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

H3 VBVR Pro v2 is a major upgrade over v1. It was trained on a subset of the VBVR Pro SFT dataset (https://huggingface.co/datasets/Video-Reason/VBVR-Pro-SFT-Video) at a higher resolution, at rank 128 (up from 16), and with better hyperparameters. It follows prompts more faithfully and consistently than v1.

Start at a strength of 1.0. That works on both stock MiniMax H3 and Eros beta 6, so v1’s advice to drop to 0.5 on Eros no longer applies. Everyone’s workflow and LoRA stack is different, so adjust from there. The example videos were made at 1.0.

H3 VBVR PRO V1 INFO BELOW

_________________________________________________________________________

The H3 VBVR Pro version was trained on 1k videos from the VBVR Pro SFT dataset https://huggingface.co/datasets/Video-Reason/VBVR-Pro-SFT-Video 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

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 "A Very Big Video Reasoning Suite" 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 added to v1

Below is the new dataset I trained on's data composition if your curious

Tier 1 — Physics and Motion (3,400 samples)

Core generators at 300 each: `G-11` (object reappearance) has a shape move off-screen in a direction and return along the same path — teaches trajectory and object persistence. `G-25` (separate object spinning) is a shape that rotates in place then translates horizontally to a target position — multi-step motion sequencing. `G-33` (visual jenga) is a stack of objects that get removed one by one from top to bottom — sequential extraction with implicit physics ordering. `O-29` (ballcolor) is ball tracking tasks with color — motion following plus identity preservation. `O-52` (traffic light) is discrete state transitions, lights switching on/off between green and gray — teaches the model that state changes are crisp, not gradual. `O-75` (communicating vessels) is fluid equalizing between connected tubes based on pressure — continuous physics simulation over time. `O-87` (fluid diffusion) is ink spreading in water — another continuous physical transformation but with expansion rather than equalization.

New additions at 250 each: `G-35` (hit target after bounce) is a ball with an initial direction that bounces off walls following reflection laws to hit a target — pure trajectory prediction with physics constraints. `O-30` (bookshelf) is book rearrangement on shelves — the specific task VBVR highlighted where their model beat Sora 2.

Multi-step transforms at 160 each: `O-7` (shape color change) is a single transformation — shape changes from one color to another. `O-8` (shape rotation) is a shape rotating by a specific angle. `O-13` (outline then move) is two sequential steps: change a shape's outline style, then move it to a new position. `O-14` (scale then outline) is also two steps: scale a shape up or down, then change its outline. These four together teach the model that instructions are ordered and each step completes before the next begins.

Tier 2 — Spatial and Reasoning (1,420 samples)

Proven generators at 100 each: `G-13` (grid number sequence) is filling in number patterns on a grid. `G-17` (grid avoid red block) is pathfinding on a grid while avoiding obstacles. `G-31` (directed graph navigation) is finding the shortest path through a directed graph. `G-41` (grid highest cost) is evaluating spatial values on a grid to find the optimal path. `O-24` (domino chain) is a sequential cascade where dominoes fall until they hit a gap — teaches causal chains and stopping conditions. `O-34` (dot to dot) is connecting numbered dots in sequence — ordered drawing. `O-47` (sliding puzzle) is tile rearrangement under constraints, like a 15-puzzle. `O-83` (planar warp) is warping a grid to align with a target quadrilateral — geometric transformation.

New reasoning diversity at 130 each: `O-1` (color mixing) is RGB additive mixing where two light sources combine and the result fills a target zone — rule-based continuous process. `O-33` (counting objects) is exactly what it sounds like — count things correctly. `G-3` (stable sort) is arranging objects by a rule while preserving relative order. `G-37` (symmetry random) is completing a pattern by mirroring across an axis. `O-21` (construction blueprint) is fitting a correct puzzle piece into a gap in a structure. `G-44` (BFS) is breadth-first search traversal of a graph — systematic layer-by-layer exploration.

The overall dataset is weighted roughly 70/30 toward physical motion and transformation tasks over abstract spatial reasoning, All of these are taken from the VBVR dataset I am not the creator of the dataset. I'm pretty new to lora training so if you have tips let me know.

REMEMBER its not X, its Y.

Disclaimer & Terms of Use

This model is provided "AS IS", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and noninfringement.

IN NO EVENT SHALL MisticRain69 BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL OR ANY OUTPUT GENERATED WITH IT.

18+ only. You, the user, are solely responsible for anything you generate with this model and for ensuring your use complies with all applicable laws and the CivitAI Terms of Service.

You may not:

produce anything illegal, non-consensual, defamatory, or used for harassment, deception, or fraud;

create any content depicting, or appearing to depict, minors (real or fictional) in a sexual context — no exceptions, ever;

use this model to depict any real, identifiable person — celebrity, public figure, or private individual — in any context, SFW or NSFW. No likeness use, period.

If you post outputs to the gallery that violate the CivitAI TOS, you'll be blocked and reported. Don't do this shit. Don't be fucking sus.

Files with SHA256: cb3e9f2e395f63e0325213c0b802f26e6d4b6b1caf3f714a969b9aa56242bccb