This is HQ Int8 Row ConvRot of LTX 2.3 v1.1 Distilled / Turbo 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 v1.1 Distilled / Turbo 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 (13)
Not exactly sure what this is, i don't have that much knowledge. But compared to regular dev int8 convrot with DMD. This delivers bloody amazing results not using any distilled loras. Never had such clarity and motion with LTX.
32 Gb RAM / 16 Gb VRAM / PlagueKind ComfyUI v6 Workflow
With 3 semi-large loras and Eros lora, 1024x1024, 315 seconds for a 20 second 24 fps clip.
It does stress my CUDA cores a bit.
thanks for comment. yes this is highest quality ever published atm.
huggingface may be nicer for download
It's so disingenuous when someone puts stuff behind paywalls without mentioning that when providing the links. I never like seeing the hate you get on reddit, I always thought it was a bit mean, but I'm starting to understand it now. Don't be a dick. Be honest,
tell me who are you working for free? where do you drive work 8 hours a day for free and return back
First, I would like to thank the creator of the model.The model is faster than "Sulphur" and showed slightly superior quality to "Sexgodpinkcherry," however, the use of LoRAs is necessary to achieve "things that don't come by default." The fp8 model still proves to be unbeatable (31,8% faster) .I am learning from you all and I welcome suggestions to improve the "workflows" at any time. Here are my results:Ltx23TurboDistilledv11_v10:
Lora: Dre4ml4y-V3---strenght 1---video---1---audio strenght--0 (very important)
Distilled lora----none (It worked fine without regular or consafe loras)
Sage attention: weight type-fp8_e4m3fn
Compute type----default
patch_cublaslinear-false
sage_attention: sageattn-qk-int8-pv-fp8-cuda++
10.75s 34.09s ----Prompt executed in 232.23 seconds
Sexgod Pinkcherry v13-fp8
Lora: distilled 1.1-ceil72-condsafe---strenght-0.80-first pass 0.55 second pass
Sage attention: weight type-fp8_e4m3fn
Compute type----default
patch_cublaslinear-false
sage_attention: sageattn-qk-int8-pv-fp8-cuda++
6,5s 21,44s----Prompt executed in 176.31 seconds
Im personally loving sexgod for any NSFW stuff, far better than sulpher. Sexgod, so far, is the only one to keep the penis identity the same throughout the entire video in most cases, even penis specific loras aren't able to keep the penis identity as well as Sexgod. But i will definitely give this one a go for things that dont have a penis! lol. Thank you for the stats!
How does this handle the sulphur lora vs just using eros 1.4?
Interesting. So I ran all three 'transformer only' models with the cloned settings just to test it myself on my 5090, and my median E2E times were bf16-26.4s, int8-16.2s, fp8-16.1s (Add on 2.5s on the 8's and 4s on the bf16 to that if we count having to load the models from the drive too)
Add on ~20% total time for the non 'transformer only' models.
this is clean. real good stuff. adding on some standard ltx 2.3 loras gave me nice generations on pretty good speed for average consumer hardware
Is the file name "ltx23TurboDistilledV11_v10" the correct one?
EDIT: NM, figured it out. The model works really well in more natural skin and facial expressions. Realized I could download it from your model downloader! Tks..
Not a bad model. I haven't checked how much its performance has increased compared to the dev model, but it responds to requests more correctly and creates a more 'responsive' image.
tested. wan22 is dead