INT8/INT4 ConvRot comfy compatible quantizations of LTX2.3 video model
INT4 Models are mixes of int8 and int4 convrot
F = Full INT4 ConvRot
B = Balanced INT4 INT8 Mixed ConvRot
Q = Quality Leaning INT4 INT8 Mixed ConvRot
Description
INT4 ConvRot Mixed Quality Quantization
FAQ
Comments (13)
Nice.
i tried to use the INT4Q. surprisingly, the stocks model loader can load this. no need to use the INT4-fast loader.
Just done a simple test on RTX3060 12GB VRAM.
736 X 1280, 5 second i2v, same prompt and seed, plus RFIE 417 double the fps and RTX super resolution X 1.5 size.
total generation time:
INT8: 171 seconds
INT4Q (Quality?): 144 seconds
INT4B (Balance?): 128 seconds
INT4F (Full int4?): 125 seconds.
Once the model size close to 12GB, the improvement in generation time will be decrease.
but quality-wise speaking, i can easily spotted that.
INT4F and INT4B would be somehow noticeably .. not usable. INT4Q would be balanced while INT8 still nice.
I named them like that as these are INT4+INT8 convrot mixes, INT4Q has more INT8 layers, INT4B has 45% INT8 layers and INT4F is a full INT4 convrot quant. Fewer INT8 layers = lower quality, lower file size, more speed boost
Uploaded a fixed quant of int4b as the previous one had 0 int8mm as i quantized it after INT4F, New quant will be slightly slower but a slight improvement to quality 😅
Even after making sure ComfyUI is updated to 0.27, any attempt to load it I get (from Diffusion Loader):
[INFO] Found quantization metadata version 1
[INFO] Detected mixed precision quantization
[INFO] Using mixed precision operations
[INFO] Native ops: float8_e5m2, float8_e4m3fn, int8_tensorwise , emulated ops: nvfp4, mxfp8
[INFO] model weight dtype torch.bfloat16, manual cast: torch.bfloat16
[INFO] model_type FLUX
[ERROR] !!! Exception during processing !!! 'convrot_w4a4'
...
[INFO] ComfyUI version: 0.27.0
[INFO] comfy-aimdo version: 0.4.10
[INFO] comfy-kitchen version: 0.2.16
Any idea's?
In manager what version are you on, im on ComfyUI: v0.27.0-56-g8b099de3 (2026-07-12)
which has the int4 support. I went to my folder ComfyUIwindowsportable\ComfyUI opened cmd prompt and typed in
git checkout master
git reset --hard origin/master
to get my install up to date with the latest commit Fix SaveVideo description: says images, saves video (#14885)
@tsolful Updating (again) did indeed solve that issue, and I do see that generated video's are for sure of a higher quality. BUT.. my generation times increased from 12-15 minutes to 2-3 *hours* (with various workflows, even the ones from https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main). Do these models need any other changes?
@dutchit288 Which GPU do you have? i get 6-8 minute 720p gen times on 3060 12gb, 32gb ram using the default diffusion model loader INT4B
@tsolful A 4070TI/OC (16G-VRAM/80GB-RAM). And yes, on any other model, even the full LTX2.3 dev, Sulphur dev, I also get 10-15 minutes (20-30 second video's, 720P+). But any of these INT8/INT4 gets me 2+ hours.
Some examples: https://civitai.red/images/136660171 (2h30 minutes), https://civitai.red/images/136514559 (12 minutes). Same workflow, just different model (workflow is linked)
And by completely reinstalling Comfy (and then also a big upgrade in torch and cuda versions I now indeed get better results (and even faster, below 10 minutes).
Downside: Need to re-install every phreaking custom_nodes :) Anyway, fixed.
@dutchit288 Yeah whenever i run into big problems like that i usually setup another comfyui install to test against my current install, glad to see you got it working 💚
the model no contain vae
Nope these are transformer only models, here are the vaes and the workflow i use https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/vae
https://huggingface.co/RuneXX/LTX-2.3-Workflows/blob/main/LTX-2.3_-_I2V_T2V_Basic.json