Switch to dev

app.py

@@ -32,7 +32,7 @@ except json.JSONDecodeError:
 dtype = torch.bfloat16 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 
-pipe = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-distilled", torch_dtype=dtype)
+pipe = LTXConditionPipeline.from_pretrained("Lightricks/LTX-Video-0.9.7-dev", torch_dtype=dtype)
 pipe_upsample = LTXLatentUpsamplePipeline.from_pretrained("Lightricks/ltxv-spatial-upscaler-0.9.7", vae=pipe.vae, torch_dtype=dtype)
 pipe.to(device)
 pipe_upsample.to(device)
@@ -93,7 +93,7 @@ def get_huggingface_safetensors_for_ltx(link): # Renamed for clarity
     print(f"Base model from card: {base_model}")
 
     # Validate model type for LTX
-    acceptable_models = {"Lightricks/LTX-Video-0.9.7-distilled"} # Key line for LTX compatibility
+    acceptable_models = {"Lightricks/LTX-Video-0.9.7-dev"} # Key line for LTX compatibility
     
     models_to_check = base_model if isinstance(base_model, list) else [base_model]
     
@@ -423,14 +423,14 @@ def generate(prompt,
     downscaled_height, downscaled_width = int(expected_height * downscale_factor), int(expected_width * downscale_factor)
     downscaled_height, downscaled_width = round_to_nearest_resolution_acceptable_by_vae(downscaled_height, downscaled_width)
 
-    timesteps_first_pass = [1000, 993, 987, 981, 975, 909, 725]
-    timesteps_second_pass = [1000, 909, 725, 421]
-    if steps == 8:
-        timesteps_first_pass = [1000, 993, 987, 981, 975, 909, 725, 0.03]
-        timesteps_second_pass = [1000, 909, 725, 421, 0]
-    elif 7 < steps < 8: # Non-integer steps could be an issue for these pre-defined timesteps
-        timesteps_first_pass = None 
-        timesteps_second_pass = None
+    #timesteps_first_pass = [1000, 993, 987, 981, 975, 909, 725]
+    #timesteps_second_pass = [1000, 909, 725, 421]
+    #if steps == 8:
+        #timesteps_first_pass = [1000, 993, 987, 981, 975, 909, 725, 0.03]
+       # timesteps_second_pass = [1000, 909, 725, 421, 0]
+   # elif 7 < steps < 8: # Non-integer steps could be an issue for these pre-defined timesteps
+        #timesteps_first_pass = None 
+       # timesteps_second_pass = None
     
     with calculateDuration("video generation"):
         latents = pipe(
@@ -443,7 +443,7 @@ def generate(prompt,
             num_inference_steps=steps,
             decode_timestep=0.05,
             decode_noise_scale=0.025,
-            timesteps=timesteps_first_pass,
+            #timesteps=timesteps_first_pass,
             image_cond_noise_scale=0.0,
             guidance_rescale=0.7,
             guidance_scale=guidance_scale,
@@ -471,7 +471,7 @@ def generate(prompt,
                 num_frames=num_frames,
                 guidance_scale=guidance_scale,
                 denoise_strength=0.999,
-                timesteps=timesteps_second_pass,
+                #timesteps=timesteps_second_pass,
                 num_inference_steps=10, # Or make this configurable
                 latents=upscaled_latents,
                 decode_timestep=0.05,
@@ -587,8 +587,8 @@ with gr.Blocks(css=css, theme=gr.themes.Ocean(font=[gr.themes.GoogleFont("Lexend
                 seed_number_input = gr.Number(label="Seed", value=0, precision=0)
                 randomize_seed_checkbox = gr.Checkbox(label="Randomize Seed", value=True)
             with gr.Row():
-                guidance_scale_slider = gr.Slider(label="Guidance Scale (CFG)", minimum=0, maximum=10, value=1.0, step=0.1) # LTX uses low CFG
-                steps_slider = gr.Slider(label="Inference Steps (Main Pass)", minimum=1, maximum=30, value=7, step=1) # Default steps for LTX
+                guidance_scale_slider = gr.Slider(label="Guidance Scale (CFG)", minimum=0, maximum=10, value=5.0, step=0.1) # LTX uses low CFG
+                steps_slider = gr.Slider(label="Inference Steps (Main Pass)", minimum=1, maximum=30, value=25, step=1) # Default steps for LTX
                         # num_frames_slider = gr.Slider(label="# Frames (Debug - Overridden by Duration)", minimum=9, maximum=MAX_NUM_FRAMES, value=96, step=8, visible=False) # Hidden, as duration controls it
             with gr.Row():
                 height_slider = gr.Slider(label="Target Height", value=512, step=pipe.vae_spatial_compression_ratio, minimum=MIN_DIM_SLIDER, maximum=MAX_IMAGE_SIZE, info=f"Must be divisible by {pipe.vae_spatial_compression_ratio}.")