prompting should work

hydra_node/models.py

@@ -34,6 +34,7 @@ def fix_batch(tensor, bs):
     return torch.stack([tensor.squeeze(0)]*bs, dim=0)
 
 # mix conditioning vectors for prompts
+# @aero
 def prompt_mixing(model, prompt_body, batch_size):
     if "|" in prompt_body:
         prompt_parts = prompt_body.split("|")
@@ -183,7 +184,7 @@ class StableDiffusionModel(nn.Module):
                 ], device=self.device)
         
         prompt = [request.prompt] * request.n_samples
-        prompt_condition = self.model.get_learned_conditioning(prompt)
+        prompt_condition = prompt_mixing(self.model, prompt[0], request.n_samples)
 
         uc = None
         if request.scale != 1.0:
@@ -227,13 +228,13 @@ class StableDiffusionModel(nn.Module):
                 if request.scale != 1.0:
                     uc = self.model.get_learned_conditioning(request.n_samples * [""])
                 
-                c = prompt_mixing(self.model, prompt[0], request.n_samples)#(model.get_learned_conditioning(prompts) + model.get_learned_conditioning(["taken at night"])) / 2
+                prompt_condition = prompt_mixing(self.model, prompt[0], request.n_samples)
 
                 # encode (scaled latent)
                 start_code_terped=None
                 z_enc = sampler.stochastic_encode(init_latent, torch.tensor([t_enc]*request.n_samples).to(self.device), noise=start_code_terped)
                 # decode it
-                samples = sampler.decode(z_enc, c, t_enc, unconditional_guidance_scale=request.scale,
+                samples = sampler.decode(z_enc, prompt_condition, t_enc, unconditional_guidance_scale=request.scale,
                                         unconditional_conditioning=uc,)
 
                 x_samples_ddim = self.model.decode_first_stage(samples)