Skip to content
Python

fix self-refiner for ltx-2, proper use of lists structures

Chris Maloned03352ae8f7a
diff --git a/shared/utils/self_refiner.py b/shared/utils/self_refiner.py
index 730e00f..f51ac65 100644
--- a/shared/utils/self_refiner.py
+++ b/shared/utils/self_refiner.py
@@ -3,6 +3,11 @@ import copy
import uuid
from diffusers.utils.torch_utils import randn_tensor
+default_plan = [
+ {"start": 1, "end": 5, "steps": 3},
+ {"start": 6, "end": 13, "steps": 1},
+]
+
def is_int_string(s: str) -> bool:
try:
int(s)
@@ -11,16 +16,12 @@ def is_int_string(s: str) -> bool:
return False
def normalize_self_refiner_plan(plan_input, max_plans: int = 1):
- default_plan = [
- {"start": 1, "end": 5, "steps": 3},
- {"start": 6, "end": 13, "steps": 1},
- ]
if len(plan_input) > max_plans:
return [], f"Self-refiner supports up to {max_plans} plan(s); found {len(plan_input)}."
if not plan_input or not isinstance(plan_input, list):
- return [default_plan], ""
+ return default_plan, ""
- return [plan_input], ""
+ return plan_input, ""
def ensure_refiner_list(plan_data):
if not isinstance(plan_data, list):
@@ -302,8 +303,9 @@ class PnPHandler:
return latents, sample_scheduler
def create_self_refiner_handler(pnp_plan, pnp_f_uncertainty, pnp_p_norm, pnp_certain_percentage, channel_dim: int = 1):
- plans, _ = normalize_self_refiner_plan(pnp_plan, max_plans=1)
- stochastic_plan = plans[0]
+ stochastic_plan, _ = normalize_self_refiner_plan(pnp_plan)
+ if not stochastic_plan:
+ stochastic_plan = default_plan
return PnPHandler(
stochastic_plan,
diff --git a/wgp.py b/wgp.py
index 4460b1a..4cf6ba7 100644
--- a/wgp.py
+++ b/wgp.py
@@ -762,14 +762,11 @@ def validate_settings(state, model_type, single_prompt, inputs):
if server_config.get("fit_canvas", 0) == 2 and outpainting_dims is not None and any_letters(video_prompt_type, "VKF"):
gr.Info("Output Resolution Cropping will be not used for this Generation as it is not compatible with Video Outpainting")
if self_refiner_setting != 0:
- if isinstance(self_refiner_plan, list):
- max_plans = model_def.get("self_refiner_max_plans", 1)
- _, error = normalize_self_refiner_plan(self_refiner_plan, max_plans=max_plans)
- if len(error):
- gr.Info(error)
- return ret()
- else:
- self_refiner_plan = []
+ norm_plan, error = normalize_self_refiner_plan(self_refiner_plan, max_plans=model_def.get("self_refiner_max_plans", 1))
+ if len(error):
+ gr.Info(error)
+ return ret()
+ self_refiner_plan = norm_plan
if not model_def.get("motion_amplitude", False): motion_amplitude = 1.
if "vae" in spatial_upsampling: