My LTX-2.3 I2V workflow inspired by nomadoor. I've tweaked a few things that work for me and it may not work in some cases.
The workflow generates two videos: a ~0.25 megapixels low resolution video ("small_XXX") after the first stage and a 2x upscaled ~1.0 megapixels final video ("big_XXX") after the second (upscale) stage. After the first stage you could play the "small" video and if you don't like it you could cancel the second stage and restart the generation. Also a low-quality preview is generated on the fly (via LTX2.3 Tiny VAE), so you could cancel even before the first stage is complete. Note that the 1.5x or 2x upscale stage for LTX-2.3 is by desigh and is not optional (unlike Wan for example), although tests show that you can do one-stage generation with high resolution and skip the upscale if your vid is relatively static.
Tips:
The default sampler in the workflow is "lcm", but "res_2s" is better, especially for vids with lots of action and smaller faces, if you don't mind waiting twice as long (res_2s is slow).
If you have enough VRAM + RAM (always set your virtual RAM to at least 150GB, even 250GB if you have enough disk space), you could set length to 241 for 25 fps (or 513 for 50 fps) or even more (at some point you'll get OOM error).
"megapixels 1.00" (or "1.50" or "1.70" etc) in the "Resize Image/Mask" node is an arbitrary safe value, but any value between 2 and 4 is better for finer details in the upscaled final video - the bigger the better as long as you have enough memory. In theory you can go up to 8.3 megapixels (4K resolution).
The workflow defaults to random seeds (seed -1), but it’s better to use fixed seeds (click 'new fixed random' for both stages before each run). This allows you to change only the second-stage seed and when you click 'Run' again, only the second stage will regenerate with slight variations in video and audio. If you encounter an OOM error while using fixed seeds, simply click 'Run' again — the second attempt sometimes succeeds.
Model:
ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors (Kijai)
SHA256: f0b8f92f8a33daf7e7f508878db6d6f25f9c8bc7b90fcb5650fb402a9d9e9f13
or
ltx-2.3-22b-dev_transformer_only_bf16.safetensors (Kijai; heavy, but recommended - works even on 12GB VRAM if you set your virtual RAM to 100GB or more)
SHA256: 0c83957f8c79d49c52ff5e3c0980a759d4dd532563d62cab35fda2b247606bd2
Video vae:
LTX23_video_vae_bf16.safetensors (Kijai)
SHA256: 01ea62d09bc139f95c5dee7b5c062ad6a3e6cd8be910a1983ac02e7eb5b8ee3b
Audio vae:
LTX23_audio_vae_bf16.safetensors (Kijai)
SHA256: 5bc10fa4adecf99dda132d916e23048cbd56797702c5fa50eb5d2079048a38c3
Lora:
ltx-2.3-22b-distilled-lora-384.safetensors
SHA256: 2943ab994f3c9d88052e5a2a34cca14e4a2dfc36b1d8c407931d52d5c25dd72b
Text encoders:
gemma-3-12b-it-IQ4_XS.gguf (unsloth, more)
SHA256: e209a303d9d0841c1ab29a728b930ef630ba27766aa0067258f9e400f7ae8c65
ltx-2.3_text_projection_bf16.safetensors
SHA256: 911d59bb4cb7708179c9a0045ea0fe41212ecfb77aed3a02702b7c0a8274911f
Latent_upscale_model:
ltx-2.3-spatial-upscaler-x2-1.0.safetensors
SHA256: 93800de87dbc448b5b31f3c5c3a1579ba6335151de061a564f6f026b0fc770ad
or
ltx-2.3-spatial-upscaler-x1.5-1.0.safetensors
SHA256: b2b3193e68cb04b4701e1d59bec4ed5d5e3e84506d9a42d5a129d37d39823df7
LTX2.3 Tiny VAE
taeltx2_3.safetensors (madebyollin)
SHA256: f0773b4e3e57318e6aa4dd4a35e1d16213a5f160fbc0376163f06888bbcbe246
Model Storage Location:
📂ComfyUI/
└── 📂models/
├── 📂diffusion_models/
│ └── ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors
OR
│ └── ltx-2.3-22b-dev_transformer_only_bf16.safetensors
├── 📂latent_upscale_models/
│ └── ltx-2.3-spatial-upscaler-x2-1.0.safetensors
OR
│ └── ltx-2.3-spatial-upscaler-x1.5-1.0.safetensors
├── 📂loras/
│ └── ltx-2.3-22b-distilled-lora-384.safetensors
└── 📂text_encoders/
│ └── gemma-3-12b-it-IQ4_XS.gguf
│ └── ltx-2.3_text_projection_bf16.safetensors
└── 📂vae/
└── LTX23_video_vae_bf16.safetensors
└── LTX23_audio_vae_bf16.safetensors
└── 📂vae_approx
└── taeltx2_3.safetensors