TLDR: A multipass process which potentially improves outputs, particularly from the distilled model.
Stemming from my experiments with custom sigma curves, I've stumbled upon an interesting technique which can change the dynamics of an initial video, increasing overall motion, improving fine detail, or adhering to prompts a bit better.
Before you get too excited, however, know that it will typically only do one of these things at a time, and sometimes will do the exact opposite. Even when it works, it's only a subtle improvement. This is not the panacea for distilled LTX's shortcomings, but I've found it useful enough that it's become my go-to generator.
The way it works is a tad convoluted and is more thoroughly explained by instructions in the workflow itself, but it essentially boils down to the simple expedient of incorporating two extra samplers, powered by the RES4LYF Clown Schedulers, one at low steps and one at low denoising. These help preserve the creativity and obedience of LTX's early steps while preventing too much degradation in overall quality. They are driven by an initial sampling pass at the default values in order to generate usable audio (and an optional template video), and then are further refined by the standard spatial upscaling process.
Note that these extra samplers do naturally make the process slower, but since this is all happening with a low step count before the upscaler, it generally only takes about half again as much time as the default workflow.
I should also add that this is just an expansion on my general day-to-day workflow and it therefore comes equipped with several functions which can interact with the extra samplers. The greatest impact is on T2V, but I2V, A2V, and V2V are all present in various capacities. I have also included a few switches to partially or completely bypass the extra samplers to make it easier to run comparisons. Or indeed, to use this as a normal workflow, if you happen to like the setup.
Which brings us to the clarification that I still have an irrational hatred for subgraphs, so don't necessarily be intimidated by the Great Wall of Nodes, as most of it is just backend. I have included copious notes to try and demystify all adjustable parameters and I advise making judicious use of Comfy's Hide Links function located in the lower right under the minimap.
Required custom nodes:
ComfyUI-GGUF: https://github.com/city96/ComfyUI-GGUF
rgthree-comfy: https://github.com/rgthree/rgthree-comfy
ComfyUI-KJNodes: https://github.com/kijai/ComfyUI-KJNodes
ComfyUI-VideoHelperSuite: https://github.com/Kosinkadink/ComfyUI-VideoHelperSuite
RES4LYF: https://github.com/ClownsharkBatwing/RES4LYF
As always, feel free to incorporate or repurpose or whatever you like with this workflow.
My thanks again to all those who have uploaded workflows for LTX 2.3. I continue to learn a great deal from you and this certainly wouldn't have been possible without referencing your excellent work.
