Qwen-image-edit training with 4-steps lora training settings

[rex-29]

Hi, i am trying to train the qwen-image-edit-2509 using GRPO to enhance ID consistent in generation. As i am using 4 steps in product env, is it possible to train the model with 4 step lora (merge as base model first)? Like klein-base-9b model those already being distilled model. Thx for your time!!!

[Jayce-Ping]

Theoretically, these RL algorithms cannot apply to distilled models, which are not Flow-Matching models. However, I've tried to do so brutally, which worked a little bit. You can always have a try.

You just need to modify the:

model:
   model_name_or_path: "path/to/your/pretrained_ckpt"

and load your merged weights directly.

[rex-29]

Theoretically, these RL algorithms cannot apply to distilled models, which are not Flow-Matching models. However, I've tried to do so brutally, which worked a little bit. You can always have a try.

You just need to modify the:

model:
model_name_or_path: "path/to/your/pretrained_ckpt"
and load your merged weights directly.

Got it, THX! There is another question that can i use the 4steps lora in sampling then train in the base model? Which can reduce a lot of time.

[Jayce-Ping]

You can do on DiffusionNFT. But you might need to modify the code a little bit.

[rex-29]

THX! I will check it later.