search

cumulo-autumn / StreamDiffusion

StreamDiffusion: A Pipeline-Level Solution for Real-Time Interactive Generation
Apache License 2.0
9.73k stars 697 forks source link

poor image quality compared to AUTOMATIC1111 #28

Open NaCloudAI opened 10 months ago

NaCloudAI commented 10 months ago

image image

This reason seems to be lack of skip_clip and no negative prompt support, for txt2img will this be support in future?

teftef6220 commented 10 months ago

Thanks for the comment.

Try changing the value of t_index_list. By default cfg_type is set to noneChange it to full and try it. (I recommend using a guidance_scale value around 1.0 ~ 1.5.)

We will discuss the implementation of clip_skip with our development team.

NaCloudAI commented 10 months ago

Thanks for the comment.

Try changing the value of t_index_list. By default cfg_type is set to noneChange it to full and try it. (I recommend using a guidance_scale value around 1.0 ~ 1.5.)

We will discuss the implementation of clip_skip with our development team.

Thanks for the reply.

for txt2img it seems I am required to use cfg = none. Is there some way I could use cfg in txt2img?

for clip_skip, I found I can change it in pipeline.py in encoding_prompt function.

teftef6220 commented 10 months ago

Thanks for your insight.

Changing the cfg_type in txt2img is a bit of a challenge since it requires modifying the contents of the pipeline, but We will discuss this among developers. (Sorry, I misunderstood your question as img2img...)

As for clip_skip, if you can submit a PR, we will consider it and try to merge it into the clip_skip.

thank you