LTX2.3 FP4 Models
Because no one uploads FP4 here. and they should (fp4 works on 3xxx and above for vram savings.)
LTX2.3 Distilled FP4ME - Distilled FP4 Mixed Extreme - 14.1GB**
LTX2.3 Full Dev FP4ME - Full Dev FP4 Mixed Extreme - 14.1GB**
LTX2.3 Official NVFP4 - From Lightricks - Updated 3/17/2026 - 21GB
LTX2.3 Dev NVFP4 - from Hippotes / LTX-2.3-various-formats - 18.15GB
LTX2.3 Dev FP4 STFO - from Kijai's Transformers only Scaled FP8 - 16.63GB**
LTX2.3 Distilled FP4 - My first FP4 - Transformers Only - 18.15GB**
** Are transformers only, or I broke vae/text projection :) needs seperate Vae and Text projection downloads.
I have a fix for lora issue on FP4ME models. Comfy Bathroom, custom lora loader with presets as well as custom config. https://huggingface.co/MrReclusive/LTX-2.3-FP4
Diffusion model loader/unet loader seems to get confused with these, use the checkpoint loader even though it has no clip/vae
Description
Extreme FP4/FP8 Mixed
FAQ
Comments (11)
Karmaşık ve işe yaramaz bozuk nodeler ve tekrar yapılandırma zinciri ile uğraşmaya gerek yok! workflow güncellenirse tekrar deneyeceğim
any chances of an int8 version for 3xxx gpus?
check with the people posting ggufs, i haven't done much of anything with int8 outside llm's and from my understand int8 is horrible for image/videos models. the fp4's as far as I know do still work on 3x series cards for memory savings, just no speed boost that you would get on the blackwell cards, i don't even have a blackwell card, im on 4x series, so only memory savings, well, do get some speed boost vs fp8 or higher because don't need to swap during generation because my extreme varient is 14gb.
@MrReclusive666 yeah i'm using it right now and the memory save is nice, i'm using less than with a q4 gguf, but for now the quality is way worse. do you have a workflow i can try?
@ZelashZelash all my videos have workflows attached. they are all multi gpu setups. first thing I would suggest, turn off any lora's, this is a big issue with my extreme fp4 model, working on that, i don't really run that one with lora's. and it really prefers low steps, with dpm++ 2m sde heun gpu it runs at 4 steps, quality degrades at 5, euler a likes about 6 steps. (these are either beta11 or sgm uniform) this is my custom sigma i've been runing with dpm++ 2m sde heun gpu, 1.0, 0.925, 0.725, 0.421875, 0.0. play with the 0.925, between 0.95 and 0.90, depending need. to get that small size its very compressed so very fickle.
@MrReclusive666 thanks, i'm having good quality results with your workflow at 4 steps even with loras! i was using a workflow with a 8 steps low res first pass and then 4 steps for refining. i guess the latent upscaling messes up with the quality a lot
TBH I am using INT8 and I have used FP4 LTX 2.3 (not this guys models, i downloaded one and it was terrible output, it was the extreme one, im sure one of the others would have worked fine but i got the fp4 elsewhere after) Anyway, I feel the speeds between INT8 and FP4 seem comparable at least if you dont have enough memory to hold the INT8 file at once, I have 32gb vram and 8gb ram so the FP4 fits nicely (using cache none to dump other models and cache) but the INT8 is too big so I think it slows it down for me to where the end result in speed is similar to FP4.
Just my 2 cents, would be interested to hear your experience if you've tried the regular INT8 there's a distilled version on silveroxides quantops huggingface and winnougan has just converted a dev version.
@MrTitsworth id be curious to, i just know ive read a lot that dit models don't do well in int8, and was that a typo or are you really running a 5090 on 8gb ram?
i guess its a trade now, prices are as equally insane.
@ZelashZelash hey, just wanted to follow up on this. i run custom sigmas for stage 2 upscale, euler ancestor.
for quick run - 0.91218, 0.725, 0.0
for more detailed run - 0.91218, 0.909375, 0.725, 0.10546875, 0.052734375, 0.0
that second one helps when going from reaaly low res base, as it helps add detail and fix anatomical issues.
the reason for the reaaly crazy low number ones, is for some reason this, 0.421875, is mathematically significant, and those low numbers i based on that.
@MrReclusive666 Nah im on a 3070ti and no isues with the models.
I've been comparing, INT8 for me is similar in speed to FP4 models on my system where the FP4 is smaller so loads faster but the INT8 is optimized but is larger, so the fp4 can be slightly faster but the INT8 is better quality as its uncompressed.
If anyone is interested in INT8 just google silveroxides and Winnougan has made some conversions of the Dev and Transformers version of distilled so far.
@MrTitsworth yeah, its all about need, i made these 14gb fp4's because I wanted the model fully on card with no swap while still getting decent resolution/length outputs.
and the main reason is my 4090's are on 1x pcie risers (this machine was built for 3d rendering), so active swapping is a huge no for me, so 30gb models, i don't even look at.