Tag a non-broken release

Good riddance to the old stuff

Makefile

@@ -1,8 +1,8 @@
 all: format check
 	@echo "Done"
 check:
-	pyflakes *.py */*.py
+	find . -name "*.py" | xargs pyflakes
 format:
-	black -l 120 *.py */*.py
+	find . -name "*.py" | xargs black -l 120
 
 .PHONY: check format all

README.md

@@ -1,25 +1,60 @@
 # ComfyUI prompt control
 
-Nodes for LoRA and prompt scheduling that make basic operations in ComfyUI completely prompt-controllable.
+Control LoRA and prompt scheduling, advanced text encoding, regional prompting, and much more, through your text prompt. Generates dynamic graphs that are literally identical to handcrafted noodle soup.
 
-LoRA and prompt scheduling should produce identical output to the equivalent ComfyUI workflow using multiple samplers or the various conditioning manipulation nodes. If you find situations where this is not the case, please report a bug.
+## Prompt Control v2
+
+Prompt control has been almost completely rewritten. It now uses ComfyUI's lazy execution to build graphs from the text prompt at runtime. This has some advantages:
+
+- ComfyUI will not re-run unchanged parts of generated graphs. This is especially useful for two-pass workflows where previously you'd be forced to re-run the first sampling pass even with filtering. That is no longer the case and it does the right thing.
+- The generated graph is often exactly equivalent to a manually built workflow using native ComfyUI nodes. There are no more weird sampling hooks that could cause problems with other nodes
+
+Prompt Control also comes with `PCTextEncode`, which provides advanced text encoding with many additional features compared to ComfyUI's base `CLIPTextEncode`.
+
+### Removed features
+
+- Prompt interpolation syntax; it was too cumbersome to maintain
+- LoRA block weight integration; ditto, for now.
+
+
+### Is it stable now?
+
+Unless I run into bugs or significant annoyances that require changing the interface, it probably won't change too much, but until I tag 2.0, everything can change.
+
+### Everything broke, where are the old nodes?
+
+If you really need them, you can install the [legacy nodes](https://github.com/asagi4/comfyui-prompt-control-legacy). However, I will not fix bugs in those nodes, and I strongly recommend just migrating your workflows to the new nodes.
+
+You can have both installed at the same time; none of the nodes conflict.
 
 ## What can it do?
 
-Things you can control via the prompt:
-- Prompt editing and filtering without multiple samplers
-- LoRA loading and scheduling (including LoRA block weights)
-- Prompt masking and area control, combining prompts and interpolation
-- SDXL parameters
-- Other miscellaneous things
+See [features](#features) below. Things you can control via the prompt:
+- Prompt editing and filtering without noodle soup
+- LoRA loading and scheduling via ComfyUI's hook system
+- Masking, composition and area control (regional prompting)
+- Prompt operations like `BREAK` and `AND`
+- Weight interpretation types (comfy, A1111, etc.)
+- Prompt masking with [cutoff](#cutoff)
+- And a bunch more
 
-[This example workflow](workflows/example.json?raw=1) implements a two-pass workflow illustrating most scheduling features. (Note: the example still uses deprecated nodes; update pending).
+See the [syntax documentation](doc/syntax.md)
+
+If you find prompt scheduling inconvenient for some reason, `PCTextEncode` can be used as a drop-in replacement for `CLIPTextEncode` to get everything else.
+
+
+[This workflow](workflows/example-lazy.json?raw=1) shows LoRA scheduling and prompt editing and compares it with the same prompt implemented with built-in ComfyUI nodes.
+
+[Here](workflows/example-2pass.json?raw=1) is a two-pass workflow illustrating more features, including custom masks and filtering.
 
 The tools in this repository combine well with the macro and wildcard functionality in [comfyui-utility-nodes](https://github.com/asagi4/comfyui-utility-nodes)
 
+
 ## Requirements
 
-You need to have `lark` installed in your Python environment for parsing to work (If you reuse A1111's venv, it'll already be there)
+For LoRA scheduling to work, you'll need at least version 0.3.7 of ComfyUI (0.3.36 of ComfyUI desktop).
+
+You need to have `lark` installed in your Python environment for parsing to work (If you reuse A1111's venv, it'll already be there).
 
