This is HQ Int8 Row ConvRot of LTX 2.3 Dev / Raw / Base model. Not the Distilled / Turbo fast model.
Made from official BF16 model with SECourses Musubi Trainer Quantization app
You can download and use Musubi Trainer app for both training and quantization from here : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634
To be able to use this model with very best performance please use our Torch 2.13 CUDA 13 ComfyUI installer with ready presets : https://www.patreon.com/SECourses/posts/download-comfyui-installers-and-presets-105023709
I also recommend our SwarmUI installer with ready SwarmUI presets : https://www.patreon.com/posts/download-swarmui-installer-and-presets-114517862
With our HQ Int8 Row ConvRot quant conversation and app and preset, the model quality is able to surpass GGUF Q8
With our ComfyUI backend, Int8 Row ConvRot is able to generate faster than FP8 Scaled literally 100% faster on RTX 3000, 4000 and 5000 series GPUs
Model quantization is taking around 3-4 hours on RTX 5090 since we do training like quantization with prodigy optimizer
Check model screenshots to see and learn more
Description
This is HQ Int8 Row ConvRot of LTX 2.3 Dev / Raw / Base model. Not the Distilled / Turbo fast model.
Made from official BF16 model with SECourses Musubi Trainer Quantization app
You can download and use Musubi Trainer app for both training and quantization from here : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634
To be able to use this model with very best performance please use our Torch 2.13 CUDA 13 ComfyUI installer with ready presets : https://www.patreon.com/SECourses/posts/download-comfyui-installers-and-presets-105023709
I also recommend our SwarmUI installer with ready SwarmUI presets : https://www.patreon.com/posts/download-swarmui-installer-and-presets-114517862
With our HQ Int8 Row ConvRot quant conversation and app and preset, the model quality is able to surpass GGUF Q8
With our ComfyUI backend, Int8 Row ConvRot is able to generate faster than FP8 Scaled literally 100% faster on RTX 3000, 4000 and 5000 series GPUs
Model quantization is taking around 3-4 hours on RTX 5090 since we do training like quantization with prodigy optimizer
Check model screenshots to see and learn more
FAQ
Comments (16)
How is this different from the existing INT8 ConvRot quants of the base model on Huggingface? What does "HQ" actually mean here, and how is this distinguished from standard INT8 ConvRot quants?
this took over 3 hours to compile on RTX 5090 and made with accurate best preset
Gotcha, so it was computed with more rigorous quality. Cool, I'll give it a go!
@Bbbrrr yep and thanks
@Bbbrrr It's also a conversion with the VAE still intact, not just the 'transformer_only' like Kijai uploaded. So you don't need special nodes to load separate audio / video VAEs.
@JonXL thanks for extra info you are right
@SECourses Hey Furkan, So, does that mean the VAE and CLIP are also quantized with INT8 ConvRot?
I could confirm it immediately by downloading the file, but I'm curious.
There must be many users who do not want the VAE and CLIP to be quantized as well. Especially for a 5090, there is no real reason to quantize them, right?
This is the part that is confusing me.
:
I have another question. Between the Distill LoRA 384 rank v1.1 and the Codsafe90 ceil72 LoRA, which one do you prefer for I2V?
Also, if I merge this Distill LoRA into a pre-LoRA using good weights (0.5 / 0.6 / 0.7) and then convert it into INT8 ConvRot, would it provide any additional advantage in terms of processing speed?
Or would the speed improvement be negligible compared to simply applying the LoRA during inference?
I’m curious about this because I wonder whether pre-merging the LoRA before quantization can actually improve performance.@SECourses
how many steps? and any lora to decrase steps needed?
8-20 steps
unless you want garbage i wouldnt go under 8 steps.
make sure to use distill lora as well
Please add a transformer only version. Thanks
Can I ask what happened to your other trained models? I was wanting to test them but now I see they're hidden/deleted.