CornLogic

/

10EROS-INT8

like 14

Hi, I'll see if I can get to it. Full disclosure I am working remotely so honestly you could have yourself up and running in five minutes if you just rewire your workflow a little. Waiting for me to get to this and then uploading over a slow connection will take far longer.

Make a backup of your workflow before you do so if you are unsure of wiring you can switch between tabs to double check.

all vae's, text projection are found here: https://huggingface.co/Kijai/LTX2.3_comfy
it includes a picture showing you what nodes to use on the model card. If that helps.

Thank you for the quick reply.
That was very helpful, as I wasn't sure which text projections file was the right one.
It does work with the minor workflow adjustments you suggested, but with the non-DMD version 1.4 checkpoint file, I was able to use that same checkpoint file for both the VAE Loader and Dual Clip Loader; I found that setup a bit easier to understand.

Hi.

BF16 and FP8 models are usually packaged as full checkpoints all-in-one including the VAEs, text encoders, (and in the case of LTX2.3 and derivatives the vocoder).
This is done because all that additional material is usually stored in BF16 format for maximum quality, and usually those are never quantized to lower precisions, because the loss of quality is big and the space and memory savings are low. Usually only the Transformer tensors are quantized.

Now, Comfyui can load mixed precision BF16/FP8 checkpoints natively, but I’m not sure it can load mixed INT8+otherQ files (and same with other more exotic quantizations like GGUF)
To read a model that includes a INT8 quantized transformer and a full quality BF16 VAE and encoders packaged together you might need to use a specialized loader node.

Edited: since July 2026 Comfyui does support loading mixed int8/bf16 checkpoints

It’s much easier to give the transformer only when publishing a quantization other than FP8, and then load the VAEs and text encoder independently

Here you have a screenshot of how you can replace the regular full 10Eros checkpoint loading nodes with CornLogic int8 quantization instead and independent aux files:
https://huggingface.co/TenStrip/LTX2.3-10Eros/discussions/63#6a6d278757d2c16b7a491f17

I'm in a rainforest right now so finding a ssd was interesting but its uploaded

re what LokkenJP said:
That's incorrect. A bf16/int8 checkpoint works fine. You might get a clip text encoder not in checkpoint error but it then goes on to complete with no issues. You can make one in convert to quant or you can just pack together the individual components together with a python script as long as you don't trip over a few gotcha's along the way like audio vae/vocoder or not porting across the metadata from all the pieces. You'll get a bit identical safetensor checkpoint.

=== diff ===
A: I:/ComfyUI SF/models/checkpoints/LTX 2.3/10Eros_v1.5_DMD_int8_convrot.safetensors
B: I:/ComfyUI SF/models/checkpoints/LTX 2.3/10Eros_v1.5_int8_DMD_chpnt.safetensors

Common tensors: 8411
Only in A: 0
Only in B: 0
Shape/dtype mismatches: 0

Headers are identical (names, shapes and dtypes all match).

slot tensors size prefix how

model 6908 23.31 GB (none) inferred from reference (48/48 sampled keys matched)
text-projection 4 2.15 GB (none) inferred from reference (4/4 sampled keys matched)
vae 170 1.35 GB vae. inferred from reference (48/48 sampled keys matched)
audio-vae 1329 347.77 MB (none) inferred from reference (48/48 sampled keys matched)

total 8411 27.16 GB

You’re right. I’ve just checked and indeed since 1 July update about a month ago with the 0.27 convrot update, the standard Load Checkpoint node of comfyui also supports mixed int8/bf16 safetensors files, much in the same way as the mixed floating point loading was added last year.

So my info was outdated. Sorry about that 👍🏻👍🏻👍🏻

I’ve edited my post accordingly