LTX 2.3 Sneaky Drop! Plus: A New AI Video Model!

AI Video moves fast — and LTX 2.3 just quietly added some very interesting new video-to-video controls inside LTX Studio and likely Open Sourced. In this video, I’m testing what LTX 2.3 can actually do with video-to-video, pose, depth, edge controls, HDR support, and stylization workflows. Some of it works surprisingly well. Some of it gets weird. And yes, we get at least one classic AI body-horror moment, because nature is healing. I also take a look at a powerful open-source Prompt Relay / LoRA workflow for more advanced users, a brand new AI video model called Bach from Video Rebirth, some early hints about a mystery image model, Seedance’s upcoming Cameos/Cast feature, and a free open-source tool for building your own AI video training datasets. LINKS MENTIONED: LTX Studio: https://app.ltx.studio Prompt Relay / LoRA workflow: https://civitai.com/models/2553704/ltx23-all-in-one-prompt-relay-id-lora-controlnet-detailer-upscaler-custom-audio-keyframes Bach / Video Rebirth: https://www.bach.art/ Open-source AI video dataset tool: https://huggingface.co/dx8152/Video-dataset-slices Dataset tool tutorial: https://youtu.be/h27Sedb_v08?si=gyIjnE9cK-YEMKXl 00:00 What’s in today’s AI video roundup 00:35 LTX 2.3’s under-the-radar update 00:46 Why HDR support matters 00:58 New LTX 2.3 video controls 01:18 Are the new controls open source yet? 01:39 Quiet launch inside LTX Studio 01:49 Burning credits on vanilla LTX 2.3 02:00 First video-to-video test 02:13 GPT Image 2 stylization pass 02:30 Hands, lip sync, and identity drift 02:49 Starship Troopers Roughnecks test 03:10 Why this old CGI clip is useful 03:35 LTX 2.3 video-to-video result 03:52 Best result so far — but not perfect 04:16 The backwards video workaround 04:31 Why strong references matter 04:58 Short-shot problems in LTX 05:21 Pose, depth, and edge modes explained 05:58 Pose-control stress test 06:17 Flamethrower girl replacement test 06:46 Depth surprisingly works better 06:57 Edge mode gets weird 07:08 Vanilla model’s fast-motion limits 07:22 Stylization transfer works better 07:41 Turning live action back into anime 08:16 Final credit-burning test 08:52 The two-second tail workaround 09:19 Why LTX still matters 09:36 Seedance 2.0 comparison 09:53 Why LTX pricing is interesting 10:08 Three open-source pieces coming together 10:27 Prompt Relay, ID LoRA, and IC LoRA 10:39 The workflow that actually impressed me 11:22 Comedy gold + workflow link 11:39 ComfyUI anxiety warning 12:07 Why platform users should still care 12:28 New AI video model: Bok 12:44 Bach’s focus: character consistency 13:02 First-look limitations and credits 13:24 Montage and style preset features 13:46 No-cherry-picking first test 14:07 Man in blue suit result 14:25 Character consistency claims 14:36 Celebrity likeness tests 15:08 Where Bach starts to break down 15:22 Don’t use celebrity likenesses 15:37 Pricing and should I review it more? 15:48 Mystery image model tease 16:01 Early-access impressions 16:12 Seedance Cameos / Cast quick hit 16:34 Watch out for fake app listings 16:55 Open-source video dataset tool 17:19 What the dataset tool actually does 17:43 Tutorial, language notes, and UI 18:03 Who should try this 18:16 Final thoughts / AI woke up today 18:28 What’s coming next Subscribe for more grounded coverage of AI video, AI filmmaking, image generation, creative AI tools, and the weird little corners of this space that usually matter more than the hype cycle.