 If you use the portable version of ComfyUI on Windows with its embedded Python, you must open a terminal in the ComfyUI installation directory and run the command:
 ```
@@ -28,229 +63,71 @@ If you use the portable version of ComfyUI on Windows with its embedded Python,
 
 Then restart ComfyUI afterwards.
 
-## Notable changes
+# Core nodes
 
-I try to avoid behavioural changes that break old prompts, but they may happen occasionally.
+## PCLazyTextEncode and PCLazyTextEncodeAdvanced
 
-- 2024-12-03 ComfyUI merged support for model/conditioning hooks. There are two new nodes, `PCEncodeSchedule` and `PCLoraHooksFromSchedule` that can be used in combination with the hook nodes. Some functionality is still missing from them, but going forward, these nodes will be the only nodes supported; **I will not spend significant time fixing bugs in the old monkeypatched nodes anymore.**
-- 2024-02-02 The node will now automatically enable offloading LoRA backup weights to the CPU if you run out of memory during LoRA operations, even when `--highvram` is specified. This change persists until ComfyUI is restarted.
-- 2024-01-14 Multiple `CLIP_L` instances are now joined with a space separator instead of concatenated.
-- 2024-01-09 AITemplate support dropped. I don't recommend or test AITemplate anymore. Use Stable-Fast instead (see below for info)
-- 2024-01-08 Prompt control now enables in-place weight updates on the model. This shouldn't affect anything, but increases performance slightly. You can disable this by setting the environment variable `PC_NO_INPLACE_UPDATE` to any non-empty value.
-- 2023-12-28 MASK now uses ComfyUI's `mask_strength` attribute instead of calculating it on its own. This changes its behaviour slightly.
-- 2023-12-06: Removed `JinjaRender`, `SimpleWildcard`, `ConditioningCutoff`, `CondLinearInterpolate` and `StringConcat`. For the first two, see [this repository](https://github.com/asagi4/comfyui-utility-nodes) for mostly-compatible implementations.
-- 2023-10-04: `STYLE:...` syntax changed to `STYLE(...)`
+`PCLazyTextEncode` uses ComfyUI's lazy graph execution mechanism to generate a graph of `PCTextEncode` and `SetConditioningTimestepRange` nodes from a prompt with schedules. This has the advantage that if a part of the schedule doesn't change, ComfyUI's caching mechanism allows you to avoid re-encoding the non-changed part.
 
-## Note on how schedules work
+for example, if you first encode `[cat:dog:0.1]` and later change that to `[cat:dog:0.5]`, no re-encoding takes place.
+
+for added fun, put `NODE(NodeClassName, textinputname)` in a prompt to generate a graph using **any other node** that's compatible. The node can't have required parameters besides a single CLIP parameter (which must be named `clip`) and the text prompt, and it must return a `CONDITIONING` as its first return value. The "default" values are `PCTextEncode` and `text`.
+
+For example, if you for some reason do not want the advanced features of `PCTextEncode`, use `NODE(CLIPTextEncode)` in the prompt and you'll still get scheduling with ComfyUI's regular TE node.
+
+The advanced node enables filtering the prompt for multi-pass workflows.
+
+## PCLazyLoraLoader and PCLazyLoraLoaderAdvanced
+
+This node reads LoRA expressions from the scheduled prompt and constructs a graph of `LoraLoader`s and `CreateHookLora`s as necessary to provide the necessary LoRA scheduling. Just use it in place of a `LoRALoader` and use the output normally.
+
+The Advanced node gives you access to the generated hooks. If you have `apply_hooks` set to true, you **do not** need to apply the `HOOKS`  output to a CLIP model separately; it's provided in case you want to use it elsewhere. The advanced node also enables filtering the prompt for multi-pass workflows.
+
+## PCTextEncode
+
+Encodes a single prompt with advanced (non-scheduling) syntax enabled. This is what actually does most of the work under the hood.
+
+Note: `PCTextEncode` **does not** ignore `<lora:...:1>` and will treat it as part of the prompt. To use a combined prompt for LoRAs and your input, use `PCLazyTextEncode` and `PCLazyLoraLoader`
+
+## PCAddMaskToCLIP
+
+This node attaches masks to a `CLIP` model so that they can be referred to when using the `IMASK` custom mask function of `PCTextEncode`.
+
+## PCSetTextEncodeSettings
+
+This node configures `PCTextEncode` default values for some functions by attaching the information to a `CLIP` model.
+
+# Features
+## Scheduling and LoRA loading
+
+Prompt control provides a way to easily schedule different prompts and control LoRA loading.
+
+See the [syntax documentation](doc/syntax.md)
+
+### Note on how schedules work
 
 ComfyUI does not use the step number to determine whether to apply conds; instead, it uses the sampler's timestep value which is affected by the scheduler you're using. This means that when the sampler scheduler isn't linear, the schedules generated by prompt control will not be either.
 
-Currently there doesn't seem to be a good way to change this.
-
-You can try using the `PCSplitSampling` node to enable an alternative method of sampling.
-
-# Scheduling syntax
-
-Syntax is like A1111 for now, but only fractions are supported for steps.
-
-```
-a [large::0.1] [cat|dog:0.05] [<lora:somelora:0.5:0.6>::0.5]
-[in a park:in space:0.4]
-```
-
-You can also use `a [b:c:0.3,0.7]` as a shortcut. The prompt be `a` until 0.3, `a b` until 0.7, and then `a c`. `[a:0.1,0.4]` is equivalent to `[a::0.1,0.4]`
-
-## LoRA loading
-
-LoRAs can be loaded by referring to the filename without extension and subdirectories will also be searched. For example, `<lora:cats:1>`. will match both `cats.safetensors` and `sd15/animals/cats.safetensors`. If there are multiple LoRAs with the same name, the first match will be loaded.
-
-Alternatively, the name can include the full directory path relative to ComfyUI's search paths, without extension: `<lora:XL/sdxllora:0.5>`. In this case, the *full* path must match.
-
-If no match is found, the node will try to replace spaces with underscores and search again. That is, `<lora:cats and dogs:1>` will find `cats_and_dogs.safetensors`. This helps with some autocompletion scripts that replace underscores with spaces.
-
-Finally, you can give the exact path (including the extension) as shown in `LoRALoader`.
-
-
-## Alternating
-
-Alternating syntax is `[a|b:pct_steps]`, causing the prompt to alternate every `pct_steps`. `pct_steps` defaults to 0.1 if not specified. You can also have more than two options.
-
-## Sequences
-
-The syntax `[SEQ:a:N1:b:N2:c:N3]` is shorthand for `[a:[b:[c::N3]:N2]:N1]` ie. it switches from `a` to `b` to `c` to nothing at the specified points in sequence.
-
-Might be useful with Jinja templating (see https://github.com/asagi4/comfyui-utility-nodes). For example:
-```
-[SEQ<% for x in steps(0.1, 0.9, 0.1) %>:<lora:test:<= sin(x*pi) + 0.1 =>>:<= x =><% endfor %>]
-```
-generates a LoRA schedule based on a sinewave
-
-## Tag selection
-Using the `FilterSchedule` node, in addition to step percentages, you can use a *tag* to select part of an input:
-```
-a large [dog:cat<lora:catlora:0.5>:SECOND_PASS]
-```
-Set the `tags` parameter in the `FilterSchedule` node to filter the prompt. If the tag matches any tag `tags` (comma-separated), the second option is returned (`cat`, in this case, with the LoRA). Otherwise, the first option is chosen (`dog`, without LoRA).
-
-the values in `tags` are case-insensitive, but the tags in the input **must** be uppercase A-Z and underscores only, or they won't be recognized. That is, `[dog:cat:hr]` will not work.
-
-For example, a prompt
-```
-a [black:blue:X] [cat:dog:Y] [walking:running:Z] in space
-```
-with `tags` `x,z` would result in the prompt `a blue cat running in space`
-
-## Prompt interpolation
-
-Note: Not currently supported by `PCEncodeSchedule`
-
-`a red [INT:dog:cat:0.2,0.8:0.05]` will attempt to interpolate the tensors for `a red dog` and `a red cat` between the specified range in as many steps of 0.05 as will fit.
-
-
-## SDXL
-
-The nodes do not treat SDXL models specially, but there are some utilities that enable SDXL specific functionality.
-
-You can use the function `SDXL(width height, target_width target_height, crop_w crop_h)` to set SDXL prompt parameters. `SDXL()` is equivalent to `SDXL(1024 1024, 1024 1024, 0 0)` unless the default values have been overridden by `PCScheduleSettings`.
-
-To set the `clip_l` prompt, as with `CLIPTextEncodeSDXL`, use the function `CLIP_L(prompt text goes here)`.
-
-Things to note:
-- Multiple instances of `CLIP_L` are joined with a space. That is, `CLIP_L(foo)CLIP_L(bar)` is the same as `CLIP_L(foo bar)`
-- Using `BREAK` isn't supported in it; it'll just parse as the plain word BREAK.
-- similarly, `AND` inside `CLIP_L` does not do anything sensible; `CLIP_L(foo AND bar)` will parse as two prompts `CLIP_L(foo` and `bar)`
-- `CLIP_L` and `SDXL` have no effect on SD 1.5.
-- The rest of the prompt becomes the `clip_g` prompt.
-- If there is no `CLIP_L` or `SDXL`, the prompts will work as with `CLIPTextEncode`.
-
-# Other syntax:
-
-- `<emb:xyz>` is alternative syntax for `embedding:xyz` to work around a syntax conflict with `[embedding:xyz:0.5]` which is parsed as a schedule that switches from `embedding` to `xyz`.
-
