DiffSynth-Studio

DiffSynth-Studio copied to clipboard

[Feature Request] Add reinforcement learning (RL) training support for Qwen-Image-Edit-2509 in DiffSynth.

1

Is there a plan to add reinforcement learning (RL) training support for Qwen-Image-Edit-2509 in DiffSynth?

zk1009

When training LoRA on Wan2.1-14B for image-to-video, the training loss keeps increasing instead of decreasing, and the per-step loss fluctuates heavily. Has anyone else run into the same issue?

lin076

Wan2.1 lora training loss

12

对于wan2.1，在一些开源video-caption数据集上，比如vidgen，koala36m，大概最终的runnning_mean_loss是多少呢 For wan2.1 trained on open-source datasets such as vidgen and koala36m using lora, what is the ultimate running_mean_loss?

AliothChen

Hi, Thanks for your excellect framework! I’m fine-tuning Qwen-Image-Edit-2509 on a self-built image editing dataset, but the training speed is extremely slow. Environment: single node with 8 GPUs Dataset size:...

pd162

Request the support for FastVideo

2

Lifedecoder

When using gradient accumulation, does the order of optimizer.zero_grad() affect training?

2

if I use accelerate+deepspeed to train a model, and I set deepspeed_config: gradient_accumulation_steps: 8 offload_optimizer_device: cpu offload_param_device: cpu zero3_init_flag: false zero_stage: 2 does the order of the order of backward(),...

polestarss

CPU memory explosion during LoRA training — all training processes duplicate full model weights (no shared memory / lazy loading)

1

Hi team, I encountered a severe CPU memory overflow issue when fine-tuning LoRA on the Qwen-Image-Edit-2509 model using the default examples/qwen_image/model_training/train.py script. **💻 Environment** OS: Ubuntu 22.04 GPUs: 1 ×...

KAI4816

Wan 2.2 S2V 训练中没有 motion_video 的输入嘛？

1

dataset 里面没有关于 motion_video 的输入，WanVideoUnit_S2V 中对 motion_video 全部进行 0 初始化。但是推理的时候，却保留 motion_videos 。 1、请问训练的 motion_video 需要自己写输入嘛？ 2、如果全部是 0 初始化，训练推理有一定 gap 是否合理？

wangjue-wzq

qwen image lora training problem.

1

I can not find any code of lora parameters doing the weight+A*B operation at training forward, am I missing some code of step. is this the key operation ?

Vincento-Wang

Wan 2.2 TI2V lora训练速度

1

你好，TI2V 5B lora的训练速度很慢是正常吗？平均下来bz 16 12s/step。是哪里需要再修改什么吗？

xuxiaoxxxx