Using Stable Diffision models for Colorization

[Dan64]

Hello,

   I released HAVC Installer 2.6.0 (for the full list of changes see: What's new)

   Now the supported DiT models are 6 (for models description see: Backends):

     1)  qwen-nunchaku (fp4/int4)
     2)  qwen-2511 (int8/q3)
     3)  longcat (q3/q4/q5/q6)
     4)  qwen21-viggle [sharp/soft] (int8/q4) 
     5)  qwen21-turbo [fast] (int8)
     6)  flux2_dev (int4/fp8/q2/q3)

  The one with the best colors remains qwen-nunchaku, while the one with the more stable colors (see image below), ideal for coloring long clips, is qwen21-viggle [sharp/soft].

   

  Thanks to recent code optimizations: Dynamic VRAM and Text-Encoder (TE) split, the qwen21-viggle model is the fastest (1.7 s/image) and has the lowest minimum hardware requirements (12 GB of VRAM, 32 GB of RAM).
  It is the default color model in HAVCServerDiT.

With these improvements, it is finally possible to colorize black-and-white films, leveraging the advanced quality provided by DiT models, on consumer-grade PCs.

    
Dan