DeepHomage

Synthetic media artist. Iconic stars re-imagined with face swapping/deepfake and VFX. Have a project in mind? Contact me: [email protected]

[DeepHomage @DeepHomage]

@NVIDIAGeForceDE I really need an RTX 5090 or RTX 5000 Pro for making 4K videos, but an RTX 5080 will be gratefully accepted if I'm the lucky commenter.

[DeepHomage @DeepHomage]

@MayorKingAI I made this to test musubi-tuner-ltx2's CREPA and dino options when training a video LoRa. The original paper used an animated character: https://t.co/PdHyb5ldxv

[DeepHomage @DeepHomage]

@carolletta OK, I'll take a chance. Any fans of old Hollywood torch songs in the house? https://t.co/cbdiOtqTHc

[DeepHomage @DeepHomage]

@RJthedarkside @expertwith_AI Send

[DeepHomage @DeepHomage]

@ostrisai @ltx_model @ComfyUI I like the video and I've used the toolkit to make LoRas locally, but I don't see any config examples for LTX-2.3 LoRa training on your github site. Do you have another branch for LTX-2.3 LoRa training in development?

[DeepHomage @DeepHomage]

@ruu_1010 @bdsqlsz Better to search for a pre-made Windows wheel that matches your python version with pytorch 2.11 and CUDA 13. Complicated to compile, can't explain on Twitter/X, but ChatGPT or similar can help you.

[DeepHomage @DeepHomage]

@ruu_1010 @bdsqlsz Turns out that today's flash attention compile actually succeeded, after an hour or so. So, I've got a Windows wheel file for python 2.12, pytorch 2.11 and CUDA 12.6. If there's any interest from fellow 30x owners, I can post it on my github page.

[DeepHomage @DeepHomage]

@ruu_1010 @bdsqlsz No. I'm using python 3.12, pytorch 2.11 and CUDA 12.8 with an Ampere 30x GPU. No joy installing the wheel, but thanks for making it. It's possible to compile flash-attention with MS Build Tools installed and some patience, but I've had no luck compiling this combination.

[DeepHomage @DeepHomage]

@rasmalai I'm using a trial of Perplexity Pro, but people who code for a living tell me that Claude can write usable code, and the others can't. I've avoided Copilot -- reminds me of Clippy - and Microsoft harvests too much Windows data already.

[DeepHomage @DeepHomage]

@carolletta I suspect this experimental music video is an acquired taste, but I'm interested in all reactions, positive and negative.

[DeepHomage @DeepHomage]

@MattVidPro LTX uploaded a new version of the 19B distilled model today, updating "Distilled checkpoint with correct VAE" https://t.co/QLlyZ7mX9X. Might fix some nagging issues, but existing workflows might need updating too.

[DeepHomage @DeepHomage]

@KingSTOOPID I went to work for longer than some of the people here have existed on Earth, but I don't go to a 9 to 5 day job these days.

[DeepHomage @DeepHomage]

@brockpierson If you mean boosting small accounts, I'm for it, but if you mean something else, hard pass.

[DeepHomage @DeepHomage]

@ltx_model I've used both Wan 2.2 and tried LTX-2 to a lesser extent, and while the native audio is appealing, there's more maturity/usability in the the Wan 2.2 ecosystem at the moment, and LTX-2 has some growing pains. especially in ComfyUI. But competition is good for open source.

[DeepHomage @DeepHomage]

@NvidiaANZ “GeForce Season” is here. At a crossroads between RTX 5000 Pro (48 Gb.) and RTX 5090 (32 Gb.) for content creation. More memory vs. more performance. Winning one would make the choice easy.

[DeepHomage @DeepHomage]

"Almost Human — AI Cyborg Music Video | Cyberpop / Hyperhop | LYRA-09” https://t.co/L8U62Q0frQ via @YouTube

[DeepHomage @DeepHomage]

@NVIDIAGeForce GeForce RTX 5090, I want a signed one. But I will accept the 48 Gb. Max-Q a6000 refresh for training models as a consolation prize. Please release it soon and send one to me.

[DeepHomage @DeepHomage]

@bressane @freepik Nano Banana asserts that they can colorize and restore photos better than the open source Kontext model, something to be expected from a paid service, but I haven't seen any side-by side comparisons.

[DeepHomage @DeepHomage]

@ScarabOfficial You can compile xformers for pytorch 2.8 on Windows, but it's not a simple task. I expect that a pytorch 2.8 compatible xformers version will released soon. In the meantime, it's pytorch 2.7.1

[DeepHomage @DeepHomage]

@cfryant Challenge accepted. I'm also looking for some inspiration and trying to master prompting text to video models. Alas, only 2-5 second clips possible with a local video model: