Drop Quality mode; distilled-only with new defaults (distill 0.7, 5s, crossview 1.5, cfg 1, 8 steps)

README.md

@@ -31,23 +31,24 @@ Upload a reference video and pick a new camera angle (azimuth, elevation, distan
 - **Pipeline**: `LTX2InContextPipeline` with `LTX2ReferenceCondition` for in-context video conditioning
 - **LoRA**: applied via standard `pipe.load_lora_weights()` + `set_adapters()`
 
-## Sampling modes
+## Sampling
 
-Two paths are available from the **Sampling mode** selector:
+Distilled sampling, mirroring the validated ComfyUI workflow for this LoRA family:
+the dev transformer with the [distilled speed LoRA](https://huggingface.co/Kijai/LTX2.3_comfy)
+(rank-dynamic resize) stacked on top.
 
-| Mode | Setup | Steps | Guidance |
-|---|---|---|---|
-| **Fast (distilled)** | dev + [distilled speed LoRA](https://huggingface.co/Kijai/LTX2.3_comfy) (rank-dynamic resize) at strength 0.6 | 8 | cfg 1, distilled sigma schedule |
-| **Quality (dev)** | dev model alone | 30 (slider) | cfg 3 (slider), STG on |
+| Control | Default |
+|---|---|
+| Duration | 5 s |
+| CrossView LoRA strength | 1.5 |
+| Distill LoRA strength | 0.7 |
+| Inference steps | 8 |
+| Guidance scale | 1 |
 
-Fast mode mirrors the validated ComfyUI workflow for this LoRA family: the dev
-transformer with the distilled LoRA stacked on top, sampled at cfg 1 over the
-distilled sigma schedule. Quality mode is the original path and still honours the
-step and guidance sliders.
-
-**Distill LoRA strength** is adjustable under Advanced (Fast mode only). `0.6`
-matches the ComfyUI workflow; lower values (0.25-0.5) preserve more motion at the
-cost of needing more steps.
+All five are adjustable under **Advanced**. At 8 steps the distilled sigma
+schedule is used; any other step count falls back to the scheduler's own spacing.
+Practitioners often run the distill LoRA at 0.25-0.5 — a stronger distill trades
+away motion, and a weaker one usually wants more steps.
 
 > The distilled LoRA used here is rank-dynamic, and the diffusers LTX-2 loader
 > discards per-layer alphas (`network_alphas=None`), which would otherwise scale
@@ -69,4 +70,4 @@ Use the dropdowns in the UI to build a valid prompt — the app constructs it au
 
 - **Small angle changes work best.** For larger viewpoint shifts, chain multiple small steps (feed the generated view back in as the new reference).
 - **LoRA strength** of 1.0–1.5 is recommended. Higher values produce a stronger viewpoint shift.
-- Fast mode generates a 3-second video in well under a minute; Quality mode takes ~2-4 minutes at 30 steps.
+- Generation takes well under a minute at the default 8 steps.

app.py

@@ -27,12 +27,7 @@ DISTILL_LORA_REPO = "Kijai/LTX2.3_comfy"
 DISTILL_LORA_WEIGHT_NAME = (
     "loras/ltx-2.3-22b-distilled-1.1_lora-dynamic_fro09_avg_rank_111_bf16.safetensors"
 )
-DISTILL_LORA_SCALE = 0.6
-
-MODE_FAST = "Fast (distilled, 8 steps)"
-MODE_QUALITY = "Quality (dev, 30 steps)"
-MODE_CHOICES = [MODE_FAST, MODE_QUALITY]
-DEFAULT_MODE = MODE_FAST
+DISTILL_LORA_SCALE = 0.7
 
 # Camera vocabulary (from captions_all_63.txt)
 AZIMUTH_CHOICES = [
@@ -52,9 +47,11 @@ DEFAULT_WIDTH = 768
 DEFAULT_HEIGHT = 512
 DEFAULT_NUM_FRAMES = 81
 DEFAULT_FPS = 24
-DEFAULT_STEPS = 30
-DEFAULT_GUIDANCE = 3.0
-DEFAULT_LORA_SCALE = 1.0
+DEFAULT_STEPS = len(DISTILLED_SIGMA_VALUES)  # 8
+DEFAULT_GUIDANCE = 1.0
+DEFAULT_LORA_SCALE = 1.5
+DEFAULT_DURATION = 5.0
+
 
 def _normalize_dynamic_rank_lora(state_dict):
     """Pre-compensate for the diffusers LTX-2 loader dropping per-layer alphas.
@@ -129,7 +126,6 @@ def generate(
     guidance_scale,
     num_steps,
     negative_prompt,
-    mode=DEFAULT_MODE,
     distill_strength=DISTILL_LORA_SCALE,
     progress=gr.Progress(track_tqdm=True),
 ):
@@ -152,21 +148,15 @@ def generate(
     ref_frames = load_video(reference_video)
     ref_cond = LTX2ReferenceCondition(frames=ref_frames, strength=1.0)
 
-    # Select the sampling path. Fast mode stacks the distill LoRA on the dev
-    # transformer and samples at cfg 1 over the distilled sigma schedule (the
-    # ComfyUI workflow's settings); quality mode runs the bare dev model.
-    if mode == MODE_FAST:
-        pipe.set_adapters(["crossview", "distill"], [lora_scale, float(distill_strength)])
-        steps = len(DISTILLED_SIGMA_VALUES)
-        sigmas = DISTILLED_SIGMA_VALUES
-        cfg = 1.0
-        stg = 0.0
-    else:
-        pipe.set_adapters("crossview", lora_scale)
-        steps = int(num_steps)
-        sigmas = None
-        cfg = guidance_scale
-        stg = 1.0
+    # Distilled sampling: the distill LoRA stacked on the dev transformer, as in
+    # the validated ComfyUI workflow. DISTILLED_SIGMA_VALUES is an 8-value
+    # schedule, so it only applies at its native step count; any other step count
+    # falls back to the scheduler's own spacing.
+    pipe.set_adapters(["crossview", "distill"], [lora_scale, float(distill_strength)])
+    steps = int(num_steps)
+    sigmas = DISTILLED_SIGMA_VALUES if steps == len(DISTILLED_SIGMA_VALUES) else None
+    cfg = guidance_scale
+    stg = 0.0
 
     # Run inference (return_dict=False gives (video, audio) tuple)
     video, audio = pipe(
@@ -260,25 +250,16 @@ with gr.Blocks(title="LTX CrossView-Prompt - Video Camera Control by Prompt", cs
                         info="Camera distance to subject",
                     )
 
-                mode = gr.Radio(
-                    choices=MODE_CHOICES,
-                    value=DEFAULT_MODE,
-                    label="Sampling mode",
-                    info="Fast stacks the distilled speed LoRA on the dev model "
-                         "(8 steps, cfg 1). Quality runs the dev model alone and "
-                         "honours the step / guidance sliders below.",
-                )
-
                 with gr.Accordion("Advanced", open=False):
                     distill_strength = gr.Slider(
                         minimum=0.0, maximum=1.0, value=DISTILL_LORA_SCALE, step=0.05,
-                        label="Distill LoRA strength (Fast mode only)",
+                        label="Distill LoRA strength",
                         info="0.6 matches the ComfyUI workflow. Practitioners often "
-                             "run 0.25-0.5 — a stronger distill trades away motion. "
-                             "Ignored in Quality mode.",
+                             "run 0.25-0.5 — a stronger distill trades away motion, "
+                             "and a weaker one usually wants more steps.",
                     )
                     duration_seconds = gr.Slider(
-                        minimum=1, maximum=5, value=3, step=0.5,
+                        minimum=1, maximum=5, value=DEFAULT_DURATION, step=0.5,
                         label="Duration (seconds)",
                     )
                     lora_scale = gr.Slider(
@@ -293,6 +274,8 @@ with gr.Blocks(title="LTX CrossView-Prompt - Video Camera Control by Prompt", cs
                     num_steps = gr.Slider(
                         minimum=8, maximum=50, value=DEFAULT_STEPS, step=1,
                         label="Inference steps",
+                        info="8 uses the distilled sigma schedule. Any other value "
+                             "falls back to the scheduler's default spacing.",
                     )
                     negative_prompt = gr.Textbox(
                         value="",
@@ -320,13 +303,13 @@ with gr.Blocks(title="LTX CrossView-Prompt - Video Camera Control by Prompt", cs
         gr.Markdown("---\n### 📋 Examples")
         gr.Markdown("Click an example to populate the inputs, then click **Generate New View**.")
         examples = [
-            ["assets/ref_dining.mp4", "to the right", "lower", "closer", 3, 42, False, 1.0, 3.0, 30, "", MODE_FAST, 0.6],
-            ["assets/ref_dining.mp4", "to the left", "higher", "further", 3, 123, False, 1.0, 3.0, 30, "", MODE_FAST, 0.6],
-            ["assets/ref_scene01.mp4", "slightly to the left", "higher", "closer", 3, 42, False, 1.0, 3.0, 30, "", MODE_FAST, 0.6],
+            ["assets/ref_dining.mp4", "to the right", "lower", "closer", 5, 42, False, 1.5, 1.0, 8, "", 0.7],
+            ["assets/ref_dining.mp4", "to the left", "higher", "further", 5, 123, False, 1.5, 1.0, 8, "", 0.7],
+            ["assets/ref_scene01.mp4", "slightly to the left", "higher", "closer", 5, 42, False, 1.5, 1.0, 8, "", 0.7],
         ]
         gr.Examples(
             examples=examples,
-            inputs=[reference_video, azimuth, elevation, distance, duration_seconds, seed, randomize_seed, lora_scale, guidance_scale, num_steps, negative_prompt, mode, distill_strength],
+            inputs=[reference_video, azimuth, elevation, distance, duration_seconds, seed, randomize_seed, lora_scale, guidance_scale, num_steps, negative_prompt, distill_strength],
             outputs=[output_video, used_seed, used_prompt],
             fn=generate,
             cache_examples=True,
@@ -336,7 +319,7 @@ with gr.Blocks(title="LTX CrossView-Prompt - Video Camera Control by Prompt", cs
     # Wire up
     generate_btn.click(
         fn=generate,
-        inputs=[reference_video, azimuth, elevation, distance, duration_seconds, seed, randomize_seed, lora_scale, guidance_scale, num_steps, negative_prompt, mode, distill_strength],
+        inputs=[reference_video, azimuth, elevation, distance, duration_seconds, seed, randomize_seed, lora_scale, guidance_scale, num_steps, negative_prompt, distill_strength],
         outputs=[output_video, used_seed, used_prompt],
     )