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
FAQ
Comments (29)
Thank you. I'm looking forward to trying this out. Had good luck with int8 with Krea2.
Whats the difference between INT8 convot and the 1.1 distilled?
Both are INT8 convrot, 1.1 distilled is the latest distilled version of dev
two version of 2.3 1.0 and 1.1.
I downloaded this model to check if the quality would be noticeably better than the model I've been using so far (GGUF - Q4) and how much longer the generation would take. And boom! Not only did the quality skyrocket with your model, but the generation time dropped by about 30-40% at 1080p, and it's more than twice as fast at 768p! I have a potato PC with an RTX 3060 (12GB VRAM + 32GB RAM), and dropping from 10 minutes to just over 4 minutes is an absolute game-changer! Thank you!
no way, 4 minutes on a 3060 with ltx? really? would you share a video with your workflow?
@hughchungus420
[INFO] Model LTXAV prepared for dynamic VRAM loading. 22404MB Staged. 0 patches attached. Force pre-loaded 608 weights: 3303 KB.
100%|████████████████████████████████████████████████████████████████████████████████████████████████| 8/8 [00:53<00:00, 6.64s/it]
[INFO] Model LTXAV prepared for dynamic VRAM loading. 22404MB Staged. 0 patches attached. Force pre-loaded 608 weights: 3303 KB.
100%|████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [01:49<00:00, 36.54s/it]
[INFO] 0 models unloaded.
[INFO] Model VideoVAE prepared for dynamic VRAM loading. 1384MB Staged. 0 patches attached.
[INFO] Prompt executed in 258.65 seconds
1344x768, 8 seconds, 24 fps (LTX 2.3 distill INT8 from here + Gemma GGUF Q2)
My workflow, here: https://drive.google.com/file/d/196C4pEUHj6MEh92b1iVYLHYXSWYKB2Op/view?usp=sharing
Btw, I generate clips for music video - so audio VAE is off.
@sheben11 damn boss real minmaxxing there with the q2'd text encoder. cool. i may give ltx one last try with how runnable it is under those circumstances.
@hughchungus420 I think it is worth trying. I will test INT8 version of Gemma tonight, I let you know how this setup works
@sheben11 Same specs here, real gamechanger using int8. How's prompt adherence using the Q2 text encoder? I use Q4 personally
@tsolful Q2 was good enough but... I tried INT8 and got some additional speed boost (not much, about 15-20 seconds - but it still makes a difference with large batches) and no OOM when using 13GB INT8 text encoder vs 5GB Q2 GGUF (!).
I use one from this hf repo: https://huggingface.co/Winnougan/LTX-2.3-INT8/tree/main/Text%20Encoder%20INT8
@hughchungus420
[INFO] Prompt executed in 238.17 seconds
It is the same batch as yesterday (i2v, 8 sec, 24 fps, 768p) but with INT8 Gemma instead my minimal Q2. 30xx loves INT8.
@sheben11 Now Winnougan have also int4 Gemma3 text encoder - 8Gb. I wonder if it be any good, and faster
@flo11ok874 Got it BUT... I re-discover (free) LTX Gemma API Text Encode node from LTXVideo node pack, and it's a way faster while no VRAM/RAM is used.
I used to think that it works only with full AiO model but it's not.
Simply copy your ltx-2.3_text_projection_bf16.safetensors file into models/checkpoints and choose it in the LTX API Text Encode node, provide your API code from here https://console.ltx.video/ (Gemma API calls are free of charge) and that's it.
You can intergate it in every LTX 2.3 workflow as I did.
LTX Gemma API speed? 7-8 s with and 3-4 s without prompt enhancement ^^
@hughchungus420 @tsolful Check it out :)
@sheben11 Thanks for reminding me, was using it when ltx 2 first came out but switched to q4 encoder 💚
@tsolful same i switched to the q4 text encoder as well, i can generate 16 seconds of video at 720p in under 5 minutes. crazy stuff.
