DiffSynth-Studio

DiffSynth-Studio copied to clipboard

Flux full train的模型转opensource format

4

Flux full train的模型在comfyui中无法加载，利用diffusers的脚本也无法转成diffusers格式，这个是否能支持

justfuwei

Probs about VACE full training setting of Num_frames

2

Hi～ 非常棒的Codebase！ 想请问一下在训练Wan2.1-VACE的full版本过程中，为什么当设置num_frames 大于 49的时候，生成的结果会是错误的？看VACE的官方版本是支持frames=81的. 当前模型是否是由于VACE的基模导致只能训练小于81的帧数视频呢，有其他基础模型可以生成长视频的吗？ Thanks!

XiaoKangW

Expected to have finished reduction in the prior iteration before starting a new one.

2

I was finetuning the [Flux.1-dev: Upscaler ControlNet]() model with my custom data, but met this problem during training: `[rank1]: RuntimeError: Expected to have finished reduction in the prior iteration before...

cxzhou35

关于Wan训练时的timestep

1

[https://github.com/modelscope/DiffSynth-Studio/blob/main/diffsynth/pipelines/wan_video_new.py#L108](url) 这个函数中timestep_id的范围是`[0, num_train_timesteps）`也就是`[0, 1000)`，而`self.scheduler.timesteps`的长度只有`num_inference_steps`，默认值是100，而在实际推理的时候可能只有40/50。 也就是说，会生成一个0到1000的随机数作为一个长度为100的数组的索引，这不会出现问题吗？ 期待您的解答，谢谢！

midcola

support of hunyuan-image-3.0?

1

请问什么时候能够支持混元 image3.0 的微调以及 lora 呢，或者有没有适配的计划？感谢！

vtars

Is there a way to combine Control and Control-Camera models?

1

Hi, thank you for the great work. Is there a way to combine the Control and Control-Camera models to have both video control and camera control? Could I, e.g., copy...

codeysun

Question about Qwen-Image-Edit prompt enhancement

1

Original Qwen-Image repo said before send to Qwen-Image-Edit model, prompt should be polished by https://github.com/QwenLM/Qwen-Image/blob/26a3635f485b0a6d85bdb2546b51fc782986fc0a/src/examples/tools/prompt_utils.py#L91 I wonder whether DiffSynth has implemented this. I do not find this. Current implementation has...

liangbingzhao

support of qwen image controlnets & Eligen in comfyUI?

1

is it possible to use the controlnets released: https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-EliGen & https://www.modelscope.cn/models/DiffSynth-Studio/Qwen-Image-Blockwise-ControlNet-Depth the implementation of these in comfyUI so we can able to use them with fp8 models and

krigeta

HunyuanVideo ValueError: Image features and image tokens do not match: tokens: 1, features 2359296

2

怎么解决呢 ``` Traceback (most recent call last): File "/root/paddlejob/workspace/env_run/zwr_workspace/DiffSynth-Studio/hunyuanvideo_i2v_24G.py", line 43, in video = pipe(prompt, input_images=images, num_inference_steps=50, seed=0, i2v_resolution=i2v_resolution) File "/root/paddlejob/workspace/env_run/miniconda3/envs/Diff-S/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) File "/root/paddlejob/workspace/env_run/zwr_workspace/DiffSynth-Studio/diffsynth/pipelines/hunyuan_video.py",...

qq1343277857

Support resume training of `QwenImagePipeline` LoRA

1

May I ask how to implement the resuming training of `QwenImagePipeline` lora weights? I didn't find similar parameters :)

JunyaoHu