Hybrid model
Hi! Would it be possible to add proper support for the FL2VA/Ref2VA hybrid checkpoints, e.g. minimax_h3_hybrid_fl2va_ref2va_b20-49 or similar?
As I understand it, using either the FL2VA or Ref2VA PDD-Acc weights alone is off-label for hybrid models. Ideally, the node could load both PDD-Acc files and apply them according to the hybrid block split.
Is something like this technically feasible / planned?
Hi! Would it be possible to add proper support for the FL2VA/Ref2VA hybrid checkpoints, e.g. minimax_h3_hybrid_fl2va_ref2va_b20-49 or similar?
As I understand it, using either the FL2VA or Ref2VA PDD-Acc weights alone is off-label for hybrid models. Ideally, the node could load both PDD-Acc files and apply them according to the hybrid block split.
Is something like this technically feasible / planned?
Unfortunately, hybrid(25, 49)block_range_adaln is the wrong answer because introducing reference workflow nodes in fl2va inevitably leads to abnormal rendering and degradation issues when using Sage Attention and Kitchen. A few days after the release of Scottmudge, I published a white paper refuting this claim after developing compatibility with int8. Currently, I highly recommend replacing any hybrid model with all_adaln. For all_adaln replacement, Codex can easily write a corresponding LoRa for each activation.
Hi, I agree with ANe5s response. I've used both the hybrid models and the hybrid loader - you should have a coding agent create a lora or new model with all_adaln. PDD support should be in Comfy core soon and I think you will have the same issue there. I know there is another version of this implementation maintained by Silveroxides and I looked through that code and it will also have the same issue. I used fable to create a rebased version of the hybrid, there might be one floating around on HF (I wasn't able to find it).