-- The keyword `BREAK` causes the prompt to be tokenized in separate chunks, which results in each chunk being individually padded to the text encoder's maximum token length. This is mostly equivalent to the `ConditioningConcat` node.
-
-## Combining prompts
-`AND` can be used to combine prompts. You can also use a weight at the end. It does a weighted sum of each prompt,
-```
-cat :1 AND dog :2
-```
-The weight defaults to 1 and are normalized so that `a:2 AND b:2` is equal to `a AND b`. `AND` is processed after schedule parsing, so you can change the weight mid-prompt: `cat:[1:2:0.5] AND dog`
-
-if there is `COMFYAND()` in the prompt, the behaviour of `AND` will change to work like `ConditioningCombine`, but in practice this seems to be just slower while producing the same output.
-
-## Functions
-
-There are some "functions" that can be included in a prompt to do various things. 
-
-Functions have the form `FUNCNAME(param1, param2, ...)`. How parameters are interpreted is up to the function.
-Note: Whitespace is *not* stripped from string parameters by default. Commas can be escaped with `\,`
-
-Like `AND`, these functions are parsed after regular scheduling syntax has been expanded, allowing things like `[AREA:MASK:0.3](...)`, in case that's somehow useful.
-
-### SHUFFLE and SHIFT
-
-Default parameters: `SHUFFLE(seed=0, separator=,, joiner=,)`, `SHIFT(steps=0, separator=,, joiner=,)`
-
-`SHIFT` moves elements to the left by `steps`. The default is 0 so `SHIFT()` does nothing
-`SHUFFLE` generates a random permutation with `seed` as its seed.
-
-These functions are applied to each prompt chunk **after** `BREAK`, `AND` etc. have been parsed. The prompt is split by `separator`, the operation is applied, and it's then joined back by `joiner`.
-
-Multiple instances of these functions are applied in the order they appear in the prompt.
-
-**NOTE:** These functions are *not* smart about syntax and will break emphasis if the separator occurs inside parentheses. I might fix this at some point, but for now, keep this in mind.
-
-For example:
-- `SHIFT(1) cat, dog, tiger, mouse` does a shift and results in `dog, tiger, mouse, cat`. (whitespace may vary)
-- `SHIFT(1,;) cat, dog ; tiger, mouse` results in `tiger, mouse, cat, dog`
-- `SHUFFLE() cat, dog, tiger, mouse` results in `cat, dog, mouse, tiger`
-- `SHUFFLE() SHIFT(1) cat, dog, tiger, mouse` results in `dog, mouse, tiger, cat`
-
-- `SHIFT(1) cat,dog BREAK tiger,mouse` results in `dog,cat BREAK tiger,mouse`
-- `SHIFT(1) cat, dog AND SHIFT(1) tiger, mouse` results in `dog, cat BREAK mouse, tiger`
-
-Whitespace is *not* stripped and may also be used as a joiner or separator
-- `SHIFT(1,, ) cat,dog` results in `dog cat`
-
-
-### NOISE
-
-The function `NOISE(weight, seed)` adds some random noise into the prompt. The seed is optional, and if not specified, the global RNG is used. `weight` should be between 0 and 1.
-
-### MASK, IMASK and AREA
-
-You can use `MASK(x1 x2, y1 y2, weight, op)` to specify a region mask for a prompt. The values are specified as a percentage with a float between `0` and `1`, or as absolute pixel values (these can't be mixed). `1` will be interpreted as a percentage instead of a pixel value.
-
-Similarly, you can use `AREA(x1 x2, y1 y2, weight)` to specify an area for the prompt (see ComfyUI's area composition examples). The area is calculated by ComfyUI relative to your latent size.
-
-#### Custom masks: IMASK and `PCScheduleAddMasks`
-
-You can attach custom masks to a `PROMPT_SCHEDULE` with the `PCScheduleAddMasks` node and then refer to those masks in the prompt using `IMASK(index, weight, op)`. Indexing starts from zero, so 0 is the first attached mask etc. `PCSCheduleAddMasks` ignores empty inputs, so if you only add a mask to the `mask4` input, it will still have index 0.
-
-Applying `PCScheduleAddMasks` multiple times *appends* masks to a schedule rather than overriding existing ones, so if you need more than 4, you can just use it more than once.
-
-#### Behaviour of masks
-If multiple `MASK`s are specified, they are combined together with ComfyUI's `MaskComposite` node, with `op` specifying the operation to use (default `multiply`). In this case, the combined mask weight can be set with `MASKW(weight)` (defaults to 1.0).
-
-Masks assume a size of `(512, 512)`, unless overridden with `PCScheduleSettings` and pixel values will be relative to that. ComfyUI will scale the mask to match the image resolution. You can change it manually by using `MASK_SIZE(width, height)` anywhere in the prompt,
-
-These are handled per `AND`-ed prompt, so in `prompt1 AND MASK(...) prompt2`, the mask will only affect prompt2.
-
-The default values are `MASK(0 1, 0 1, 1)` and you can omit unnecessary ones, that is, `MASK(0 0.5, 0.3)` is `MASK(0 0.5, 0.3 1, 1)`
-
-Note that because the default values are percentages, `MASK(0 256, 64 512)` is valid, but `MASK(0 200)` will raise an error.
-
-Masking does not affect LoRA scheduling unless you set unet weights to 0 for a LoRA.
-
-### FEATHER
-
-When you use `MASK` or `IMASK`, you can also call `FEATHER(left top right bottom)` to apply feathering using ComfyUI's `FeatherMask` node. The values are in pixels and default to `0`.
-
-If multiple masks are used, `FEATHER` is applied *before compositing* in the order they appear in the prompt, and any leftovers are applied to the combined mask. If you want to skip feathering a mask while compositing, just use `FEATHER()` with no arguments.
-
-For example:
-```
-MASK(1) MASK(2) MASK(3) FEATHER(1) FEATHER() FEATHER(3) weirdmask FEATHER(4)
-```
-
-gives you a mask that is a combination of 1, 2 and 3, where 1 and 3 are feathered before compositing and then `FEATHER(4)` is applied to the composite.
-
-The order of the `FEATHER` and `MASK` calls doesn't matter; you can have `FEATHER` before `MASK` or even interleave them.
-
-# Schedulable LoRAs
-
-Note: Use `PCLoraHooksFromSchedule`. It will work better.
-
-## Old nodes
-The `ScheduleToModel` node patches a model so that when sampling, it'll switch LoRAs between steps. You can apply the LoRA's effect separately to CLIP conditioning and the unet (model).
-
-Swapping LoRAs often can be quite slow without the `--highvram` switch because ComfyUI will shuffle things between the CPU and GPU. When things stay on the GPU, it's quite fast.
-
-If you run out of VRAM during a LoRA swap, the node will attempt to save VRAM by enabling CPU offloading for future generations even in highvram mode. This persists until ComfyUI is restarted.
-
-You can also set the `PC_RETRY_ON_OOM` environment variable to any non-empty value to automatically retry sampling once if VRAM runs out.
-
-## LoRA Block Weight
-
-Note: Not supported by `PCEncodeSchedule` yet
-
-If you have [ComfyUI Inspire Pack](https://github.com/ltdrdata/ComfyUI-Inspire-Pack) installed, you can use its Lora Block Weight syntax, for example:
-
-```
-a prompt <lora:cars:1:LBW=SD-OUTALL;A=1.0;B=0.0;>
-```
-The `;` is optional if there is only 1 parameter.
-The syntax is the same as in the `ImpactWildcard` node, documented [here](https://github.com/ltdrdata/ComfyUI-extension-tutorials/blob/Main/ComfyUI-Impact-Pack/tutorial/ImpactWildcard.md)
-
-# Other integrations
 ## Advanced CLIP encoding
-Note: `perp` is not supported by `PCEncodeSchedule`
 
-You can use the syntax `STYLE(weight_interpretation, normalization)` in a prompt to affect how prompts are interpreted.
+If you use `PCTextEncode`, advanced encodings are available automatically. Thanks to BlenderNeko for the original code.
 
-Without any extra nodes, only `perp` is available, which does the same as [ComfyUI_PerpWeight](https://github.com/bvhari/ComfyUI_PerpWeight) extension.
+Use the syntax `STYLE(weight_interpretation, normalization)` in a prompt to affect how prompts are interpreted.
 
-If you have [Advanced CLIP Encoding nodes](https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb/tree/master) cloned into your `custom_nodes`, more options will be available.
+The weight interpretations available are:
+  - comfy (default)
+  - comfy++
+  - compel
+  - down_weight
+  - A1111
+  - perp
+
+Normalizations are:
+  - none (default)
+  - length
+  - mean
+
+The normalization calculations are independent operations and you can combine them with `+`, eg `STYLE(A1111, length+mean)` or `STYLE(comfy, mean+length)`, or even something silly like `STYLE(perp, mean+length+mean+length)`
 
 The style can be specified separately for each AND:ed prompt, but the first prompt is special; later prompts will "inherit" it as default. For example:
 
@@ -265,86 +142,30 @@ Will interpret the first one using the default ComfyUI behaviour, the second pro
 
 For things (ie. the code imports) to work, the nodes must be cloned in a directory named exactly `ComfyUI_ADV_CLIP_emb`.
 
-## Cutoff node integration
+## Cutoff
 
-Note: Not supported by `PCEncodeSchedule` yet.
+NOTE: Cutoff syntax might change at some point; it's pretty clunky.
 
-If you have [ComfyUI Cutoff](https://github.com/BlenderNeko/ComfyUI_Cutoff) cloned into your `custom_nodes`, you can use the `CUT` keyword to use cutoff functionality
+`PCTextEncode` reimplements cutoff from [ComfyUI Cutoff](https://github.com/BlenderNeko/ComfyUI_Cutoff).
 
 The syntax is
 ```
 a group of animals, [CUT:white cat:white], [CUT:brown dog:brown:0.5:1.0:1.0:_]
 ```
