21.10.2026, 22:57
Hi dan64,
I finally ran your own HAVC v2.8.0 with the Klein 4B server (flux2_klein_4b_int8) on the SAME test film you colorized — in my case on a Colab T4 (Linux), with default steps. One Linux note: spatial_correlation_sampler would not build there, so propagation used a pure-PyTorch fallback based on your own correlation math; everything else is your stock pipeline.
Side-by-side of your clip (left) and mine (right):
https://drive.google.com/file/d/1F90FcCr...bx_Xa/view
My full 3-second clip:
https://drive.google.com/file/d/1XKc5znr...4bs20/view
Three things puzzle me:
1. Your clip is only 630x328, mine is 940x704, yet yours looks cleaner and more vivid. Is that just the smaller scale hiding softness, or is something different in your render (steps, settings, post-processing)?
2. Yours has NO color bleed/spread at all — the colors sit exactly inside their objects. Mine is softer, with slight bleed. Could that be the compiled correlation kernel vs my PyTorch fallback, or a propagation setting?
3. Your curtains came out purple; mine are red/burgundy (and in an earlier letterboxed run they were even brownish). Same film, same model. Is that pure seed variance, or do you set something (prompt tweak, color option) that pushes such decisions?
Stills from shot B attached (yours vs mine). Any hint is appreciated — my goal is to reproduce your result 1:1.
your clip:
my clip:
Thanks!
KAKA
I finally ran your own HAVC v2.8.0 with the Klein 4B server (flux2_klein_4b_int8) on the SAME test film you colorized — in my case on a Colab T4 (Linux), with default steps. One Linux note: spatial_correlation_sampler would not build there, so propagation used a pure-PyTorch fallback based on your own correlation math; everything else is your stock pipeline.
Side-by-side of your clip (left) and mine (right):
https://drive.google.com/file/d/1F90FcCr...bx_Xa/view
My full 3-second clip:
https://drive.google.com/file/d/1XKc5znr...4bs20/view
Three things puzzle me:
1. Your clip is only 630x328, mine is 940x704, yet yours looks cleaner and more vivid. Is that just the smaller scale hiding softness, or is something different in your render (steps, settings, post-processing)?
2. Yours has NO color bleed/spread at all — the colors sit exactly inside their objects. Mine is softer, with slight bleed. Could that be the compiled correlation kernel vs my PyTorch fallback, or a propagation setting?
3. Your curtains came out purple; mine are red/burgundy (and in an earlier letterboxed run they were even brownish). Same film, same model. Is that pure seed variance, or do you set something (prompt tweak, color option) that pushes such decisions?
Stills from shot B attached (yours vs mine). Any hint is appreciated — my goal is to reproduce your result 1:1.
your clip:
my clip:
Thanks!
KAKA