Add quality presets

app.py

@@ -89,6 +89,7 @@ NODE_FIRST_FRAME = "797"
 NODE_LIKENESS_GUIDE = "806"
 NODE_LIKENESS_ANCHOR = "827"
 NODE_LATENT_ANCHOR = "731"
+PRESETS = ["original", "tuned"]
 
 DEFAULT_NEGATIVE = (
     "captions, music, transition, VR, bad quality, subtitles, text, watermark, "
@@ -459,6 +460,7 @@ def _set_slider(workflow: dict[str, Any], node_id: str, value: int | float) -> N
 def _inject_params(
     workflow: dict[str, Any],
     *,
+    preset: str,
     image_name: str,
     prompt: str,
     negative_prompt: str,
@@ -486,19 +488,102 @@ def _inject_params(
     anchor = workflow.get(NODE_LIKENESS_ANCHOR, {}).get("inputs", {})
     latent_anchor = workflow.get(NODE_LATENT_ANCHOR, {}).get("inputs", {})
 
-    if mode == "anchor only":
+    if preset == "original":
+        guide["strength"] = likeness_strength
+        guide["placement_mode"] = "silent_reference"
+        guide["face_detect"] = "manual"
+        guide["reference_mask_mode"] = "bbox_only"
+        guide["face_padding"] = 0.15
+        guide["crf"] = 24
+        guide["blur_radius"] = 0
+        guide["interpolation"] = "area"
+        guide["crop"] = "center"
+        guide["attention_strength"] = 1
+        guide["emit_latent"] = "passthrough"
+        guide["debug"] = False
+
+        anchor["strength"] = likeness_anchor_strength
+        anchor["reference_source"] = "auto"
+        anchor["similarity_threshold"] = 0.5
+        anchor["decay_with_distance"] = 0
+        anchor["bypass"] = False
+        anchor["debug"] = False
+        anchor["advanced_mode"] = False
+        anchor["depth_curve"] = "middle"
+        anchor["block_index_filter"] = ""
+        anchor["similarity_sharpness"] = 8
+        anchor["override_face_bbox"] = ""
+        anchor["skip_when_sigma_above"] = 0
+        anchor["pull_mode"] = "directional"
+        anchor["late_block_falloff"] = 0.4
+
+        latent_anchor["strength"] = latent_anchor_strength
+        latent_anchor["cache_at_step"] = 5
+        latent_anchor["similarity_threshold"] = 0.5
+        latent_anchor["decay_with_distance"] = 0.15
+        latent_anchor["energy_threshold"] = 0.3
+        latent_anchor["bypass"] = False
+        latent_anchor["debug"] = False
+        latent_anchor["advanced_mode"] = True
+        latent_anchor["cache_mode"] = "schedule"
+        latent_anchor["forwards_per_step"] = 2
+        latent_anchor["cache_warmup"] = 50
+        latent_anchor["anchor_frame"] = 0
+        latent_anchor["depth_curve"] = "flat"
+        latent_anchor["block_index_filter"] = ""
+
+        if mode == "manual bbox" and face_bbox.strip():
+            guide["face_bbox_within_reference"] = face_bbox.strip()
+            anchor["frame_0_bbox"] = face_bbox.strip()
+
+    elif mode == "anchor only":
         guide["strength"] = 0.0
         anchor["strength"] = 0.0
         latent_anchor["strength"] = latent_anchor_strength
     else:
         guide["strength"] = likeness_strength
+        guide["placement_mode"] = "silent_reference"
         anchor["strength"] = likeness_anchor_strength
         latent_anchor["strength"] = latent_anchor_strength
         guide["face_detect"] = "manual" if mode == "manual bbox" else "auto"
         guide["face_bbox_within_reference"] = face_bbox.strip()
         guide["reference_mask_mode"] = "bbox_softfade"
+        guide["face_padding"] = 0.15
+        guide["crf"] = 24
+        guide["blur_radius"] = 0
+        guide["interpolation"] = "area"
+        guide["crop"] = "center"
+        guide["attention_strength"] = 1
+        guide["emit_latent"] = "passthrough"
+        guide["debug"] = False
+
         anchor["reference_source"] = "auto"
+        anchor["similarity_threshold"] = 0.45
+        anchor["decay_with_distance"] = 0
+        anchor["bypass"] = False
+        anchor["debug"] = False
+        anchor["advanced_mode"] = True
+        anchor["depth_curve"] = "flat"
+        anchor["block_index_filter"] = ""
+        anchor["similarity_sharpness"] = 6
         anchor["override_face_bbox"] = face_bbox.strip()
+        anchor["skip_when_sigma_above"] = 0
+        anchor["pull_mode"] = "directional"
+        anchor["late_block_falloff"] = 0.4
+
+        latent_anchor["cache_at_step"] = 5
+        latent_anchor["similarity_threshold"] = 0.5
+        latent_anchor["decay_with_distance"] = 0.15
+        latent_anchor["energy_threshold"] = 0.3
+        latent_anchor["bypass"] = False
+        latent_anchor["debug"] = False
+        latent_anchor["advanced_mode"] = True
+        latent_anchor["cache_mode"] = "schedule"
+        latent_anchor["forwards_per_step"] = 2
+        latent_anchor["cache_warmup"] = 50
+        latent_anchor["anchor_frame"] = 0
+        latent_anchor["depth_curve"] = "flat"
+        latent_anchor["block_index_filter"] = ""
 