-the parameters in the `CUT` section are `region_text:target_text:weight;strict_mask:start_from_masked:padding_token` of which only the first two are required.
-If `strict_mask`, `start_from_masked` or `padding_token` are specified in more than one section, the last one takes effect for the whole prompt
+You should read the prompt as `a group of animals, white cat, brown dog`, but CUT causes the tokens in `target_tokens` to be masked off from the base prompt in `region_text`, so that their effect can be isolated, and you're less likely to get brown cats or white dogs.
 
-# Nodes
+Target tokens are treated individually, separated by space, for example, `[CUT:green apple, red apple, green leaf:green apple]` will mask *both* greens and the apple, giving you `+ +, red +, + leaf`. To mask out just `green apple`, use `[CUT:green apple, red apple:green_apple]` which will result in a masked prompt of `+ +, red apple`. Escape `_` with a `\`.
 
-## PCLoraHooksFromSchedule
+the parameters in the `CUT` section are `region_text:target_tokens:weight;strict_mask:start_from_masked:padding_token` of which only the first two are required. The default values are `weight=1.0`, `strict_mask=1.0` `start_from_masked=1.0`, `padding_token=+`
 
-Creates a ComfyUI `HOOKS` object from a prompt schedule. Can be attached to a CLIP model to perform encoding and LoRA switching
+If `strict_mask`, `start_from_masked` or `padding_token` are specified in more than one CUT, the *last* one becomes the default for any CUTs afterwards that do not explicitly set the parameters. For example, in:
 
-## PCEncodeSchedule
+`[CUT:white cat:white:0.5] and [CUT:black parrot, flying:black:1.0:0.5] and [CUT:green apple:green]`
 
-Encodes all prompts in a schedule. Pass in a `CLIP` object with hooks attached for LoRA scheduling, then use the resulting `CONDITIONING` normally
+`white cat` will a weight of 0.5, and 1.0 for all parameters, and `black parrot` and `green apple` will *both* have a `strict_mask` parameter of 0.5.
 
-## PromptToSchedule
-Parses a schedule from a text prompt. A schedule is essentially an array of `(valid_until, prompt)` pairs that the other nodes can use.
-
-## FilterSchedule
-Filters a schedule according to its parameters, removing any *changes* that do not occur within `[start, end)`.
-
-The node also does tag filtering if any tags are specified.
-
-Always returns at least the last prompt in the schedule if everything would otherwise be filtered.
-
-`start=0, end=0` returns the prompt at the start and `start=1.0, end=1.0` returns the prompt at the end.
-
-## PCScheduleSettings
-Returns an object representing **default values** for the `SDXL` function and allows configuring `MASK_SIZE` outside the prompt. You need to apply them to a schedule with `PCApplySettings`. Note that for the SDXL settings to apply, you still need to have `SDXL()` in the prompt.
-
-The "steps" parameter currently does nothing; it's for future features.
-
-## PCApplySettings
-Applies the give default values from `PCScheduleSettings` to a schedule
-
-## PCPromptFromSchedule
-
-Extracts a text prompt from a schedule; also logs it to the console.
-LoRAs are *not* included in the text prompt, though they are logged.
-
-## PCScheduleAddMasks
-
-Attaches custom masks to a `PROMPT_SCHEDULE` that can then be used in a prompt.
-
-## ScheduleToCond (deprecated)
-Produces a combined conditioning for the appropriate timesteps. From a schedule. Also applies LoRAs to the CLIP model according to the schedule.
-
-## ScheduleToModel (deprecated)
-Produces a model that'll cause the sampler to reapply LoRAs at specific steps according to the schedule.
-
-This depends on a callback handled by a monkeypatch of the ComfyUI sampler function, so it might not work with custom samplers, but it shouldn't interfere with them either.
-
-## PCSplitSampling (deprecated)
-Causes sampling to be split into multiple sampler calls instead of relying on timesteps for scheduling. This makes the schedules more accurate, but seems to cause weird behaviour with SDE samplers. (Upstream bug?)
-
-
-## PromptControlSimple (deprecated)
-This node exists purely for convenience. It's a combination of `PromptToSchedule`, `ScheduleToCond`, `ScheduleToModel` and `FilterSchedule` such that it provides as output a model, positive conds and negative conds, both with and without any specified filters applied.
-
-This makes it handy for quick one- or two-pass workflows.
-
-## Older nodes
-
-- `EditableCLIPEncode`: A combination of `PromptToSchedule` and `ScheduleToCond`
-- `LoRAScheduler`: A combination of `PromptToSchedule`, `FilterSchedule` and `ScheduleToModel`
+The parameters affect how the masked and unmasked prompts are combined to produce the final embedding. Just play around with them.
 
 # Known issues
 
-- If you use LoRA scheduling in a workflow with `LoRALoader` nodes, you might get inconsistent results. For now, just avoid mixing `ScheduleToModel` or `LoRAScheduler` with `LoRALoader`. See https://github.com/asagi4/comfyui-prompt-control/issues/36
-- Workflows using `SamplerCustom` will calculate LoRA schedules based on the number of sigmas given to the sampler instead of the number of steps, since that information isn't available.
-- `CUT` does not work with `STYLE:perp`
-- `PCSplitSampling` overrides ComfyUI's `BrownianTreeNoiseSampler` noise sampling behaviour so that each split segment doesn't add crazy amounts of noise to the result with some samplers.
-- Split sampling may have weird behaviour if your step percentages go below 1 step.
-- Interpolation is probably buggy and will likely change behaviour whenever code gets refactored.
-- If execution is interrupted and LoRA scheduling is used, your models might be left in an undefined state until you restart ComfyUI
+- None at the moment

__init__.py

@@ -1,19 +1,15 @@
+"""
+@author: asagi4
+@title: ComfyUI Prompt Control
+@nickname: ComfyUI Prompt Control
+@description: Control LoRA and prompt scheduling, advanced text encoding, regional prompting, and much more, through your text prompt. Generates dynamic graphs that are literally identical to handcrafted noodle soup.
+"""
+
 import os
 import sys
 import logging
+import importlib
 
-from .prompt_control.node_clip import EditableCLIPEncode, ScheduleToCond
-from .prompt_control.node_lora import LoRAScheduler, ScheduleToModel, PCSplitSampling, PCWrapGuider
-from .prompt_control.node_other import (
-    PromptToSchedule,
-    FilterSchedule,
-    PCScheduleSettings,
-    PCScheduleAddMasks,
-    PCApplySettings,
-    PCPromptFromSchedule,
-)
-from .prompt_control.node_aio import PromptControlSimple
-from .prompt_control.node_hooks import PCLoraHooksFromSchedule, PCEncodeSchedule
 
 log = logging.getLogger("comfyui-prompt-control")
 log.propagate = False
@@ -22,28 +18,24 @@ if not log.handlers:
     h.setFormatter(logging.Formatter("[%(levelname)s] PromptControl: %(message)s"))
     log.addHandler(h)
 
-if os.environ.get("COMFYUI_PC_DEBUG"):
+if os.environ.get("PROMPTCONTROL_DEBUG"):
     log.setLevel(logging.DEBUG)
 else:
     log.setLevel(logging.INFO)
 
-sys.path.insert(0, os.path.join(os.path.dirname(os.path.realpath(__file__)), "comfy"))
 
+NODE_CLASS_MAPPINGS = {}
+NODE_DISPLAY_NAME_MAPPINGS = {}
 
-NODE_CLASS_MAPPINGS = {
-    "PromptControlSimple": PromptControlSimple,
-    "PromptToSchedule": PromptToSchedule,
-    "PCSplitSampling": PCSplitSampling,
-    "PCScheduleSettings": PCScheduleSettings,
-    "PCScheduleAddMasks": PCScheduleAddMasks,
-    "PCApplySettings": PCApplySettings,
-    "PCPromptFromSchedule": PCPromptFromSchedule,
-    "PCWrapGuider": PCWrapGuider,
-    "FilterSchedule": FilterSchedule,
-    "ScheduleToCond": ScheduleToCond,
-    "ScheduleToModel": ScheduleToModel,
-    "EditableCLIPEncode": EditableCLIPEncode,
-    "LoRAScheduler": LoRAScheduler,
-    "PCLoraHooksFromSchedule": PCLoraHooksFromSchedule,
-    "PCEncodeSchedule": PCEncodeSchedule,
-}
+nodes = ["base", "lazy", "tools"]
+if importlib.util.find_spec("comfy.hooks"):
+    nodes.append("hooks")
+else:
+    log.warning(
+        "Your ComfyUI version is too old, can't import comfy.hooks for PCEncodeSchedule and PCLoraHooksFromSchedule. Update your installation."
+    )
+
+for node in nodes:
+    mod = importlib.import_module(f".prompt_control.nodes_{node}", package=__name__)
+    NODE_CLASS_MAPPINGS.update(mod.NODE_CLASS_MAPPINGS)
+    NODE_DISPLAY_NAME_MAPPINGS.update(mod.NODE_DISPLAY_NAME_MAPPINGS)

doc/syntax.md

@@ -0,0 +1,197 @@
+# Scheduling syntax
+
+Syntax is like A1111 for now, but only fractions are supported for steps. LoRAs are scheduled by including them in a scheduling expression.
+
+```
+a [large::0.1] [cat|dog:0.05] [<lora:somelora:0.5:0.6>::0.5]
+[in a park:in space:0.4]
+```
+
+## Scheduled prompts
+
+There are two forms of scheduled prompts.
+
+### Basic scheduling expressions
+Basic expressions take the form `[before:after:X]` where `X` is the switch point, a decimal number between 0.0 and 1.0 inclusive, representing 0 to 100% of timesteps.
+For example:
+```
+a [red:blue:0.5] cat
+```
+switches from `a red cat` to `a blue cat` at 0.5. `before` and `after` can be arbitrary prompts (`after` can also be empty), including other scheduling expressions, allowing nesting:
+```
+a [red:[blue::0.7]:0.5] cat
+```
+
+switches from `a red cat` to `a blue cat` at 0.5 and to `a cat` at 0.7
+
+
