Update README.md

README.md

@@ -1,8 +1,10 @@
-LTX2 i2v - Lora+Refinements.
+LTX2 simple i2v. 
 
-Many nodes to change/edit lora strengths and refine second pass results for high-effort generation. Use res_2s for best quality. Hard load audio from the first pass if it's good, it will usually be better than the second pass.
+Uses res2s and usually lower distill strength with fp8 model, 0.75-0.9 depending on application. Stack as many loras as you can even if barely related to the concept, lora stack drowns out the base model noise and makes the output more stable.
 
+Use 33-38 preprocess node compression strength to increase motion. Get the best motion and audio from the first pass, cancel and reroll seed if bad. If preview is kind of good stop it and refine with prompt and lora weight mixes. 
 
+Second pass uses audio directly from first seed to track and half strength distilled upscale pass based on the full size input image for max quality. Only way to get very good clear visuals is with the half distill, but passing audio latent into the half distilled sampler ruins it, so this is the neat trick.
 
 
 
@@ -20,30 +22,3 @@ Many nodes to change/edit lora strengths and refine second pass results for high
 
 
 
-
-
-
-
-
-
-
-
-
-OLD LTX i2v workflow
-
-Differences from default:
-
--Running distilled lora 1.0 on both passes with fp8 model - allows control over the distill strength if the effect is too strong. The fully distilled model was supposedly fixed, can be replaced with that. 1920x1080 241 frames possible on a 5090.
-
--Using KJ's seperated models for easier offloading.
--- https://huggingface.co/Kijai/LTXV2_comfy/tree/main
-
--dpmpp sde sampler first pass, LCM second pass. SDE adds much better motion detail, LCM adds less for the weaker second pass, only upscaling it.
-
--4 step sigmas on upscale pass, this is what the LTX2 dev post suggested. Not sure why default WF only has 3 steps.
-
--First audio latent is discarded, audio output comes from 4 step pass. This is to correct for the SDE first pass, which doesn't produce good audio. Tradeoff. The second pass basically retracks the video, it can be innaccurate to prompt but that might just be the model.
-
--Latent normalizer (LTX video github): seems to fix bad audio and burned video outputs
-
--Image upscaler before preprocessing. Full size image sent to preprocessing. Use a 2x-4x upscaler the image is not 1280 on a side for fidelity, or use a sharpener to sharpen results. They will come through and also have a visual impact on the output.