Feedback after release
Looking forward to testing it more, so far, the results have been of varying quality, but very few are aesthetically pleasing. We'll see if tuning in settings changes this.
That being said, I appreciate the work and effort put into this, and hopefully issues might be mitigated, or at the very least a good amount of learning experience has been gained.
Thank you @astralite-heart and the team for the work you have done, after I run some grids, I'll post results.
@CalamitousFelicitousness I hope he improves the situation and makes an enhanced version to encourage others to use and train it
I find style_cluster_x very powerful. Once I find a style cluster that fits my need I get quite good results. I see the total lack of style cluster documentation holding back the model by a lot as casual users like me don't have the time to compile a list of all the 2048 different style clusters so I really hope someone will create and share one. I had a much easier time getting Pony v6 based models generate aesthetically pleasing images compared to Pony v7 but if I put enough effort into it, I can get better results with Pony v7.
I waited for this model for such a long time and am really happy they finally released it. This was my most anticipated stable diffusion model and almost daily checked for news about it. Their training cost must have been massive especially because they trained in full resolution so we should really appreciate what we got. I don't think anyone else will be crazy enough to do such an extensive training run in full resolution. Now that I can run this model locally, I can finally enjoy it. I'm already looking forward to Pony v8. Keep up the great work!
@nicoboss Hi, what can you say about other models? I don't plan to dedicate my time to Pony v7, unless there's a new, better fine-tune. If someone gets normal images with the base model, I don't mind, but it appears that it's hard to use it, like SD 1
@nicoboss Hi, what can you say about other models? I don't plan to dedicate my time to Pony v7, unless there's a new, better fine-tune. If someone gets normal images with the base model, I don't mind, but it appears that it's hard to use it, like SD 1
@qpqpqpqpqpqp Here my personaly opinion about the most popolar stable diffusion models:
- SD 1.5: Despite its age still one of my favorite models. The model is so tiny that I can use massive batch sizes to generate a massive number of images at once. Quite bad with anatomy and the 512x512 resolution limit can be a limit for certain use-cases. Very uncensored base model as it was trained on unfiltered LION making it very unique.
- SD 2.0/SD 2.1: In almost any way worse than SD 1.5 or SDXL. I used it when it initial released but since the SDXL release I’m no longer using it.
- SDXL/Pony v6/Illustration: Simple to prompt and gives beautiful results with very little effort. There is a massive ecosystem around those models. No matter how niche the thing you want to generate is someone likely either already created a finetune or LoRA for your use-case. I still absolutely love to use them.
- SD 3.5: After the licensing disaster during the initial release, I completely boycott this model.
- Flux.1-Dev: Censored and relatively heavy to run but perfect at generating photo realistic images. I used it quite often to generate synthetic training data before Qwen got released.
- Chroma: Builds on top of Flux.1-Dev improving it in every way but also making it much harder to prompt so that for most use cases I still prefer the original.
- Pony v7: Extremely hard to prompt especially with the lack of style cluster documentation but can generate really good images if one spends enough effort.
- Qwen: Perfect model. Simple to prompt and really fast to run on A100 40G. Uses around 30 GiB of GPU memory at batch size 64 at 512x512 but requires around 50 steps for best possible results. An entire ecosystem starts to pop up around it so I see me soon switching from SDXL/Pony v6/Illustration to Qwen for most my use cases. Currently I mostly use it for synthetic training data generation.