Why do you provide different VAEs and Text Encoders for Dev vs. Distilled models?
Neither the offical repo https://huggingface.co/Lightricks/LTX-2.3 nor Kijai https://huggingface.co/Kijai/LTX2.3_comfy/tree/main provides different VAEs or Text Encoders for Dev vs. Distill models, so why have you?
Are the VAE and Text Encoder files actually the same (the sizes match), but for some reason have different SHA256 hashes?
On the original repository, the Diffusion and VAE components are merged into a single model file, which requires using the Checkpoint Loader in ComfyUI instead of the Diffusion Loader.
To quantize the Diffusion model without affecting the VAE, both Unsloth and Kijai separated these two components. If everything remained packed into a single file, the quantization process could break the file structure and render the model unusable.
You might notice that the VAE versions from Kijai and Unsloth have different SHA256 hashes, meaning their contents differ. However, both should work perfectly. Because the Diffusion and VAE parts are decoupled, you can freely mix and match models and VAEs from different providers (for example, using Unsloth's LTX model alongside Kijai's VAE).