+**Note:** As a special case, `[cat:0.5]` is like `[:cat:0.5]` meaning it switches from empty to `cat` at 0.5. Currently, `[:cat:0.5]` doesn't actually parse correctly, so you **must** use the shortcut form
+
+### Range expressions
+
+You can also use `a [during:after:0.3,0.7]` as a shortcut. The prompt be `a` until 0.3, `a during` until 0.7, and then `a after`. This form is equivalent to `[[during:after:0.7]:0.3]`
+For convenience, `[during:0.1,0.4]` is equivalent to `[during::0.1,0.4]`
+
+## Tag selection
+Using the `FilterSchedule` node, in addition to step percentages, you can use a *tag* to select part of an input:
+```
+a large [dog:cat<lora:catlora:0.5>:SECOND_PASS]
+```
+Set the `tags` parameter in the `FilterSchedule` node to filter the prompt. If the tag matches any tag `tags` (comma-separated), the second option is returned (`cat`, in this case, with the LoRA). Otherwise, the first option is chosen (`dog`, without LoRA).
+
+the values in `tags` are case-insensitive, but the tags in the input **must** be uppercase A-Z and underscores only, or they won't be recognized. That is, `[dog:cat:hr]` will not work.
+
+For example, a prompt
+```
+a [black:blue:X] [cat:dog:Y] [walking:running:Z] in space
+```
+with `tags` `x,z` would result in the prompt `a blue cat running in space`
+
+
+## LoRA Scheduling
+When using the lazy graph building nodes, LoRAs can be scheduled by referring to them in a scheduling expression, like so:
+
+`<lora:fulllora:1> [<lora:partialora:1>::0.5]`
+
+This will schedule `fulllora` for the entire duration of the prompt and `partiallora` until half of sampling is complete.
+
+You can refer to LoRAs by using the filename without extension and subdirectories will also be searched. For example, `<lora:cats:1>`. will match both `cats.safetensors` and `sd15/animals/cats.safetensors`. If there are multiple LoRAs with the same name, the first match will be loaded.
+
+Alternatively, the name can include the full directory path relative to ComfyUI's search paths, without extension: `<lora:XL/sdxllora:0.5>`. In this case, the *full* path must match.
+
+If no match is found, the node will try to replace spaces with underscores and search again. That is, `<lora:cats and dogs:1>` will find `cats_and_dogs.safetensors`. This helps with some autocompletion scripts that replace underscores with spaces.
+
+Finally, you can give the exact path (including the extension) as shown in `LoRALoader`.
+
+
+## Alternating
+
+Alternating syntax is `[a|b:pct_steps]`, causing the prompt to alternate every `pct_steps`. `pct_steps` defaults to 0.1 if not specified. You can also have more than two options.
+
+
+## Sequences
+
+The syntax `[SEQ:a:N1:b:N2:c:N3]` is shorthand for `[a:[b:[c::N3]:N2]:N1]` ie. it switches from `a` to `b` to `c` to nothing at the specified points in sequence.
+
+Might be useful with Jinja templating (see https://github.com/asagi4/comfyui-utility-nodes). For example:
+```
+[SEQ<% for x in steps(0.1, 0.9, 0.1) %>:<lora:test:<= sin(x*pi) + 0.1 =>>:<= x =><% endfor %>]
+```
+generates a LoRA schedule based on a sinewave
+
+
+# Basic prompt syntax
+
+This syntax is also available in outside scheduled prompts, where applicable.
+
+## LoRA loading
+
+The A111-style syntax `<lora:loraname:weight>` can be used to load LoRAs via the prompt. See LoRA scheduling above.
+
+## Combining prompts, A1111-style
+
+- The keyword `BREAK` causes the prompt to be tokenized in separate chunks, which results in each chunk being individually padded to the text encoder's maximum token length. This is mostly equivalent to the `ConditioningConcat` node.
+
+`AND` can be used to combine prompts. You can also use a weight at the end. It does a weighted sum of each prompt,
+```
+cat :1 AND dog :2
+```
+The weight defaults to 1 and are normalized so that `a:2 AND b:2` is equal to `a AND b`. `AND` is processed after schedule parsing, so you can change the weight mid-prompt: `cat:[1:2:0.5] AND dog`
+
+
+## Functions
+
+There are some "functions" that can be included in a prompt to do various things. 
+
+Functions have the form `FUNCNAME(param1, param2, ...)`. How parameters are interpreted is up to the function.
+Note: Whitespace is *not* stripped from string parameters by default. Commas can be escaped with `\,`
+
+Like `AND`, these functions are parsed after regular scheduling syntax has been expanded, allowing things like `[AREA:MASK:0.3](...)`, in case that's somehow useful.
+
+### SDXL
+
+The nodes do not treat SDXL models specially, but there are some utilities that enable SDXL specific functionality.
+
+You can use the function `SDXL(width height, target_width target_height, crop_w crop_h)` to set SDXL prompt parameters. `SDXL()` is equivalent to `SDXL(1024 1024, 1024 1024, 0 0)` unless the default values have been overridden by `PCScheduleSettings`.
+
+To set the `clip_l` prompt, as with `CLIPTextEncodeSDXL`, use the function `CLIP_L(prompt text goes here)`.
+
+Things to note:
+- Multiple instances of `CLIP_L` are joined with a space. That is, `CLIP_L(foo)CLIP_L(bar)` is the same as `CLIP_L(foo bar)`
+- Using `BREAK` isn't supported in it; it'll just parse as the plain word BREAK.
+- similarly, `AND` inside `CLIP_L` does not do anything sensible; `CLIP_L(foo AND bar)` will parse as two prompts `CLIP_L(foo` and `bar)`
+- `CLIP_L` and `SDXL` have no effect on SD 1.5.
+- The rest of the prompt becomes the `clip_g` prompt.
+- If there is no `CLIP_L` or `SDXL`, the prompts will work as with `CLIPTextEncode`.
+
+### SHUFFLE and SHIFT
+
+Default parameters: `SHUFFLE(seed=0, separator=,, joiner=,)`, `SHIFT(steps=0, separator=,, joiner=,)`
+
+`SHIFT` moves elements to the left by `steps`. The default is 0 so `SHIFT()` does nothing
+`SHUFFLE` generates a random permutation with `seed` as its seed.
+
+These functions are applied to each prompt chunk **after** `BREAK`, `AND` etc. have been parsed. The prompt is split by `separator`, the operation is applied, and it's then joined back by `joiner`.
+
+Multiple instances of these functions are applied in the order they appear in the prompt.
+
+**NOTE:** These functions are *not* smart about syntax and will break emphasis if the separator occurs inside parentheses. I might fix this at some point, but for now, keep this in mind.
+
+For example:
+- `SHIFT(1) cat, dog, tiger, mouse` does a shift and results in `dog, tiger, mouse, cat`. (whitespace may vary)
+- `SHIFT(1,;) cat, dog ; tiger, mouse` results in `tiger, mouse, cat, dog`
+- `SHUFFLE() cat, dog, tiger, mouse` results in `cat, dog, mouse, tiger`
+- `SHUFFLE() SHIFT(1) cat, dog, tiger, mouse` results in `dog, mouse, tiger, cat`
+
+- `SHIFT(1) cat,dog BREAK tiger,mouse` results in `dog,cat BREAK tiger,mouse`
+- `SHIFT(1) cat, dog AND SHIFT(1) tiger, mouse` results in `dog, cat BREAK mouse, tiger`
+
+Whitespace is *not* stripped and may also be used as a joiner or separator
+- `SHIFT(1,, ) cat,dog` results in `dog cat`
+
+### NOISE
+
+The function `NOISE(weight, seed)` adds some random noise into the prompt. The seed is optional, and if not specified, the global RNG is used. `weight` should be between 0 and 1.
+
+### MASK, IMASK and AREA
+
+You can use `MASK(x1 x2, y1 y2, weight, op)` to specify a region mask for a prompt. The values are specified as a percentage with a float between `0` and `1`, or as absolute pixel values (these can't be mixed). `1` will be interpreted as a percentage instead of a pixel value.
+
+Similarly, you can use `AREA(x1 x2, y1 y2, weight)` to specify an area for the prompt (see ComfyUI's area composition examples). The area is calculated by ComfyUI relative to your latent size.
+
+#### Custom masks: IMASK and `PCAddMaskToCLIP`
+
+You can attach custom masks to a `CLIP` with the `PC: Attach Mask` nodes and then refer to those masks in the prompt using `IMASK(index, weight, op)`. Indexing starts from zero, so 0 is the first attached mask etc. `PCSCheduleAddMasks` ignores empty inputs, so if you only add a mask to the `mask4` input, it will still have index 0.
+
+Applying the nodes multiple times *appends* masks rather than overriding existing ones, so if you need more than 4, you can just use it more than once.
+
+#### Behaviour of masks
+If multiple `MASK`s are specified, they are combined together with ComfyUI's `MaskComposite` node, with `op` specifying the operation to use (default `multiply`). In this case, the combined mask weight can be set with `MASKW(weight)` (defaults to 1.0).
+
+Masks assume a size of `(512, 512)`, unless overridden with `PC: Configure PCTextEncode` and pixel values will be relative to that. ComfyUI will scale the mask to match the image resolution. You can change it manually by using `MASK_SIZE(width, height)` anywhere in the prompt,
+
+These are handled per `AND`-ed prompt, so in `prompt1 AND MASK(...) prompt2`, the mask will only affect prompt2.
+
+The default values are `MASK(0 1, 0 1, 1)` and you can omit unnecessary ones, that is, `MASK(0 0.5, 0.3)` is `MASK(0 0.5, 0.3 1, 1)`
+
+Note that because the default values are percentages, `MASK(0 256, 64 512)` is valid, but `MASK(0 200)` will raise an error.
+
+Masking does not affect LoRA scheduling unless you set unet weights to 0 for a LoRA.
+
+### FEATHER
+
+When you use `MASK` or `IMASK`, you can also call `FEATHER(left top right bottom)` to apply feathering using ComfyUI's `FeatherMask` node. The values are in pixels and default to `0`.
+
