Advice on free local B&W film colorization — DDColor + exemplar propagation vs altern

[Dan64]

Hi Dan,
Quick follow-up — I now want to install HAVC on my own PC instead of Colab: GTX 1650 with 4GB VRAM, 16GB RAM, i5 (12th gen), Windows, SSD.
Between the v2.8.0 options (Klein 4b, Klein 9b, LongCat Q3, LongCat Q4), which one would you pick for this card — which one actually runs on 4GB VRAM and gives the fastest output? Is Klein 4b with 2 steps the right choice here, or would LongCat Q3 be faster on a 1650?
And does the official installer pick the right model for a 4GB card automatically, or do I need to select it manually?
Still curious about my three questions from Oct 22 above too, if you get a chance.
Thanks! KAKA

If you install the last version HAVC v2.9.1 it will install automatically the profile Flux2 klein 4b with 2steps, which (in Fast Pipeline) is much faster than LongCat Q3.
I don't recommend any more LongCat Q3 on small GPUs, Flux2 klein 4b is much better.
I don't have answered to your questions from Oct 22, because I don't have Linux and cannot provide any feedback, for sure the fact that was missing spatial_correlation_sampler could create some quality problems. But now you are on Windows and you should be able to note some improvement respect to Linux install.

Dan

[KAKA_colorize]

Hi Dan,

Thank you for your reply about v2.9.1 and Klein 4b.

I ran the SAME indoor test film you colorized (your Klein 4b clip in this thread) through the direct per-frame path, and the result came out broken. In my eyes, the faces changed, and the color keeps shifting throughout the clip — unlike your render of the same film, which looked clean to me. Could you help me understand why?

My clip (3 seconds, 90 frames): https://drive.google.com/file/d/1bqapWi1...p=drivesdk

Setup:
- Kaggle notebook on Linux, stock HAVCServerDiT (commit ffdecb9), no patches — server config flux2_klein_4b_int8, Fast Pipeline ON
- Direct per-frame colorize_frame, no propagation: verbatim prompt "Colorize this image, natural colors. Strictly preserve all shapes, edges and background details.", img_size 0, steps 2, seed 42
- Installed with python -m pip: torch 2.10.0+cu130, diffusers 0.37.0.dev0 (HAVCServerDiT release v1.0.0), transformers 4.57.6, accelerate 1.12.0, huggingface_hub >= 0.26.0, Pillow, scipy, av, torchsde — plus comfy-aimdo 0.6.0 (with the comfy/aimdo namespace symlink), torchvision 0.25.0+cu130, comfy-kitchen 0.2.40
- Input: 90 B&W frames, 940x704, 29.97 fps; original audio muxed onto the result
- Result: 90/90 frames colorized, 0 failed — the pipeline ran clean, but the output itself is broken

What differs between your render and mine? Is there a setting, step count, or prompt detail I'm missing — or is the direct per-frame path on Linux expected to look like this?

Thanks!
KAKA

[Dan64]

Hi KAKA,

   as I wrote I don't have linux pc and I cannot test your configuration.

  I don't understand what you mean by: "Direct coloring of an individual frame (colorize_frame), without propagation."

   In my pipeline Klein-4b is used only to provide the reference colors that are propagated by CMNET2.
   CMNET2 will propagate only the colors by maintaining the exact aspect of the reference frame (see example below).


   as you can see even if the frames 15 and 30 colorized by klein-4b are quite different from the original frames, the frames colorized by CMNET2 are equal to the original frames and the colors are properly propagated.

Dan