What changed in V3: Attention-only layers. The feedforward layers have been stripped, leaving only the attention weights. It seems like the prompt following and reasoning behavior most likely live in the attention layers, while the feedforward layers were potentially interfering with natural motion, likely by over-learning features like textures and style from the training data. Better motion, smaller file size. A LoRA that improves prompt following, temporal consistency, and motion "precision" for LTX 2.3. Reduces the floaty, drifty motion that LTX tends to add to scenes. Things that should move, move with purpose. Things that shouldn't move, move less. Also works on non-NSFW, non-Furry, realistic, animated etc. It responds well to detailed prompts. In comfyui or wan2gp lowering image strength to 0.85 can improve motion in general if you want more motion Feedback and A-B comparisons welcome. V2 and V1 was trained on 4800 videos. Recommended to run at strength 1.0-0.7 but experiment to find what works best for your setup. If you want stronger prompt adherence try strength 1.5-2.0. I have noticed the only side effect I have gotten from a high strength is the video looking like its 16fps. I have not seen the choppiness issues on i2v unless the lora is cranked to 1.5-2.0 so it may be a t2v thing. Prompting tips in non-nsfw terms so its less confusing just adapt it to nsfw: Be specific and literal. Describe what happens, in what order, step by step. Instead of "a ball bouncing around" → "A red ball moves to the right, bounces off the wall, and returns to the center" Instead of "fluid pouring" → "Water flows from the left container through the connecting tube into the right container until both levels are equal" Describe the starting state, the action, and the end state The LoRA follows prompts more literally than base LTX — precise prompts will give much better results How was it made? v0.1 and v0.2 were trained on 360 videos from the VBVR (Very Big Video Reasoning) dataset synthetic task videos where every motion is precise and intentional. No concept bleed, no style change, just tighter control. Based on the paper which demonstrated this approach on Wan 2.2. I noticed that lora helped prompt following and temporal consistency a ton with wan so I trained this for LTX, and now H3. What does it actually do? Prompt following is more faithful — the model does more of what you asked instead of improvising Motion is more deliberate and less erratic Reduces random drift and wobble in scenes Temporal consistency improved — actions follow logical sequences What it doesn't do: Doesn't change visual style Doesn't add or remove capabilities LTX OR H3 doesn't already have Not a motion LoRA — stacks with motion LoRA's Training details for v0.1 and v0.2 (if you give a shit) Rank 32 360 VBVR synthetic videos at 512x512, 81 frames <------Alot less than 1 million but still a shitload to train on this is very slow to train locally. LR 1e-4, adamw8bit Training details for V1 Training videos were increased to 4800 Resolution is the same but frames were increased to 121 Every other setting the same as v0.1 and v0.2 More training data from the VBVR dataset was