From b452488c58a0045dd03967ec8ffdc6ed327fdc83 Mon Sep 17 00:00:00 2001
From: Haoming <hmstudy02@gmail.com>
Date: Mon, 23 Feb 2026 11:54:01 +0800
Subject: [PATCH] Flux2Scheduler

---
 modules/sd_samplers_kdiffusion.py | 15 +++++++++-
 modules/sd_schedulers.py          | 46 +++++++++++++++++++++++++++++--
 2 files changed, 57 insertions(+), 4 deletions(-)

diff --git a/modules/sd_samplers_kdiffusion.py b/modules/sd_samplers_kdiffusion.py
index 85598b67..2ec7d059 100644
--- a/modules/sd_samplers_kdiffusion.py
+++ b/modules/sd_samplers_kdiffusion.py
@@ -79,7 +79,12 @@ class KDiffusionSampler(sd_samplers_common.Sampler):
 
         scheduler_name = (p.hr_scheduler if p.is_hr_pass else p.scheduler) or "Automatic"
         if scheduler_name == "Automatic":
-            scheduler_name = self.config.options.get("scheduler", None)
+            from backend.args import dynamic_args
+
+            if dynamic_args["klein"]:
+                scheduler_name = "Flux2"
+            else:
+                scheduler_name = self.config.options.get("scheduler", None)
 
         scheduler = sd_schedulers.schedulers_map.get(scheduler_name)
 
@@ -116,6 +121,14 @@ class KDiffusionSampler(sd_samplers_common.Sampler):
                 p.extra_generation_params["Beta schedule alpha"] = opts.beta_dist_alpha
                 p.extra_generation_params["Beta schedule beta"] = opts.beta_dist_beta
 
+            if scheduler.label == "Flux2":
+                if p.is_hr_pass:
+                    sigmas_kwargs["width"] = p.hr_upscale_to_x
+                    sigmas_kwargs["height"] = p.hr_upscale_to_y
+                else:
+                    sigmas_kwargs["width"] = p.width
+                    sigmas_kwargs["height"] = p.height
+
             sigmas = scheduler.function(n=steps, **sigmas_kwargs, device=devices.cpu)
 
         if discard_next_to_last_sigma:
diff --git a/modules/sd_schedulers.py b/modules/sd_schedulers.py
index 7dfb518b..a752b2c0 100644
--- a/modules/sd_schedulers.py
+++ b/modules/sd_schedulers.py
@@ -1,13 +1,14 @@
 import dataclasses
-from math import atan, pi
+from math import atan, exp, pi
 from typing import Callable
 
 import k_diffusion
 import numpy as np
 import torch
-from modules import shared
 from scipy import stats
 
+from modules import shared
+
 
 def to_d(x: torch.Tensor, sigma: float, denoised: torch.Tensor):
     """Converts a denoiser output to a Karras ODE derivative"""
@@ -201,7 +202,9 @@ def bong_tangent_scheduler(n, sigma_min, sigma_max, device, *, start=1.0, middle
 
 
 def flow_match_euler_discrete_scheduler(n, sigma_min, sigma_max, inner_model, device):
-    from diffusers.schedulers.scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler
+    from diffusers.schedulers.scheduling_flow_match_euler_discrete import (
+        FlowMatchEulerDiscreteScheduler,
+    )
 
     unet = inner_model.inner_model.forge_objects.unet
 
@@ -225,6 +228,42 @@ def flow_match_euler_discrete_scheduler(n, sigma_min, sigma_max, inner_model, de
     return torch.FloatTensor(sigmas).to(device)
 
 
+def generalized_time_snr_shift(t: torch.Tensor, mu: float, sigma: float) -> float:
+    return exp(mu) / (exp(mu) + (1 / t - 1) ** sigma)
+
+
+def compute_empirical_mu(image_seq_len: int, num_steps: int) -> float:
+    a1, b1 = 8.73809524e-05, 1.89833333
+    a2, b2 = 0.00016927, 0.45666666
+
+    if image_seq_len > 4300:
+        mu = a2 * image_seq_len + b2
+        return float(mu)
+
+    m_200 = a2 * image_seq_len + b2
+    m_10 = a1 * image_seq_len + b1
+
+    a = (m_200 - m_10) / 190.0
+    b = m_200 - 200.0 * a
+    mu = a * num_steps + b
+
+    return float(mu)
+
+
+def get_schedule(num_steps: int, image_seq_len: int) -> list[float]:
+    mu = compute_empirical_mu(image_seq_len, num_steps)
+    timesteps = torch.linspace(1, 0, num_steps + 1)
+    timesteps = generalized_time_snr_shift(timesteps, mu, 1.0)
+    return timesteps
+
+
+def flux2_scheduler(n: int, width: int, height: int, sigma_min, sigma_max, device):
+    # https://github.com/Comfy-Org/ComfyUI/blob/master/comfy_extras/nodes_flux.py
+    seq_len = width * height / (16 * 16)
+    sigmas = get_schedule(n, round(seq_len))
+    return torch.FloatTensor(sigmas).to(device)
+
+
 schedulers = [
     Scheduler("automatic", "Automatic", None),
     Scheduler("karras", "Karras", k_diffusion.sampling.get_sigmas_karras, default_rho=7.0),
@@ -242,6 +281,7 @@ schedulers = [
     Scheduler("turbo", "Turbo", turbo_scheduler, need_inner_model=True),
     Scheduler("bong_tangent", "Bong Tangent", bong_tangent_scheduler),
     Scheduler("flow_match", "FlowMatchEulerDiscrete", flow_match_euler_discrete_scheduler, need_inner_model=True),
+    Scheduler("flux2", "Flux2", flux2_scheduler),
 ]
 
 schedulers_map = {**{x.name: x for x in schedulers}, **{x.label: x for x in schedulers}}