spooknik

/

Chroma-HD-SVDQ

like 8

Not spooknik here but, you can merge the models together with the official nunchaku model merger in ComfyUI. Just grab the two files and the two .json from here and put it all in folder.

What are the file folders, e.g., "chroma-hd-int4-2" to "chroma-hd-int4-5," for?
Are the files inside them usable in any way?
Do I need to rename anything, ...
or are these just intermediate steps for temporary storage,
meaning they're not usable for users without in-depth knowledge?

Thanks for info. :)

PS: Sorry for the maybe silly question. :)

Everything in this project is WiP, so it's a bit messy.
I am testing different methods for compressing the INT4 version of the model, it's proving very tricky to maintain good quality.

Not spooknik here but, you can merge the models together with the official nunchaku model merger in ComfyUI. Just grab the two files and the two .json from here and put it all in folder.

Thank you very much!

• I did it with "chroma-hd-int4-5" - looks ok in the NunchakuModelMerger.
• Running the result in NunchakuChromaDiTLoader gives me the Error:
  "Missing key in converted state dict: transformer_blocks.0.attn.to_out.0.qweight"

My fault?

Not spooknik here but, you can merge the models together with the official nunchaku model merger in ComfyUI. Just grab the two files and the two .json from here and put it all in folder.

Thank you very much!

• I did it with "chroma-hd-int4-5" - looks ok in the NunchakuModelMerger.
• Running the result in NunchakuChromaDiTLoader gives me the Error:
  "Missing key in converted state dict: transformer_blocks.0.attn.to_out.0.qweight"

My fault?

Nope, in4-5 has contains unquantized attention layers, which the current released version of nunchaku-chroma doesn't handle.

Hey spooknik, will You bve able to do SVDQ of Qwen image edit 2511 when they will release it ?