For each model, see the version details for more information. All the models here are experimental; they often won't work perfectly, the prompt will often need improvement, the input will often need alteration, or the parameters will need adjusting. Therefore, don't treat these models as something that should work perfectly right away for everything. They will only cease to be experimental when we discover a better way to do this. <a target="_blank" rel="ugc" href="https://huggingface.co/Alissonerdx/LTX-LoRAs">Alissonerdx/LTX-LoRAs</a> <a target="_blank" rel="ugc" href="https://huggingface.co/Alissonerdx/EditAnything">Alissonerdx/EditAnything</a> <a target="_blank" rel="ugc" href="https://buymeacoffee.com/nrdx">If this model was helpful to you in any way, please consider helping me continue creating more model for the price of a coffee.</a>

EditAnything

This model was trained on 8,000 video pairs, and training is still ongoing for a few thousand more steps. It is still experimental, not trained with a fully professional production target, and the model may be updated unexpectedly as new checkpoints.The current goal is not final polished production quality, but to explore:<ul><li>edit-anything behavior</li><li>prompt-following</li><li>inference tradeoffs</li><li>synthetic dataset building, especially for style data</li></ul>The model was trained around four main prompt patterns:Add <code>Add a/an [subject/object] with [clear visual attributes], [precise location in the scene].</code>Remove <code>Remove the [subject/object] [location or identifying description].</code>Replace <code>Replace the [original subject/object] [location] with a/an [new subject/object] with [clear visual attributes].</code>Convert / Style <code>Convert the video into a [style name] style.</code>Workflow URL: <code>https://huggingface.co/Alissonerdx/LTX-LoRAs/blob/main/workflows/ltx23_edit_anything_v1.1.json</code>One important thing during inference is CFG.A good starting point is testing a distilled setup with CFG = 1. If the edit feels too weak or the model is not following the prompt well enough, increasing CFG can be the key. In some cases, increasing the LoRA strength to around 1.2 can also help.The workflow is also not fully optimized yet. It still needs more testing to find the best combination of:<ul><li>CFG</li><li>LoRA strength</li><li>number of steps</li><li>model combinations</li></ul>It may also be interesting to combine this model with other models and see what kinds of results emerge.If you can test it, please share your findings. Feedback on prompt behavior, edit strength, consistency, style transfer, and failure cases would be very helpful while training is still in progress.Another very important thing is that the Removal task should have a very clear direction indicating where you want to remove what you want to remove.Examples:Remove the black robot sitting at the table.Remove the person riding the electric scooter on the left.Remove the person with glasses and the microphone in the foreground.Remove the image of the green trees on the top left.Remove the woman and the smoking bottle.For example, if the object are in front, use foreground ... background, left, right, top, bottom.Another way to remove things that don't want to be removed is to simply add a mask, for example magenta, over the object you want to remove, and use this video as a guide. When writing the prompt, you write something like: "Remove object masked with the pink color." Sometimes this is much more precise than waiting for it to recognize what actually needs to be removed, because in this case the biggest indicator is the magenta-colored object.

Files with SHA256: 36721b3988c468afbec2f5bc52eb534fc6a3398fcf150303eba24f27d9da2974