# ONNX models (Inference) **URL:** \ **Category:** Flame Questions\ **Tags:** ai, onnx\ **Created:** [August 1, 2024, 5:12pm UTC](https://forum.logik.tv/t/onnx-models-inference/11148 "2024-08-01T17:12:58Z")\ **Posts on this page:** 9\ **Page:** 1
**Author:** ![bryanb](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.logik.tv/bryanb/32/136_2.png) [@bryanb](https://forum.logik.tv/u/bryanb)\ **Post date:** [August 1, 2024, 5:12pm UTC](https://forum.logik.tv/t/onnx-models-inference/11148/1 "2024-08-01T17:12:58Z")
Anyone that has tried the new inference node (flame 2024.1): – any ONNX models to recommend? – where do you go to check out available models? This is all new to me. ---
**Author:** ![philm](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.logik.tv/philm/32/9286_2.png) [@philm](https://forum.logik.tv/u/philm)\ **Post date:** [August 1, 2024, 5:53pm UTC](https://forum.logik.tv/t/onnx-models-inference/11148/2 "2024-08-01T17:53:28Z")
@bryanb - here is the [homepage](https://onnx.ai/). It’s quite a deep rabbit hole and not everything is applicable to flame. ---
**Author:** ![Josh\_Laurence](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.logik.tv/josh_laurence/32/1243_2.png) [@Josh\_Laurence](https://forum.logik.tv/u/Josh_Laurence)\ **Post date:** [August 2, 2024, 3:02am UTC](https://forum.logik.tv/t/onnx-models-inference/11148/3 "2024-08-02T03:02:50Z")
I’m also headed down that same rabbit hole (but I have no clue). @tpo posted a General Improvement a few days ago with a request for the Inference Node to handle RGB inputs ([fl-03327](https://feedback.autodesk.com/project/feedback/view.html?cap=5afe6c84-5cb3-447a-b36c-cbd7f0688f84&uf=56772e3a-aba5-49f0-b888-27f63188aed4&slsid=59106b1f-15b5-4740-8a65-c0a920ff605e)) which also has some interesting links - one to a frame IO presentation that’s just bonkers. And another to one of the onnx links from [huggingface.io](https://huggingface.co/Xenova/vitmatte-base-composition-1k/tree/main). Which gets an error in Flame… Here’s the link to [Hugging Face](https://huggingface.co/) which seems to be a place to find this stuff… Maybe there’s more sites I’ll find once I really dig into this one. I’ve tried a few onnx files and get the same error, but, again, I’m just scratching the surface by finding any .onnx file to download and trying to load it in the node. Usually resulting in an error message. I love the idea of generating mattes the way these examples show, but honestly need a lot more time reading up on this stuff to figure out how to leverage them inside that node. I watched @Jeff’s FLC video on the node, and as usual, it’s informative and very lucid, but I’m left with more questions than answers… ---
**Author:** ![Fabi](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.logik.tv/fabi/32/14709_2.png) [@Fabi](https://forum.logik.tv/u/Fabi)\ **Post date:** [August 2, 2024, 7:25am UTC](https://forum.logik.tv/t/onnx-models-inference/11148/4 "2024-08-02T07:25:18Z")
it’s all new to me too but i have found this site and am trying my hand at it. > **[Models - Hugging Face](https://huggingface.co/models?sort=trending)** > > We’re on a journey to advance and democratize artificial intelligence through open source and open science. I am also in the process of finding a reasonable upscaler > **[OpenModelDB](https://openmodeldb.info/)** > > OpenModelDB is a community driven database of AI Upscaling models. We aim to provide a better way to find and compare models than existing sources. ---