+If multiple masks are used, `FEATHER` is applied *before compositing* in the order they appear in the prompt, and any leftovers are applied to the combined mask. If you want to skip feathering a mask while compositing, just use `FEATHER()` with no arguments.
+
+For example:
+```
+MASK(1) MASK(2) MASK(3) FEATHER(1) FEATHER() FEATHER(3) weirdmask FEATHER(4)
+```
+
+gives you a mask that is a combination of 1, 2 and 3, where 1 and 3 are feathered before compositing and then `FEATHER(4)` is applied to the composite.
+
+The order of the `FEATHER` and `MASK` calls doesn't matter; you can have `FEATHER` before `MASK` or even interleave them.
+
+## Miscellaneous
+- `<emb:xyz>` is alternative syntax for `embedding:xyz` to work around a syntax conflict with `[embedding:xyz:0.5]` which is parsed as a schedule that switches from `embedding` to `xyz`.

prompt_control/adv_encode.py

@@ -0,0 +1,242 @@
+import torch
+import numpy as np
+import itertools
+
+
+def _grouper(n, iterable):
+    it = iter(iterable)
+    while True:
+        chunk = list(itertools.islice(it, n))
+        if not chunk:
+            return
+        yield chunk
+
+
+def _norm_mag(w, n):
+    d = w - 1
+    return 1 + np.sign(d) * np.sqrt(np.abs(d) ** 2 / n)
+    # return  np.sign(w) * np.sqrt(np.abs(w)**2 / n)
+
+
+def weights_like(weights, emb):
+    return torch.tensor(weights, dtype=emb.dtype, device=emb.device).reshape(1, -1, 1).expand(emb.shape)
+
+
+def divide_length(word_ids, weights):
+    sums = dict(zip(*np.unique(word_ids, return_counts=True)))
+    sums[0] = 1
+    weights = [[_norm_mag(w, sums[id]) if id != 0 else 1.0 for w, id in zip(x, y)] for x, y in zip(weights, word_ids)]
+    return weights
+
+
+def shift_mean_weight(word_ids, weights):
+    delta = 1 - np.mean([w for x, y in zip(weights, word_ids) for w, id in zip(x, y) if id != 0])
+    weights = [[w if id == 0 else w + delta for w, id in zip(x, y)] for x, y in zip(weights, word_ids)]
+    return weights
+
+
+def scale_to_norm(weights, word_ids, w_max):
+    top = np.max(weights)
+    w_max = min(top, w_max)
+    weights = [[w_max if id == 0 else (w / top) * w_max for w, id in zip(x, y)] for x, y in zip(weights, word_ids)]
+    return weights
+
+
+def mask_word_id(tokens, word_ids, target_id, mask_token):
+    new_tokens = [[mask_token if wid == target_id else t for t, wid in zip(x, y)] for x, y in zip(tokens, word_ids)]
+    mask = np.array(word_ids) == target_id
+    return (new_tokens, mask)
+
+
+def batched_clip_encode(tokens, length, encode_func, num_chunks):
+    embs = []
+    for e in _grouper(32, tokens):
+        enc, pooled = encode_func(e)
+        enc = enc.reshape((len(e), length, -1))
+        embs.append(enc)
+    embs = torch.cat(embs)
+    embs = embs.reshape((len(tokens) // num_chunks, length * num_chunks, -1))
+    return embs
+
+
+def from_masked(tokens, weights, word_ids, base_emb, length, encode_func, m_token=266):
+    pooled_base = base_emb[0, length - 1 : length, :]
+    wids, inds = np.unique(np.array(word_ids).reshape(-1), return_index=True)
+    weight_dict = dict((id, w) for id, w in zip(wids, np.array(weights).reshape(-1)[inds]) if w != 1.0)
+
+    if len(weight_dict) == 0:
+        return torch.zeros_like(base_emb), base_emb[0, length - 1 : length, :]
+
+    weight_tensor = weights_like(weights, base_emb)
+
+    # m_token = (clip.tokenizer.end_token, 1.0) if  clip.tokenizer.pad_with_end else (0,1.0)
+    # TODO: find most suitable masking token here
+    m_token = (m_token, 1.0)
+
+    ws = []
+    masked_tokens = []
+    masks = []
+
+    # create prompts
+    for id, w in weight_dict.items():
+        masked, m = mask_word_id(tokens, word_ids, id, m_token)
+        masked_tokens.extend(masked)
+        masks.append(weights_like(m, base_emb))
+
+        ws.append(w)
+
+    # batch process prompts
+    embs = batched_clip_encode(masked_tokens, length, encode_func, len(tokens))
+    masks = torch.cat(masks)
+
+    embs = base_emb.expand(embs.shape) - embs
+    pooled = embs[0, length - 1 : length, :]
+
+    embs *= masks
+    embs = embs.sum(axis=0, keepdim=True)
+
+    pooled_start = pooled_base.expand(len(ws), -1)
+    ws = torch.tensor(ws).reshape(-1, 1).expand(pooled_start.shape)
+    pooled = (pooled - pooled_start) * (ws - 1)
+    pooled = pooled.mean(axis=0, keepdim=True)
+
+    return ((weight_tensor - 1) * embs), pooled_base + pooled
+
+
+def mask_inds(tokens, inds, mask_token):
+    clip_len = len(tokens[0])
+    inds_set = set(inds)
+    new_tokens = [
+        [mask_token if i * clip_len + j in inds_set else t for j, t in enumerate(x)] for i, x in enumerate(tokens)
+    ]
+    return new_tokens
+
+
+def down_weight(tokens, weights, word_ids, base_emb, length, encode_func, m_token=266):
+    w, w_inv = np.unique(weights, return_inverse=True)
+
+    if np.sum(w < 1) == 0:
+        return base_emb, tokens, base_emb[0, length - 1 : length, :]
+    # m_token = (clip.tokenizer.end_token, 1.0) if  clip.tokenizer.pad_with_end else (0,1.0)
+    # using the comma token as a masking token seems to work better than aos tokens for SD 1.x
+    m_token = (m_token, 1.0)
+
+    masked_tokens = []
+
+    masked_current = tokens
+    for i in range(len(w)):
+        if w[i] >= 1:
+            continue
+        masked_current = mask_inds(masked_current, np.where(w_inv == i)[0], m_token)
+        masked_tokens.extend(masked_current)
+
+    embs = batched_clip_encode(masked_tokens, length, encode_func, len(tokens))
+    embs = torch.cat([base_emb, embs])
+    w = w[w <= 1.0]
+    w_mix = np.diff([0] + w.tolist())
+    w_mix = torch.tensor(w_mix, dtype=embs.dtype, device=embs.device).reshape((-1, 1, 1))
+
+    weighted_emb = (w_mix * embs).sum(axis=0, keepdim=True)
+    return weighted_emb, masked_current, weighted_emb[0, length - 1 : length, :]
+
+
+def scale_emb_to_mag(base_emb, weighted_emb):
+    norm_base = torch.linalg.norm(base_emb)
+    norm_weighted = torch.linalg.norm(weighted_emb)
+    embeddings_final = (norm_base / norm_weighted) * weighted_emb
+    return embeddings_final
+
+
+def recover_dist(base_emb, weighted_emb):
+    fixed_std = (base_emb.std() / weighted_emb.std()) * (weighted_emb - weighted_emb.mean())
+    embeddings_final = fixed_std + (base_emb.mean() - fixed_std.mean())
+    return embeddings_final
+
+
+def perp_weight(weights, unweighted_embs, empty_embs):
+    unweighted, unweighted_pooled = unweighted_embs
+    zero, zero_pooled = empty_embs
+
+    weights = weights_like(weights, unweighted)
+
+    if zero.shape != unweighted.shape:
+        zero = zero.repeat(1, unweighted.shape[1] // zero.shape[1], 1)
+
+    perp = (
+        torch.mul(zero, unweighted).sum(dim=-1, keepdim=True) / (unweighted.norm(dim=-1, keepdim=True) ** 2)
+    ) * unweighted
+
+    over1 = weights.abs() > 1.0
+    result = unweighted + weights * perp
+    result[~over1] = (unweighted - (1 - weights) * perp)[~over1]
+    result[weights == 0.0] = zero[weights == 0.0]
+
+    return result, unweighted_pooled
+
+
+def advanced_encode_from_tokens(
+    tokenized,
+    token_normalization,
+    weight_interpretation,
+    encode_func,
+    m_token=266,
+    length=77,
+    w_max=1.0,
+    return_pooled=False,
+    apply_to_pooled=False,
+    **extra_args
+):
+    tokens = [[t for t, _, _ in x] for x in tokenized]
+    weights = [[w for _, w, _ in x] for x in tokenized]
+    word_ids = [[wid for _, _, wid in x] for x in tokenized]
+
+    for op in token_normalization.split("+"):
+        op = op.strip()
+        if op == "length":
+            # distribute down/up weights over word lengths
+            weights = divide_length(word_ids, weights)
+        if op == "mean":
+            weights = shift_mean_weight(word_ids, weights)
+
+    pooled = None
+
+    if weight_interpretation == "comfy":
+        weighted_tokens = [[(t, w) for t, w in zip(x, y)] for x, y in zip(tokens, weights)]
+        weighted_emb, pooled_base = encode_func(weighted_tokens)
+        pooled = pooled_base
+    else:
+        unweighted_tokens = [[(t, 1.0) for t, _, _ in x] for x in tokenized]
+        base_emb, pooled_base = encode_func(unweighted_tokens)
+
+    if weight_interpretation == "A1111":
+        weighted_emb = base_emb * weights_like(weights, base_emb)  # from_zero
+        weighted_emb = (base_emb.mean() / weighted_emb.mean()) * weighted_emb  # renormalize
+        pooled = pooled_base
+
+    if weight_interpretation == "compel":
+        pos_tokens = [[(t, w) if w >= 1.0 else (t, 1.0) for t, w in zip(x, y)] for x, y in zip(tokens, weights)]
+        weighted_emb, _ = encode_func(pos_tokens)
+        weighted_emb, _, pooled = down_weight(pos_tokens, weights, word_ids, weighted_emb, length, encode_func)
+
+    if weight_interpretation == "comfy++":
+        weighted_emb, tokens_down, _ = down_weight(unweighted_tokens, weights, word_ids, base_emb, length, encode_func)
+        weights = [[w if w > 1.0 else 1.0 for w in x] for x in weights]
+        # unweighted_tokens = [[(t,1.0) for t, _,_ in x] for x in tokens_down]
+        embs, pooled = from_masked(unweighted_tokens, weights, word_ids, base_emb, length, encode_func)
+        weighted_emb += embs
+
+    if weight_interpretation == "down_weight":
+        weights = scale_to_norm(weights, word_ids, w_max)
+        weighted_emb, _, pooled = down_weight(unweighted_tokens, weights, word_ids, base_emb, length, encode_func)
+
