Hi Anyflow team,
Thanks for open-sourcing AnyFlow, very cool work — I really enjoy it.
I'd like to continue on-policy training on top of your released stage2
1.3B checkpoint. From the configs, real_cfg / discriminator_cfg in
options/train/anyflow//onpolicy/.yml both load the stage0 teacher
checkpoint (e.g. train_wan1b_teacher_shift5_81f_480p_lr5e-5_6k_b32/models/
step_6000.pt), which isn't in the HF release.
If I understand correctly, this teacher is your official Wan2.1-T2V-1.3B
base fine-tuned with LoRA (standard flow-matching loss) on your own
training data. Would you be able to share this teacher checkpoint (1.3B
is fine, no need for 14B)? It would save me from retraining it from
scratch just to get real_score/discriminator initialization working.
Really appreciate your help — thanks again for the great work!
Best wishes,
Zongze
Hi Anyflow team,
Thanks for open-sourcing AnyFlow, very cool work — I really enjoy it.
I'd like to continue on-policy training on top of your released stage2
1.3B checkpoint. From the configs, real_cfg / discriminator_cfg in
options/train/anyflow//onpolicy/.yml both load the stage0 teacher
checkpoint (e.g. train_wan1b_teacher_shift5_81f_480p_lr5e-5_6k_b32/models/
step_6000.pt), which isn't in the HF release.
If I understand correctly, this teacher is your official Wan2.1-T2V-1.3B
base fine-tuned with LoRA (standard flow-matching loss) on your own
training data. Would you be able to share this teacher checkpoint (1.3B
is fine, no need for 14B)? It would save me from retraining it from
scratch just to get real_score/discriminator initialization working.
Really appreciate your help — thanks again for the great work!
Best wishes,
Zongze