Correcting the Record
Added 2025-03-07 19:59:56 +0000 UTCHey, I was up way too late last night generating videos with Wan 2.1 and LTXV. And I posted about it and wanted to correct some things on the record.
Correction: For Wan 2.1, only the 14B model has i2v capability. Grok was confused and I was unclear in my prompts (and had some typos). There is no "1.3" i2v model, only t2v (though this runs great and fast on my rig). But, I'm having good success generating using wan2.1_i2v_480p_14B_fp8_e4m3fn.safetensors with only 8 GB of VRAM despite Grok's claims that this shouldn't be possible. Still, it does take about 1.5 hours to generate 5 seconds of video typically for me.
But, so far, the results are very promising, with some of the best stuff generating after I'd posted yesterday. Here are a couple of my favorites attached to this post and included in the release archive now under:
Releases\2025\013-Wan2.1LTXExperiments\
This feels close to the likes of Kling, and this is all with just the base model (which seems to understand things like futanari and semen/fluids just fine) and no training/use of LoRA's. So, super impressive. Granted, with my crappy hardware I'm sure the FP8 quantization and low resolution/number of frames is costing me a bit of quality. All the same, the fact that I can generate this stuff at all locally seems like a minor miracle. And this was using native nodes in ComfyUI and way easier to set up than Hunyuan while using less VRAM and following prompts better.