     return workflow
 
@@ -565,6 +650,7 @@ def get_gpu_duration(
     image_path: str,
     prompt: str,
     negative_prompt: str,
+    preset: str,
     seconds: float,
     max_width: int,
     max_height: int,
@@ -592,6 +678,7 @@ def generate(
     image_path: str,
     prompt: str,
     negative_prompt: str,
+    preset: str,
     seconds: float,
     max_width: int,
     max_height: int,
@@ -624,6 +711,7 @@ def generate(
 
         workflow = _inject_params(
             _workflow_template(),
+            preset=preset,
             image_name=image_name,
             prompt=prompt.strip(),
             negative_prompt=negative_prompt.strip() or DEFAULT_NEGATIVE,
@@ -674,6 +762,24 @@ if os.environ.get("SKIP_STARTUP_SETUP") != "1":
     _ensure_models()
 
 
+def apply_preset(preset: str):
+    if preset == "original":
+        return (
+            gr.update(value="auto face"),
+            gr.update(value=0.9),
+            gr.update(value=0.5),
+            gr.update(value=0.11),
+            gr.update(value=0.77),
+        )
+    return (
+        gr.update(value="auto face"),
+        gr.update(value=0.9),
+        gr.update(value=0.15),
+        gr.update(value=0.08),
+        gr.update(value=0.82),
+    )
+
+
 with gr.Blocks(title="LTX image-to-video") as demo:
     gr.Markdown("# LTX image-to-video")
     with gr.Row():
@@ -681,6 +787,7 @@ with gr.Blocks(title="LTX image-to-video") as demo:
             image = gr.Image(label="reference image", type="filepath")
             prompt = gr.Textbox(label="prompt", lines=4)
             negative = gr.Textbox(label="negative prompt", value=DEFAULT_NEGATIVE, lines=2)
+            preset = gr.Dropdown(PRESETS, value="tuned", label="preset")
             seconds = gr.Slider(1.0, 10.0, value=4.0, step=0.5, label="duration")
             with gr.Row():
                 max_width = gr.Slider(512, 1536, value=1120, step=32, label="max width")
@@ -707,6 +814,7 @@ with gr.Blocks(title="LTX image-to-video") as demo:
             image,
             prompt,
             negative,
+            preset,
             seconds,
             max_width,
             max_height,
@@ -722,6 +830,12 @@ with gr.Blocks(title="LTX image-to-video") as demo:
         outputs=[video, status, used_seed],
     )
 
+    preset.change(
+        fn=apply_preset,
+        inputs=[preset],
+        outputs=[mode, likeness_strength, likeness_anchor_strength, latent_anchor_strength, first_frame_strength],
+    )
+
 demo.queue(default_concurrency_limit=1)
 
 if __name__ == "__main__":

requirements.txt

@@ -14,6 +14,7 @@ scipy
 numpy
 pillow
 opencv-python-headless
+mediapipe
 av
 kornia<0.8.0
 psutil