Lightricks

/

LTX-2.5

Like 11.7k

Follow

LTX.io 6.74k

Hi, I’m looking for a reproducible local ComfyUI workflow for five-second, medium-wide dialogue shots. I want the speaker’s face, eyes and teeth to keep their shape through speech, head turns and the reaction afterward. I’ve seen visually stable low-resolution Grok Imagine clips, so higher resolution alone doesn’t seem to be the answer. Those reference files have no audio track, so I’m using them as a visual benchmark only.

Setup: Windows 11 Pro, RTX 3060 12 GB VRAM, 32 GB system RAM, ComfyUI 0.42.0, PyTorch 2.11.0+cu130, low-VRAM mode. LTX-2.5 22B Distilled INT8 ConvRot, Gemma 4 12B INT8 ConvRot, LTX-2.5 Conv Video VAE BF16 and Audio VAE BF16.

Baseline: Image-to-video with an initial frame, fixed seed and native audio. Euler ancestral, CFG 1, four-step shortened distilled schedule (1.0, 0.975, 0.725, 0.421875, 0.0), 24 fps, 121 frames. Tiled decode with spatial tile 512 / overlap 64 and temporal overlap 8. I use temporal window 256 for the current tests.

What I tried

• INT8 vs BF16 Distilled: Same seed, reference and settings at 480p, 720p and 1080p. BF16 ran slower with no clear improvement in facial identity or expressions in the shots I inspected. For a single-character 720p shot, INT8 took 3:14 and BF16 took 4:06.
• Sampling: On a 480p two-person shot, 4 steps took 1:39, the standard 8-step schedule took 2:26, and an experimental 20-step subdivision took 4:05. More steps changed the acting and gaze without reliably improving the eyes or mouth.
• Decoder: At 30 fps and 153 frames, decoding the same latent with temporal window 256 reduced double contours near 4 s compared with 128. The audio was identical. That helped with a visible artifact, but facial articulation is still an issue.
• Two stages: Four initial steps, 2× latent spatial upscaling, then three refinement steps at 720p, 24 fps and window 256. I also shortened the prompt to action and audio. The dialogue transcribed correctly, but this combined test showed no clear improvement in eyes or skin.
• Audio-driven INT8: Encoded a prepared speech track, masked its latent from denoising and muxed the original audio. In two 720p attempts the line was present, but the mouth barely articulated in places and the head turned away. MuseTalk on a silent LTX render preserved the head and eyes, but left the mouth looking pursed and softened the surrounding skin.
• First/last-frame conditioning: I tested 720 × 1280 with first and last frames from the same visual reference, 4 steps, 24 fps and window 256. The last head pose followed the guide, but the model still generated unwanted on-screen text during the line. This run took 3:18.

Has anyone here achieved natural lip articulation, stable eyes and smooth head movement on a 12 GB GPU with LTX-2.5? A reproducible workflow or a short example would help. In particular, I’d like to know which checkpoint, audio conditioning method, sampler schedule and VAE you used, and whether you tested a medium-wide shot at 480p or 720p. I’m especially interested in setups that keep the face stable throughout a spoken line.

Thanks for any ideas or examples.