Still amazingly fast renders, but..
I am having a general quality issue with LTX-2.3.. It is similar to the way I was feeling about LTX-2, and I am wondering if I am doing something particularly wrong, but on both, the output quality is nothing even close to usable. I am using the ComfyUI workflow with the ltx-2.3-22b-dev.safetensors" checkpoint model and a "gemma_3_12B_it.safetensors" text_encoder, rendering output at 1920x1088, but the output is just not hitting the mark for quality. It still seems to struggle with detail, warping, and most noticeably teeth anomalies. I went through all of the adjustments I could think of to try to produce higher quality output, but was surprised that I just can't seem to get there. Regardless, even at 50 steps, it is insanely fast, but I think I must be missing something for the image quality end of the spectrum. I also tried using various input images with different resolutions and varying amounts of detail, but just can't really seem to get a clean output. If anyone has any absolute "Do's" or "Dont's" I'd love to hear them.
i really think this needs a detailer like ltx2 had, and the ltx2 detailer on 2.3 is less then Ideal.
It feels like a toy, just play with it. The model is still too poor, even the first generation version of Ke Ling from a few years ago can't compare
i really think this needs a detailer like ltx2 had, and the ltx2 detailer on 2.3 is less then Ideal.
Hopefully we'll get a 2.3 detailer lora before next week.
I'm not sure what it needs.. Or what I need to do, but it seems like regardless of setting the output to 1920x1088, it just produces relatively blotchy output. I have been looking at the "Resize Images by Longer Edge" node in the workflow, but haven't seen anything suggesting to change that value dependent on output resolution settings. It has been set at 1536 since I loaded the workflow and through every test I have run so far. I have seen some other people seemingly producing some really great results, just can't figure out where I'm going wrong.
I'm not sure what it needs.. Or what I need to do, but it seems like regardless of setting the output to 1920x1088, it just produces relatively blotchy output. I have been looking at the "Resize Images by Longer Edge" node in the workflow, but haven't seen anything suggesting to change that value dependent on output resolution settings. It has been set at 1536 since I loaded the workflow and through every test I have run so far. I have seen some other people seemingly producing some really great results, just can't figure out where I'm going wrong.
Pls check ur sigma values they're different from ltx 2 and all other preprocesser values. I recommend to use the wfs from the ltx repo but to be honest I haven't tried the comfy built-in templates not and I won't, because I generating through kijajs transformers only. U could/should try this too. My outputs are splendid. Especially with ltx guidance parameter nodes.
It feels like a toy, just play with it. The model is still too poor, even the first generation version of Ke Ling from a few years ago can't compare
Ure from the Kling team right?? Never read such bs before
Pls check ur sigma values they're different from ltx 2 and all other preprocesser values. I recommend to use the wfs from the ltx repo but to be honest I haven't tried the comfy built-in templates not and I won't, because I generating through kijajs transformers only. U could/should try this too. My outputs are splendid. Especially with ltx guidance parameter nodes.
Alright cool, I was definitely considering that it might be a ComfyUI template issue.. I'll give the official repo a shot.. Thanks @Sikaworld1990
Seem to do a great job my end. In fact, I find that LTX-2.3 works very well, and quite a bump up from LTX-2.0 (that was already good)
And a huge thanks to Lightricks for sharing such an amazing model with is all ;-)
(and the video above is not even at max quality settings. Basic ComfyUI run with 8 steps, distilled model... less than a minute to render. That's pretty amazing)