So just creating a general OFX thread. I’ve made a lot of OFX over the last few months and I just like you all so much I am letting you have some fun too. Here is a new color match with AI ability it can match at various levels and the AI well is AI so roll the dice. Also based on Joels Matchbox I built a new OFX additive keyer. Please help me out and provide some thoughts because thats how we make this better.
I keep updating them and getting better so I will just leave the github home so everyone grabs the latest.
Just tested Upscale Native on my M2 Ultra / Flame 2027 setup.
Tested HD → 4K (Factor 2x) on a 10-frame clip. SeedVR2 3B processed 1920×1080 → 3840×2160 in ~759s. The stored result in the bg/ folder is genuinely impressive — 19.35x sharper than a simple Lanczos resize (Laplacian variance comparison). Side-by-side crop comparison clearly shows the AI upscaling adding real detail vs interpolation.
One thing though — I couldn’t see the result in the Flame viewer (F4). The node status says “upscale stored” but the output looks identical to the resized input. I had to go into the bg/ folder and pull the EXR files directly to verify the result was actually there and correct (it was). Could this be a technical issue with the plugin not outputting the stored result during render? Or is there something I need to do on the Flame side to make it read the stored result? Anyone else seeing this?
Tested Additive Key Native with Spill Difference / additive mode — generally working well! Edges are picking up the tint from the background nicely. Didn’t have the ideal material to push it really hard (need a proper blue/green screen plate with fine hair detail to stress test it), but from what I can see the additive blending is doing its job…Great work @space_monkey
The non-AI version seems to work well! It matches colors and tones nicely, and I think it would be great for matching two similar shots. I don’t have suitable footage to test right now. I tried two very different shots, and it copied the colors and tones from the reference, but the result didn’t look very good because the scenes were too different. It would be better to test it with two similar shots.
The AI mode doesn’t seem to be working yet. It looks like the system is still using LTX instead of Qwen. The required Qwen model and LoRAs are already installed, so maybe the AI mode just needs a small fix.Thank you @space_monkey
That is the result is an OFX limitation and I will circle back to that issue and try to python update it , just frame forward or back and it will refresh. If you all could vote on the OFX requests I made on ADSK that might help move it forward.
Just posted a few more, back ground which you can do backgrounds in with Nano and a few other models, matte works for the holdout of the background. And since you’ll need to blur it in depth I also added the Z-Depth OFX. Have fun!
thats a good one, the challenge with OFX is its limitations you cant have multiple outs but I could create an internal toggle to select which pass to run , let me look into it
Here is a fun knew one that I thought you all would like, stop trying to find that element and generate it, the text box you can describe what you want.
IMPORTANT: Please install the hotfix py since that was deselecting clips (it was to refresh the render status) now it only pings when render is done and turns off.
Multiple node instance all linked to the same cache files. Instance 1 does the compute, Instance 2+ just reads cache to outputs.
Some small json file for the instances to coordinate on state. Use a hash key parameter to tie them together as a set. Can even use Flame link feature to sync hash key.
Since we live in a time where it’s raining OFX nodes daily for breakfast…
Sometimes you just want to run a model. You know like Nano Banana. Without Fuzz.
Yes, there are cool ComfyUI workflows, and they’re powerful and cool. But you need to install a whole bunch of stuff, worry about security, how to get files back and forth…
Beak wrote a nice frame work for running larger workflows right out of batch: BCE
But what about a one-node setup. A simple GenAi or InPaint.
There’s now an OFX to run partner nodes. It uses the Comfy Router API. You need to buy tokens through Comfy Org, if you don’t have any yet and get an API Key. But most importantly, you don’t need ComfyUI installed locally to use this. Just this OFX node.
Once during setup you have to enter the API Key. It will be saved in a per-user config file.
The node calls out to the API to render it and then caches the result locally. Color space transforms will be handled on both ends based on your settings. You can provide up to 5 inputs for reference.
Right now this only supports two nodes:
Krea3 (new image edit model, released very recently)
Nano Banana Pro (image generation & edit)
These are both image, not video models. So the node repeats the first frame for the duration of the batch. Video models and other models can be added over time as there’s interest. Ping me, and we’ll see what’s possible.
In the right most column it shows you the cost per render based on your settings.
This does use a published API. But as with all APIs, their owners may change them without warning. So these nodes work as is. The providers behind these are the same ones Comfy Cloud lists, a combination of Google, fal.ai, Runware, etc. Read the license/readme. This is a ‘bring your own key’ solution, no other warranties made here.
If you need something more sturdy for your studio, I recommend iNodes or other commercial setups.
May it make your day easier if just need to run a node without too much fuzz.
friday fun I exported a bunch more, I really like the motion estimation which you can export and it automatically imports the tracked masks. I did make some changes to the engine so be sure to do your updates.
