Any plan to fix the typo and train again?
Based on Comfyanonymous's comment on the PR, it seems like a typo has been made before the training on the 2.0 version which result in the model being slower than how it should perform, any plans to fix this?
Hi team, just checking in on this discussion, is there any update or plan to address the typo mentioned by ComfyAnonymous and potentially retrain a corrected version 2.1? The model is great, but fixing the speed issue would make it even better! Thanks.
I said v1 was large, they decided to make it even larger
Hello, I've identified this issue and tried to address it with a workaround, but it didn't succeed. I think if there's demand, I could retrain this part of the model as version 2.1 to better meet the community's needs. If the current model's performance meets the requirements, I believe it would be acceptable to invest additional resources into retraining it.
I said v1 was large, they decided to make it even larger
Hello, I saw your comment. Actually, I also tried making it smaller, but the performance was very poor, so I had no choice but to abandon that approach.
Hello, I've identified this issue and tried to address it with a workaround, but it didn't succeed. I think if there's demand, I could retrain this part of the model as version 2.1 to better meet the community's needs. If the current model's performance meets the requirements, I believe it would be acceptable to invest additional resources into retraining it.
I think, you should wait for their base model, and train for that model. So not to waste compute. Since most people will use the base model with turbo extracted as a lora anyways.
Hello, I've identified this issue and tried to address it with a workaround, but it didn't succeed. I think if there's demand, I could retrain this part of the model as version 2.1 to better meet the community's needs. If the current model's performance meets the requirements, I believe it would be acceptable to invest additional resources into retraining it.
Unrelated to Z-Image can we get reward loras/fun models for the Kandinsky 5 Pro T2V/I2V 19B models? Since it's currently the #1 open model on LMArena having same points as Luma Ray3 and Hailuo 2.3 and surpasses Seedance. I believe with reward loras it can surpass even Kling 2.6 since it's not that far in points.
Hello, I've identified this issue and tried to address it with a workaround, but it didn't succeed. I think if there's demand, I could retrain this part of the model as version 2.1 to better meet the community's needs. If the current model's performance meets the requirements, I believe it would be acceptable to invest additional resources into retraining it.
Since the release date of the base model is unknown, I’d support retraining it as long as it doesn’t require significant resources.