From 48810e6a15104fbe01966d028291a0e7d75be8a4 Mon Sep 17 00:00:00 2001
From: asagi4 <130366179+asagi4@users.noreply.github.com>
Date: Sun, 31 Dec 2023 19:43:37 +0200
Subject: [PATCH] Note about ComfyUI_stable_fast

---
 0001-Limit-batch-chunk-size-to-2.patch | 32 --------------------------
 README.md                              |  8 +++----
 2 files changed, 4 insertions(+), 36 deletions(-)
 delete mode 100644 0001-Limit-batch-chunk-size-to-2.patch

diff --git a/0001-Limit-batch-chunk-size-to-2.patch b/0001-Limit-batch-chunk-size-to-2.patch
deleted file mode 100644
index ebc6888..0000000
--- a/0001-Limit-batch-chunk-size-to-2.patch
+++ /dev/null
@@ -1,32 +0,0 @@
-From cc7531d1eb4c9042e614f24857df8b764c0c4ea6 Mon Sep 17 00:00:00 2001
-From: asagi4 <130366179+asagi4@users.noreply.github.com>
-Date: Mon, 14 Aug 2023 21:51:39 +0300
-Subject: [PATCH] Limit batch chunk size to 2
-
-This makes AITemplate more reliable
----
- comfy/samplers.py | 8 +-------
- 1 file changed, 1 insertion(+), 7 deletions(-)
-
-diff --git a/comfy/samplers.py b/comfy/samplers.py
-index 28cd466..c92bbdc 100644
---- a/comfy/samplers.py
-+++ b/comfy/samplers.py
-@@ -206,13 +206,7 @@ def sampling_function(model_function, x, timestep, uncond, cond, cond_scale, con
-                         to_batch_temp += [x]
- 
-                 to_batch_temp.reverse()
--                to_batch = to_batch_temp[:1]
--
--                for i in range(1, len(to_batch_temp) + 1):
--                    batch_amount = to_batch_temp[:len(to_batch_temp)//i]
--                    if (len(batch_amount) * first_shape[0] * first_shape[2] * first_shape[3] < max_total_area):
--                        to_batch = batch_amount
--                        break
-+                to_batch = to_batch_temp[:2]
- 
-                 input_x = []
-                 mult = []
--- 
-2.39.3
-
diff --git a/README.md b/README.md
index 9204e60..422977e 100644
--- a/README.md
+++ b/README.md
@@ -114,19 +114,19 @@ Similarly, you can use `AREA(x1 x2, y1 y2, weight)` to specify an area for the p
 Masking does not affect LoRA scheduling unless you set unet weights to 0 for a LoRA.
 
 ## Schedulable LoRAs
-The `ScheduleToModel` node patches a model such that when sampling, it'll switch LoRAs between steps. You can apply the LoRA's effect separately to CLIP conditioning and the unet (model)
+The `ScheduleToModel` node patches a model such that when sampling, it'll switch LoRAs between steps. You can apply the LoRA's effect separately to CLIP conditioning and the unet (model).
 
 For me this seems to be quite slow without the --highvram switch because ComfyUI will shuffle things between the CPU and GPU. YMMV. When things stay on the GPU, it's quite fast.
 
 ## AITemplate support
-Note: The old AITemplate project is unmaintaned. I have a fork at https://github.com/asagi4/AIT that I'll keep on minimal life support until I can fix https://github.com/asagi4/comfyui-prompt-control/issues/15
+Note: Instead of AITemplate, try using [ComfyUI_stable_fast](https://github.com/gameltb/ComfyUI_stable_fast). It seems to work correctly when applied *before* `ScheduleToModel`.
+
+The old AITemplate project is unmaintaned. I have a fork at https://github.com/asagi4/AIT that I'll keep on minimal life support until I can fix https://github.com/asagi4/comfyui-prompt-control/issues/15
 
 LoRA scheduling supports AITemplate.
 
 Due to sampler patching, your AITemplate nodes must be cloned to a directory called `AIT` under `custom_nodes` or the hijack won't find it.
 
-AITemplate sometimes breaks because ComfyUI creates batches larger than what AITemplate can handle. You can apply [this patch](0001-Limit-batch-chunk-size-to-2.patch) to ComfyUI to make AITemplate more reliable. The patch is a hack though, and may slightly deoptimize non-AIT sampling
-
 ## Advanced CLIP encoding
 You can use the syntax `STYLE(weight_interpretation, normalization)` in a prompt to affect how prompts are interpreted.