Please note: this is not a full LTX workflow and cannot generate videos on its own.
UPDATE: While browsing custom nodes for a different project, I encountered an additional tool that provides a visual method for making your own sigma curves. This is by no means a necessary addition, particularly since this workflow wound up being something of a folly anyway, but I found it interesting enough to showcase nonetheless.
ComfyUI-NKD-Sigmas-Curve: https://github.com/Nekodificador/ComfyUI-NKD-Sigmas-Curve
In addition, with appreciation to @redlittlerabbit for making me realize this, I have been terribly unclear as to what this is all in service of, at least to those unfamiliar with the default LTX 2.3 distilled workflow. So, with the caveat that I've only just learned about this stuff myself and am still wrapping my head around it, let me try to explain in typical overwrought fashion:
In most workflows for the distilled version of LTX 2.3, there are typically two SamplerCustomAdvanced nodes into which are connected all the bits that make a sampler actually work, including two ManualSigmas nodes. The strings of numbers in each of these sigma nodes are what determine the number of sampling steps, as well as the level of denoising at each step. Each number (except 0.0) represents a step and the value of each number represents the level of creativity the model is permitted for changing the image in that step. Hence why the sigma values produce a curve on the graph. You generally want high creativity initially that is then reduced over time to eventually resolve into a stable image.
So why change these values at all? My original intent was to glean more quality by slightly increasing the number of steps, but that didn't really seem to work. Thus, while testing does continue, I now view this project as more of a toy than a legitimate tool. It's fun to see what different curves make, but ultimately, none of them seem to work better than the defaults. Which makes sense, given the defaults were presumably put in by the LTX team themselves.
That said, there is one potential use case I noticed right before uploading this update in that certain sigma curves which drop in "chunks" (ie: two steps in the 1.0-0.9 range, then, say, three steps in 0.8-0.7, etc...) tend to retain more motion from the initial sampling steps. I attempted to demonstrate this phenomenon in the example video, which is generated from the same seed as the prior examples, though I admittedly had to change a couple prompt details to correct some disobedience. Regardless, the thing to take away is that while the quality appears to be worse overall, the added motion could perhaps prove useful in certain applications. Especially I2V.
ORIGINAL POST: So I sort of worked out an easy, if slightly janky, way to get new manual sigma values which can then be applied to the distilled workflow for LTX 2.3. My intent was merely to adjust the sampler step count so I could squeeze out a little more quality without jumping all the way up to the dev model, but it turns out this method can also produce values which mimic other schedulers, which has been rather interesting to play with.
Now for it's intended purpose of adjusting steps, it wound up making virtually no difference to quality, so don't expect this to magically fix LTX's shortcomings. That said, I certainly haven't tried every mixture of sampler and scheduler, so maybe there's some combination which will actually improve things. But, despite how I categorized it, I'd still think of this less as a tool and more as a toy.
Instructions are included via notes in the workflow. And I actually do have the slightest idea what I'm doing this time, so I can (probably) answer any questions here as well.
The required custom nodes are:
ComfyUI-GGUF (if using a GGUF model): https://github.com/city96/ComfyUI-GGUF
RES4LYF: https://github.com/ClownsharkBatwing/RES4LYF
As ever, feel free to repurpose/expand upon/incorporate/whatever this workflow in any way you might wish.
