Difference with the previous Heretic "x"
is this one like the 'heretic x' that you made for ltx2? or this one is better? or just different?
also i want to thank you for the other models too
tested a bit and this one feels a bit better than heretic-x but might be just me, tested only with ltx2.3 distilled nvfp4
is this one like the 'heretic x' that you made for ltx2? or this one is better? or just different?
also i want to thank you for the other models too
tested a bit and this one feels a bit better than heretic-x but might be just me, tested only with ltx2.3 distilled nvfp4
Hi! Thank you for the kind words and for testing the models! You have a very good eye – it's not just you, this new version actually is technically superior for NVFP4, and here is exactly why:
They are different in their foundational math.
The older "Heretic-X" model was an aggressive fine-tune compressed via PTQ (Post-Training Quantization). When you force a model that only knows 16-bit math into a tight 8-bit or 4-bit space, the extreme numbers get slightly "clipped." While Heretic-X is great for raw motion, this clipping can sometimes cause micro-jitters or slight temporal instability in the video.
This new model, however, is built on QAT (Quantization-Aware Training).
During its training, artificial 4-bit noise was already simulated. The model knew it was going to be compressed! Because of this, when we crush the middle layers into the tiny NVFP4 format, the weights lock in perfectly without taking any structural damage.
So, while Heretic-X provides a loud, raw "push" for motion, this new QAT-Abliterated edition gives you the absolute best of both worlds: It is completely uncensored to prevent frozen videos, but it maintains a flawless, mathematically clean signal for the LTX-2.3 cross-attention. It is simply the cleanest and most stable text embedding you can feed the new NVFP4 engine right now!
is this one like the 'heretic x' that you made for ltx2? or this one is better? or just different?
also i want to thank you for the other models too
tested a bit and this one feels a bit better than heretic-x but might be just me, tested only with ltx2.3 distilled nvfp4Hi! Thank you for the kind words and for testing the models! You have a very good eye – it's not just you, this new version actually is technically superior for NVFP4, and here is exactly why:
They are different in their foundational math.
The older "Heretic-X" model was an aggressive fine-tune compressed via PTQ (Post-Training Quantization). When you force a model that only knows 16-bit math into a tight 8-bit or 4-bit space, the extreme numbers get slightly "clipped." While Heretic-X is great for raw motion, this clipping can sometimes cause micro-jitters or slight temporal instability in the video.
This new model, however, is built on QAT (Quantization-Aware Training).
During its training, artificial 4-bit noise was already simulated. The model knew it was going to be compressed! Because of this, when we crush the middle layers into the tiny NVFP4 format, the weights lock in perfectly without taking any structural damage.
So, while Heretic-X provides a loud, raw "push" for motion, this new QAT-Abliterated edition gives you the absolute best of both worlds: It is completely uncensored to prevent frozen videos, but it maintains a flawless, mathematically clean signal for the LTX-2.3 cross-attention. It is simply the cleanest and most stable text embedding you can feed the new NVFP4 engine right now!
waos, quite interesting indeed, thanks for explaning