EDIT: Updated to a 2-Stage workflow to increase quality, gen time, and usability.
LTX notoriously treats your prompts as suggestions rather than as instructions. So quit beating your head against the wall trying to prompt better! The path to success (and less frustration) is to be able to see the results of many seeds quickly, and when you find one you like, to be able to take that promising "golden" seed to finalization. I call this SEED HUNTING, and it's your new best friend in AI filmmaking and content creation. This workflow also has a functionality I call MULTIROLL, allowing you to regen Stage 2 and Stage 3 as many times as you like, getting slight variations that help you build iteratively on your perfect result, without requiring you to always be genning from scratch and wasting your valuable time!
Watch the video below that covers both basic to advanced use-cases for FFLF. It is part 2 in my tutorial series on AI filmmaking, where I deep dive into every topic from this, to audio mastering in Davinci Resolve/Audacity.
NOTE: This is a FIRST FRAME LAST FRAME workflow, it is not designed for single inputs. If you want to gen from a single image/video input source, use my other workflow:
https://civarchive.com/models/2676452/ltx-23-seed-hunter-multiroll-workflow
This workflow uses ComfyUI-mxToolkit for Sliders, and requires Nodes 2.0 to be turned off in comfyUI's settings for it to work.
Description
Updates from a 3-stage workflow to a 2-stage workflow, increasing quality and gen speed.
FAQ
Comments (5)
Thanks for this, getting smooth results on the first try. HOWEVER!!! I used your recommendation in the workflow of "Continuous fluid shot" and the video rendered with thick white fluid spraying all over the subject...
"a continuous fluid shot from the beginning to end, the camera pulls out to reveal the chubby Caucasian female with dark hair."
It was just a quick prompt to test the first/last, and it worked... But she's covered in thick white fluid. I took out the "fluid" and it worked :D
if you're being serious or not, that is hilarious. thanks for that mental image lmao
@foxydits I'm serious. I just had her walking toward the camera, and all 3 preview windows showed 40 gallons of thick white liquid shot at her. I admit, it was a crappy prompt, but took me far too long to workout what was going on... Checking the metadata of the start image to see if there was anything in there etc... I was genuinely impressed when her eyes flickered down to look at it briefly.
Steps... Where the hell do I set them? I'm trying to use the dev instead of distilled and I want to increase the number of steps. But I cannot find the setting in any of the subgraphs. Have you hard-coded the steps somewhere? OR am I missing them everytime?
They're manual sigmas from the original LTX team's distilled workflow. Inside each sampler subgraph you'll see them. First stage has 8, 2nd has 4. You'd delete the manual sigmas and replace with BasicScheduler node.