+    if weight_interpretation == "perp":
+        weighted_emb, pooled = perp_weight(
+            weights, (base_emb, pooled_base), encode_func(extra_args["tokenizer"].tokenize_with_weights(""))
+        )
+
+    if return_pooled:
+        if apply_to_pooled:
+            return weighted_emb, pooled
+        else:
+            return weighted_emb, pooled_base
+    return weighted_emb, None

prompt_control/cutoff.py

@@ -0,0 +1,222 @@
+import torch
+import copy
+import re
+
+import numpy as np
+import logging
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+def replace_embeddings(max_token, prompt, replacements=None):
+    """Replaces embedding tensors in a token array and replaces them with increasing IDs past max_token"""
+
+    if replacements is None:
+        emb_lookup = []
+    else:
+        emb_lookup = replacements.copy()
+        max_token += len(emb_lookup)
+
+    def get_replacement(embedding):
+        for e, n in emb_lookup:
+            if torch.equal(embedding, e):
+                return n
+        return None
+
+    tokens = []
+    for x in prompt:
+        row = []
+        for i in range(len(x)):
+            emb = x[i][0]
+            if not torch.is_tensor(emb):
+                row.append(emb)
+            else:
+                n = get_replacement(emb)
+                if n is not None:
+                    row.append(n)
+                else:
+                    max_token += 1
+                    row.append(max_token)
+                    emb_lookup.append((emb, max_token))
+        tokens.append(row)
+    tokens = np.array(tokens)[:, 1:-1].reshape(-1)
+    return (tokens, emb_lookup)
+
+
+def unpad_prompt(pad_token, prompt):
+    res = np.trim_zeros(prompt, "b")
+    return np.trim_zeros(res - pad_token, "b") + pad_token
+
+
+def get_sublists(super_list, sub_list):
+    positions = []
+    for candidate_ind in (i for i, e in enumerate(super_list) if e == sub_list[0]):
+        if super_list[candidate_ind : candidate_ind + len(sub_list)] == sub_list:
+            positions.append(candidate_ind)
+    return positions
+
+
+def cutoff_add_region(
+    clip_regions, tokenizer, region_text, target_text, weight, strict_mask, start_from_masked, mask_token
+):
+    """Adds a cut region to the clip_regions dictionary. It is modified in place"""
+    base_tokens = clip_regions["base_tokens"]
+    region_outputs = []
+    target_outputs = []
+    if strict_mask is not None:
+        clip_regions["strict_mask"] = float(strict_mask)
+    if start_from_masked is not None:
+        clip_regions["start_from_masked"] = float(start_from_masked)
+    if mask_token is not None:
+        clip_regions["mask_token"] = tokenizer.tokenizer(mask_token)["input_ids"][1]
+    if weight is None:
+        weight = 1.0
+    else:
+        weight = float(weight)
+
+    region_text = region_text.strip()
+    target_text = target_text.strip()
+
+    strict_mask = clip_regions["strict_mask"]
+    start_from_masked = clip_regions["start_from_masked"]
+    mask_token = clip_regions["mask_token"]
+    log.info(f"CUT region {region_text=} {target_text=} {weight=} {strict_mask=} {start_from_masked=} {mask_token=}")
+
+    pad_token = tokenizer.end_token
+
+    prompt_tokens, emb_lookup = replace_embeddings(pad_token, base_tokens)
+
+    for rt in region_text.split("\n"):
+        region_tokens = tokenizer.tokenize_with_weights(rt)
+        region_tokens, _ = replace_embeddings(pad_token, region_tokens, emb_lookup)
+        region_tokens = unpad_prompt(pad_token, region_tokens).tolist()
+
+        # calc region mask
+        region_length = len(region_tokens)
+        regions = get_sublists(list(prompt_tokens), region_tokens)
+
+        region_mask = np.zeros(len(prompt_tokens))
+        for r in regions:
+            region_mask[r : r + region_length] = 1
+        region_mask = region_mask.reshape(-1, tokenizer.max_length - 2)
+        region_mask = np.pad(region_mask, pad_width=((0, 0), (1, 1)), mode="constant", constant_values=0)
+        region_mask = region_mask.reshape(1, -1)
+        region_outputs.append(region_mask)
+
+        # calc target mask
+        targets = []
+        for target in target_text.split(" "):
+            # deal with underscores
+            target = re.sub(r"(?<!\\)_", " ", target)
+            target = re.sub(r"\\_", "_", target)
+
+            target_tokens = tokenizer.tokenize_with_weights(target)
+            target_tokens, _ = replace_embeddings(pad_token, target_tokens, emb_lookup)
+            target_tokens = unpad_prompt(pad_token, target_tokens).tolist()
+
+            targets.extend([(x, len(target_tokens)) for x in get_sublists(region_tokens, target_tokens)])
+        targets = [(t_start + r, t_start + t_end + r) for r in regions for t_start, t_end in targets]
+
+        targets_mask = np.zeros(len(prompt_tokens))
+        for t_start, t_end in targets:
+            targets_mask[t_start:t_end] = 1
+        targets_mask = targets_mask.reshape(-1, tokenizer.max_length - 2)
+        targets_mask = np.pad(targets_mask, pad_width=((0, 0), (1, 1)), mode="constant", constant_values=0)
+        targets_mask = targets_mask.reshape(1, -1)
+        target_outputs.append(targets_mask)
+
+    # prepare output
+    region_mask_list = clip_regions["regions"].copy()
+    region_mask_list.extend(region_outputs)
+    target_mask_list = clip_regions["targets"].copy()
+    target_mask_list.extend(target_outputs)
+    weight_list = clip_regions["weights"].copy()
+    weight_list.extend([weight] * len(region_outputs))
+
+    clip_regions["regions"] = region_mask_list
+    clip_regions["targets"] = target_mask_list
+    clip_regions["weights"] = weight_list
+
+
+def create_masked_prompt(weighted_tokens, mask, mask_token):
+    mask_ids = list(zip(*np.nonzero(mask.reshape((len(weighted_tokens), -1)))))
+    new_prompt = copy.deepcopy(weighted_tokens)
+    for x, y in mask_ids:
+        new_prompt[x][y] = (mask_token,) + new_prompt[x][y][1:]
+    return new_prompt
+
+
+def process_cuts(encode, extra, tokens):
+    if not extra.get("cuts"):
+        return encode(tokens)
+
+    base = {
+        "base_tokens": tokens,
+        "regions": [],
+        "targets": [],
+        "weights": [],
+        "strict_mask": 1.0,
+        "start_from_masked": 1.0,
+        "mask_token": extra["tokenizer"].tokenizer("+")["input_ids"][1],
+    }
+
+    for cut in extra["cuts"]:
+        cutoff_add_region(base, extra["tokenizer"], *cut)
+
+    return encode_regions(base, encode, extra["tokenizer"])
+
+
+def debug_tokens(label, prompt, tokenizer):
+    log.debug("Tokens for %s", label)
+    for tokens in prompt:
+        tokens = (t for t in tokens if not torch.is_tensor(t[0]))
+        log.debug(" ".join(f"{x[0][0]} {x[1]}" for x in tokenizer.untokenize(tokens) if x[0][0] != tokenizer.end_token))
+
+
+def encode_regions(clip_regions, encode, tokenizer):
+    base_weighted_tokens = clip_regions["base_tokens"]
+    start_from_masked = clip_regions["start_from_masked"]
+    mask_token = clip_regions["mask_token"]
+    strict_mask = clip_regions["strict_mask"]
+
+    # calc base embedding
+    base_embedding_full, pool = encode(base_weighted_tokens)
+
+    # Avoid numpy value error and passthrough base embeddings if no regions are set.
+
+    # calc global target mask
+    global_target_mask = np.any(np.stack(clip_regions["targets"]), axis=0).astype(int)
+
+    # calc global region mask
+    global_region_mask = np.any(np.stack(clip_regions["regions"]), axis=0).astype(float)
+    regions_sum = np.sum(np.stack(clip_regions["regions"]), axis=0)
+    regions_normalized = np.divide(1, regions_sum, out=np.zeros_like(regions_sum), where=regions_sum != 0)
+
+    # mask base embeddings
+    base_masked_prompt = create_masked_prompt(base_weighted_tokens, global_target_mask, mask_token)
+    debug_tokens("base_masked", base_masked_prompt, tokenizer)
+    base_embedding_masked, _ = encode(base_masked_prompt)
+    base_embedding_start = base_embedding_full * (1 - start_from_masked) + base_embedding_masked * start_from_masked
+    base_embedding_outer = base_embedding_full * (1 - strict_mask) + base_embedding_masked * strict_mask
+
+    region_embeddings = []
+    for region, target, weight in zip(clip_regions["regions"], clip_regions["targets"], clip_regions["weights"]):
+        region_masking = torch.tensor(
+            regions_normalized * region * weight, dtype=base_embedding_full.dtype, device=base_embedding_full.device
+        ).unsqueeze(-1)
+
+        region_prompt = create_masked_prompt(base_weighted_tokens, global_target_mask - target, mask_token)
+        debug_tokens("region", region_prompt, tokenizer)
+        region_emb, _ = encode(region_prompt)
+        region_emb -= base_embedding_start
+        region_emb *= region_masking
+
+        region_embeddings.append(region_emb)
+    region_embeddings = torch.stack(region_embeddings).sum(axis=0)
+
+    embeddings_final_mask = torch.tensor(
+        global_region_mask, dtype=base_embedding_full.dtype, device=base_embedding_full.device
+    ).unsqueeze(-1)
+    embeddings_final = base_embedding_start * embeddings_final_mask + base_embedding_outer * (1 - embeddings_final_mask)
+    embeddings_final += region_embeddings
+    return embeddings_final, pool

prompt_control/hijack.py

@@ -1,160 +0,0 @@
-from .utils import get_callback, unpatch_model
-import sys
-
