...

#5
by premiumfp2026 - opened
This comment has been hidden (marked as Resolved)

Still working on both. Instant action starts at a random point in latent space and works backwards towards the input image like it's an end frame. That ones for just animating an image differently. The latent anchor still caches the input image as a frame 0 start so if you increase it's strength it will deviate less. Its just a different kind of i2v.

I2v_v2 is supposed to just be normal start input image but there is a problem with fade on upscale pass. Both produce decent outputs still.

Hi, Thanks for the model and appreciate your work, quick question, Is there any specific trigger words I can use to enhance the motion and prompt adherence? using natural language doesn't work mostly especially when it comes to ending the NSFW act of a couple.

I2v_v2 is supposed to just be normal start input image but there is a problem with fade on upscale pass. Both produce decent outputs still.

Does the i2v_v2 basically replace the "triple sampler" workflow then when it's done?

premiumfp2026 changed discussion title from I2V v2 and InstantAction workflows to ...

Does the i2v_v2 basically replace the "triple sampler" workflow then when it's done?

Yeah the triple sampler does 8 extra steps for very little improvement, mainly because the latent gets more manipulated by upsample than I thought, those inpainted details all get stretched out and diminished. Worth maybe still combining it all for a max quality WF though I'd have to direct test them. It still works you can try v2 for now.

man for the love of god can you share a basic I2V workflow? No upsampling, no weird nvidia RTX nodes that dont install, no spaggeti comfy. A very simple with 2-3 loras workflow and thats it.

man for the love of god can you share a basic I2V workflow? No upsampling, no weird nvidia RTX nodes that dont install, no spaggeti comfy. A very simple with 2-3 loras workflow and thats it.

To be fair the Comfy template as well as LTX's own template work absolutely fine. That's what I've been using while keep an eye on this repo. Just replace the model and distill lora.

Yeah these are mostly in-progress sampling experiments btw, not recommended for normal use and all are being updated. As far as over-complicated though they're pretty simple next to some of WFs on civit, LTX is a node jungle to set up.

Sign up or log in to comment