TinyTerra / ComfyUI_tinyterraNodes

A selection of nodes for Stable Diffusion ComfyUI
GNU General Public License v3.0
421 stars 45 forks source link

TTN samplers incompatible with Flux hybrid workflows #149

Closed rjgoif closed 1 month ago

rjgoif commented 2 months ago

When trying to use a TTN Ksampler in a flux workflow, it throws this error:

ttN KSampler_v2

It seems that models and clips are mixed and interconnected between SDXL Base, SDXL Refiner, SD1.x, and SD2.x. Please verify.

For now, I will jank it together with an upscaler that does 1 step and 0 denoise to get access to the XYZ plotter, but it would be nice if this could be fixed to work with Flux (and maybe be more model-flexible for future developments).

Attached example workflow: TTN error

TinyTerra commented 2 months ago

I downloaded those same models/clips and used your workflow and the generation works fine... The only difference being that i used Integer primitive / Empty SD3 latent instead of the comfyroll node.

Is your comfy up to date?

rjgoif commented 2 months ago

Ah, the empty SD3 latent node does the trick. Funny, in my normal Flux workflow, I use the Comfy node but run the latent through the SamplerCustomAdvanced node standard to Flux workflows. I thought that was just an aggregator, but it clearly does some latent processing as well.

Problem solved, I guess!

I will point out that I wanted to do this in order to access the XY capabilities of the TTN Ksampler (still the best XY in town, by far). How hard would it be to spin off the XYZ into a standalone node, that takes an image and outputs the image + the grid, but then also has the input for XY configuration like the current KSampler does?

So 2 inputs: Images, adv_XYplot 2 outputs: Images, Plot Image

And it would give access to the XY plotter to all upstream nodes, like it currently does.

If it's a pain to make this happen, then what we have now will work. Feel free to close this issue when you weigh in on it.

TinyTerra commented 2 months ago

Interesting idea.

I've pushed an update adding this functionality (albeit hastily). Let me know if that's what you had in mind / if you come across any unexpected issues. I haven't tested it very thoroughly.

rjgoif commented 1 month ago

Looks like it's working just as I imagined!

You now have an even better generalized XY tester for Flux than Matteo's, which is a pretty high bar to clear. I didn't even notice the combo picker you put in there, with the range option to test LoRA training step progression - amazing!