@sheben11 You were able to run the int8 (nearly 22GB) on a 12GB 3060?! I just want to confirm that this is not a typo, as this could mean that I might be able to run it too (I can run the Q4 on my 16GB card). The sample videos on this page show much higher looking quality from the int8 as compared to the int4s, and if both, speed and quality, are improved, that would be amazing!
@SkyDood to put my two cents in here almost a month later; i suggest if you're on a modern card (4000 and up) just grab the fp8 version of whatever you're trying to run, save yourself the headache of hours worth of failgens lol. LTX is extremely sensitive to quantization and speed ups don't really matter if it takes several tries to get something usable.
@hughchungus420 I appreciate your input! I actually did just try the dev Int8 model above, and was surprised that it completed without OOM, especially because I had retrofit it into an LTX Director workflow that I'd been running the q4 on with a fair amount of success at not-too-painful speeds, but the first run of the Int8 took near 32 min for a 30 second clip, which is painful. LOL! I have a feeling I'd OOM, or at least see even longer gen times with an fp8 model. I'm gonna try one the Int4s to see if there is speed and quality boost above q4 next. Thanks again! :)
@SkyDood yeah huge recommendation there to be honest; just drop your runtime to half that, 16 seconds at most, anything more is going to be cripplingly slow even on high end cards, i cap mine out at 12 seconds to get less than 5 minute gen times at 720p 25fps with loras enabled.
@hughchungus420 I reverted back to the q4 and it took less than a third of the time around 9 minutes, which I think is pretty good for a 30 second clip (the int4 was about 15 min on first run, with substantial degradation in quality, btw). Oddly, I've deleted the two models I'd tried from this page, but for some reason, my HDD is not showing the space as having been reclaimed. I'd noticed it after deleting the first (dev int8) and was surprised to see the same thing after deleting the dev int4... Maybe they're cached somewhere, but this is alarming because I'm kinda low on space as it is LOL. Maybe a reboot will clear it up, but I've tested and deleted models right after in the past, and always saw the space they'd occupied cleared up. Maybe it is an int model thing. Hopefully a reboot will give me back that 30-something gigs.
Follow-up to my last comment mentioning space reclamation: Rebooting comfy seemed to clear it up. :) I have a feeling I may have had better results with these dev models if I wasn't pushing them through the distil lora I'd been using for the q4, but I think I've experimented enough with LTX for today. I may try one of the distilled models next time.
It is possible to make INT8 version of spacial upscaler model? 🙏
I'll run a few tests with the int8 version of it, but I think the performance increase will be minimal, as the upscaler upscales the latent, which then gets passed to the sampler
Where to find a workflow compatible with it ?
Any workflow is compatible with int8 models as comfyui natively supports it, i currently use this workflow from runexx https://huggingface.co/RuneXX/LTX-2.3-Workflows/blob/main/LTX-2.3_-_I2V_T2V_Basic.json, place in diffusion_models folder and use Load Diffusion Model, You will need text projection (https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/text_encoders) and https://huggingface.co/Comfy-Org/ltx-2/tree/main/split_files/text_encoders gemma 3 12b as this is a transformer-only model.
Amazing.
I just tried a lots of INT8 model in image generation after comfyUI natively support INT8. and suddenly realized that the LTX could also use INT8.
i used in use GGUF Q5KM (RTX 3060). (15GB File size)
Same prompt same resolution and duration, Just changed this INT8 (22GB File size), the generation time is way faster....
originally, the first 8 steps generation is 18s/it., now is 6.8s/it.
amazing.
The 3 upscale step also jump from 63s/it to 40s/it.
now i am testing the memory efficiency. Not sure if i can generate longer video without error.
Same runtime for me rtx3060 12gb 32gb ram, I've done 10 second generations, haven't pushed past that yet
@tsolful I'm also on 3060 plus 32GB ram before, then i decided to buy an SODIMM adapter to make use of my old SODIMM, so now is 64GB RAM.. Depends on resolution. for 864 X 1536, i can do up to ... i think 30 second.
32GB RAM is just not enough for swapping....and you will eventually use SSD cache, which is way...tooo slow for inference..... 48GB would be great.... it just didn't reach 50GB RAM for 30second...