search

cocktailpeanut / fluxgym

Dead simple FLUX LoRA training UI with LOW VRAM support
1.12k stars 88 forks source link

Problem with lora preemption in comfy ui and its different result #96

Open jonysilverstone opened 1 month ago

jonysilverstone commented 1 month ago

TEST In short, I've been training a LoRA model several times and I see an issue with its application strength in ComfyUI. So, why do I get such terrible results when generating in Comfy, compared to the quite good results in Flux Gym? I specifically tried generating these images with the same prompt and seed, but I get such different results. Seed 42, resolution 512x512, 20 steps, and I chose the exact same settings in ComfyUI. Why is there such a strong difference? Maybe I’m not connecting the LoRA to the ksampler correctly, or I’ve also tried it with the XLabs sampler. Please help me resolve this issue, because even if I adjust the strength to 1.5 or 2, there is no significant difference, and the image becomes very noisy

jonysilverstone commented 1 month ago

lora key not loaded text_encoder_2.text_model.encoder.layers.31.self_attn.out_proj.lora_linear_layer.up.weight

Make sure you update/update_comfyui.bat on the standalone or git pull and they probably will work.