**Author:** ![finnjaeger](https://avatars.discourse-cdn.com/v4/letter/f/e99b99/32.png) [@finnjaeger](https://forum.logik.tv/u/finnjaeger)\ **Post date:** [March 10, 2025, 11:21am UTC](https://forum.logik.tv/t/onnx-models-inference/11148/5 "2025-03-10T11:21:53Z")
I also sturggle to get any ONNX models to load, just goes " ![image](https://us1.discourse-cdn.com/flex020/uploads/thedepartmentofexternalservices/original/2X/2/25cfd140729e446555dd3d2ab541890e70df5068.png) and in the terminal log \*\*\*ErrorMsg: Failed to load model because protobuf parsing failed Update: make sure to install git LFS before you clone from huggingface or else the onnx files will be empty.. Hugging Face uses Git LFS for large files, so you need to ensure you download the actual model. ### Solution: 1. **Install Git LFS (if not installed)** - On macOS: ```auto brew install git-lfs ``` - On Linux: ```auto sudo apt install git-lfs # Debian/Ubuntu sudo dnf install git-lfs # Fedora ``` - On Windows: Download and install from [Git LFS website](https://git-lfs.github.com/) 1. **Enable Git LFS** (Only needed once) ```auto git lfs install ``` 1. **Clone the Hugging Face Model Repo with LFS Files** ```auto git clone https://huggingface.co/your-model-repo ``` 1. **Manually Pull the ONNX Model** If you already cloned but only got the pointer file: ```auto git lfs pull ``` ---
**Author:** ![finnjaeger](https://avatars.discourse-cdn.com/v4/letter/f/e99b99/32.png) [@finnjaeger](https://forum.logik.tv/u/finnjaeger)\ **Post date:** [March 10, 2025, 11:49am UTC](https://forum.logik.tv/t/onnx-models-inference/11148/6 "2025-03-10T11:49:42Z")
but yea… not having any luck, this isnt fun. All models i have downloaded from cattery worked oob in nuke ![image](https://us1.discourse-cdn.com/flex020/uploads/thedepartmentofexternalservices/original/2X/2/2a69d5955fefda3a66470f935e873ece0973bc0e.jpeg) ---
**Author:** ![Ben](https://avatars.discourse-cdn.com/v4/letter/b/df788c/32.png) [@Ben](https://forum.logik.tv/u/Ben)\ **Post date:** [March 10, 2025, 12:12pm UTC](https://forum.logik.tv/t/onnx-models-inference/11148/7 "2025-03-10T12:12:23Z")
Ive just got on 2025 and not having much luck either. Feels like another doomed new feature ---
**Author:** ![allklier](https://sea2.discourse-cdn.com/flex020/user_avatar/forum.logik.tv/allklier/32/4309_2.png) [@allklier](https://forum.logik.tv/u/allklier)\ **Post date:** [March 10, 2025, 12:14pm UTC](https://forum.logik.tv/t/onnx-models-inference/11148/8 "2025-03-10T12:14:42Z")
Just to test, grab some of the files of the LogikPortal. The DepthAnythingV2 is a good test case, as a pretty stable model. There are still a lot of variables around ONNX models (as opposed to tools that use models internally), as you need to feed the model with the image in the exact color splace, channel order, etc. that it was trained in, or even if it runs, it will be gibberish on the other end. Also, these models are extremely memory hungry and have not been optimized for your average machine like the models Boris or ADSK embed. The same is true for the Cattery models btw. I’ve run through all the available ones the other day, and half of them are hard to use or do not produce usable results for some of the same reasons. While the thought of loading late-breaking models of hugging face and just plugging them into batch is intriguing, I think the reality will be more that folks with applicable experience need to test and massage them, then publish via LogikPortal for daily consumption. The good thing is that we do have the platform, so this can become an extension of Flame like matchbox shaders eventually. Which also required a few folks to roll up their sleeves and do the dishes first. ---
**Author:** ![chlowden](https://avatars.discourse-cdn.com/v4/letter/c/c2a13f/32.png) [@chlowden](https://forum.logik.tv/u/chlowden)\ **Post date:** [May 12, 2026, 11:45am UTC](https://forum.logik.tv/t/onnx-models-inference/11148/9 "2026-05-12T11:45:04Z")
Has anyone compared the Auto-matte 2027 using other ONNX files and the ADSK models? I found RMBG-2-Matting onnx files > **[ZQL9711/RMBG-2-Matting at main](https://huggingface.co/ZQL9711/RMBG-2-Matting/tree/main)** > > We’re on a journey to advance and democratize artificial intelligence through open source and open science. It takes me about 7mins to load the 1GB model and analyse a 200 frame 6k shot. These models are 1024x1024. The main model did a good job isolating a man from a backgound. Much better than the body isolation node. Slightly better in my case than the ADSK model, but frankly it is debatable. The couple of other models I tried gave very different / unsatisfactory results . Anybody tried to get the SAM3 model to work ?