Open nikolaiusa opened 4 months ago
Yes, I think they are fully compatible. They are, in fact, the same model, except that for "Kandinsky Flash" pipeline an additional projection layer to the text encoder was added. They used a distillation methodology similar to LCM and SDXL Turbo and applied the approach described in "Adversarial Diffusion Distillation" by training a GAN to accelerate generation. This additional layer (as far as I understand, containing cross-attention layers for FLAN-UL2 embeddings) is required for the Flash pipeline to function properly. Therefore, even if any text encoder other than the default is chosen in Settings, the projection layer will still be used from the https://huggingface.co/ai-forever/Kandinsky3.1 text encoder (because, obviously, any other text encoder repos do not contain this layer).
thanks for the answer. can I ask another question? how to choose one of several GPU adapters? general: device: cuda1?
Right, but it must be "cuda:1" (with colon).
here's an idea for constant generation, each time with a new seed)
here's an idea for constant generation, each time with a new seed)
OK, I'll think about it. I never used this function in Auto1111, but it's easy to implement.
or are they different?