-import logging
-import gc
-import comfy.model_management
-import os
-
-log = logging.getLogger("comfyui-prompt-control")
-
-
-def has_hijack(obj):
-    return hasattr(obj, "pc_hijack_done")
-
-
-def hijack(obj, attr, replacement):
-    setattr(obj, attr, replacement)
-    setattr(replacement, "pc_hijack_done", True)
-
-
-def hijack_sampler(module, function, is_custom):
-    mod = sys.modules[module]
-    orig_sampler = getattr(mod, function)
-    if has_hijack(orig_sampler):
-        return
-
-    from comfy.k_diffusion.sampling import BrownianTreeNoiseSampler
-
-    def pc_sample(*args, **kwargs):
-        model = args[0]
-        cb = get_callback(model)
-        BrownianTreeNoiseSampler.pc_reset(
-            model.model_options.get("pc_split_sampling"),
-            kwargs.get("force_full_denoise") or kwargs.get("denoise", 1.0) >= 1.0,
-        )
-        if cb:
-            try:
-                try:
-                    r = cb(orig_sampler, is_custom, *args, **kwargs)
-                except comfy.model_management.OOM_EXCEPTION:
-                    if not os.environ.get("PC_RETRY_ON_OOM"):
-                        raise
-                    log.error("Got OOM while sampling, freeing memory and retrying once...")
-                    unpatch_model(model)
-                    BrownianTreeNoiseSampler.pc_reset(False)
-                    gc.collect()
-                    comfy.model_management.soft_empty_cache()
-                    r = cb(orig_sampler, is_custom, *args, **kwargs)
-            except Exception:
-                log.error("Exception occurred during callback, unpatching model.")
-                unpatch_model(model)
-                BrownianTreeNoiseSampler.pc_reset(False)
-                raise
-        else:
-            r = orig_sampler(*args, **kwargs)
-        BrownianTreeNoiseSampler.pc_reset()
-        return r
-
-    hijack(mod, function, pc_sample)
-
-
-def hijack_ksampler(module, cls):
-    mod = sys.modules[module]
-    orig_sampler = getattr(mod, cls)
-    if has_hijack(orig_sampler):
-        return
-
-    class HijackedKSampler(orig_sampler):
-        def sample(
-            self,
-            noise,
-            positive,
-            negative,
-            cfg,
-            latent_image=None,
-            start_step=None,
-            last_step=None,
-            force_full_denoise=False,
-            denoise_mask=None,
-            sigmas=None,
-            callback=None,
-            disable_pbar=False,
-            seed=None,
-        ):
-            if sigmas is None:
-                sigmas = self.sigmas
-
-            from comfy.k_diffusion.sampling import BrownianTreeNoiseSampler
-
-            BrownianTreeNoiseSampler.set_global_sigmas(self.sigmas)
-
-            return super().sample(
-                noise,
-                positive,
-                negative,
-                cfg,
-                latent_image,
-                start_step,
-                last_step,
-                force_full_denoise,
-                denoise_mask,
-                sigmas,
-                callback,
-                disable_pbar,
-                seed,
-            )
-
-    hijack(mod, cls, HijackedKSampler)
-
-
-def hijack_browniannoisesampler(module, cls):
-    mod = sys.modules[module]
-    orig_sampler = getattr(mod, cls)
-    if has_hijack(orig_sampler):
-        return
-
-    class PCBrownianTreeNoiseSampler(orig_sampler):
-        global_instance = None
-        use_global_sigmas = False
-        global_sigmas = None
-        force_full_denoise = False
-
-        @classmethod
-        def pc_reset(cls, use_global_sigmas=False, force_full_denoise=False):
-            cls.global_instance = None
-            cls.global_sigmas = None
-            cls.use_global_sigmas = use_global_sigmas
-            cls.force_full_denoise = force_full_denoise
-
-        @classmethod
-        def set_global_sigmas(cls, sigmas):
-            if cls.global_sigmas is None and cls.use_global_sigmas:
-                cls.global_sigmas = (0 if cls.force_full_denoise else sigmas[sigmas > 0].min(), sigmas.max())
-                log.info(
-                    "Initializing BrownianTreeNoiseSampler instance with global sigmas %s, %s",
-                    cls.global_sigmas,
-                    cls.force_full_denoise,
-                )
-
-        def __init__(self, x, sigma_min, sigma_max, **kwargs):
-            if self.global_sigmas is not None:
-                sigma_min, sigma_max = self.global_sigmas
-            if not self.global_instance:
-                super().__init__(x, sigma_min, sigma_max, **kwargs)
-                PCBrownianTreeNoiseSampler.global_instance = self
-
-        def __call__(self, *args, **kwargs):
-            if self.global_instance and self != self.global_instance:
-                return self.global_instance(*args, **kwargs)
-            else:
-                return super().__call__(*args, **kwargs)
-
-    hijack(mod, cls, PCBrownianTreeNoiseSampler)
-
-
-def do_hijack():
-    hijack_browniannoisesampler("comfy.k_diffusion.sampling", "BrownianTreeNoiseSampler")
-    hijack_sampler("comfy.sample", "sample", False)
-    hijack_sampler("comfy.sample", "sample_custom", True)
-    hijack_ksampler("comfy.samplers", "KSampler")

prompt_control/node_aio.py

@@ -1,48 +0,0 @@
-from .node_clip import control_to_clip_common
-from .node_lora import schedule_lora_common
-from .parser import parse_prompt_schedules
-
-
-class PromptControlSimple:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "clip": ("CLIP",),
-                "positive": ("STRING", {"multiline": True}),
-                "negative": ("STRING", {"multiline": True}),
-            },
-            "optional": {
-                "tags": ("STRING", {"default": ""}),
-                "start": ("FLOAT", {"min": 0.0, "max": 1.0, "step": 0.1, "default": 0.0}),
-                "end": ("FLOAT", {"min": 0.0, "max": 1.0, "step": 0.1, "default": 1.0}),
-            },
-        }
-
-    RETURN_TYPES = ("MODEL", "CONDITIONING", "CONDITIONING", "MODEL", "CONDITIONING", "CONDITIONING")
-    RETURN_NAMES = ("model", "positive", "negative", "model_filtered", "pos_filtered", "neg_filtered")
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, model, clip, positive, negative, tags="", start=0.0, end=1.0):
-        lora_cache = {}
-        cond_cache = {}
-        pos_sched = parse_prompt_schedules(positive)
-        pos_cond = pos_filtered = control_to_clip_common(clip, pos_sched, lora_cache, cond_cache)
-
-        neg_sched = parse_prompt_schedules(negative)
-        neg_cond = neg_filtered = control_to_clip_common(clip, neg_sched, lora_cache, cond_cache)
-
-        new_model = model_filtered = schedule_lora_common(model, pos_sched, lora_cache)
-
-        if [tags.strip(), start, end] != ["", 0.0, 1.0]:
-            pos_filtered = control_to_clip_common(
-                clip, pos_sched.with_filters(tags, start, end), lora_cache, cond_cache
-            )
-            neg_filtered = control_to_clip_common(
-                clip, neg_sched.with_filters(tags, start, end), lora_cache, cond_cache
-            )
-            model_filtered = schedule_lora_common(model, pos_sched.with_filters(tags, start, end), lora_cache)
-
-        return (new_model, pos_cond, neg_cond, model_filtered, pos_filtered, neg_filtered)

prompt_control/node_clip.py

@@ -1,703 +0,0 @@
-import logging
-import re
-import torch
-from . import utils as utils
-from .parser import parse_prompt_schedules, parse_cuts
-from .utils import Timer, equalize, safe_float, get_function, parse_floats
-from .perp_weight import perp_encode
-from comfy_extras.nodes_mask import FeatherMask, MaskComposite
-from node_helpers import conditioning_set_values
-
-
-log = logging.getLogger("comfyui-prompt-control")
-
-try:
-    from custom_nodes.ComfyUI_ADV_CLIP_emb.adv_encode import (
-        advanced_encode_from_tokens,
-        encode_token_weights_l,
-        encode_token_weights_g,
-        prepareXL,
-        encode_token_weights,
-    )
-
-    have_advanced_encode = True
-    AVAILABLE_STYLES = ["comfy", "A1111", "compel", "comfy++", "down_weight"]
-    AVAILABLE_NORMALIZATIONS = ["none", "mean", "length", "length+mean"]
-except ImportError:
-    have_advanced_encode = False
-    AVAILABLE_STYLES = ["comfy"]
-    AVAILABLE_NORMALIZATIONS = ["none"]
-
-try:
-    from custom_nodes.Vector_Sculptor_ComfyUI.nodes import vector_sculptor_tokens
-
-    can_sculpt = True
-    log.info("Vector sculptor extension detected, can use SCULPT()")
-except ImportError:
-    can_sculpt = False
-
-
-AVAILABLE_STYLES.append("perp")
-log.info("Use STYLE(weight_interpretation, normalization) at the start of a prompt to use advanced encodings")
-log.info("Weight interpretations available: %s", ",".join(AVAILABLE_STYLES))
-log.info("Normalization types available: %s", ",".join(AVAILABLE_NORMALIZATIONS))
-
-
-def linear_interpolate_cond(
-    start, end, from_step=0.0, to_step=1.0, step=0.1, start_at=None, end_at=None, prompt_start="N/A", prompt_end="N/A"
-):
-    count = min(len(start), len(end))
-    if len(start) != len(end):
-        log.info(
-            "Length of conds to interpolate does not match (start=%s != end=%s), interpolating up to %s.",
-            len(start),
-            len(end),
-            count,
-        )
-
-    all_res = []
-    for idx in range(count):
-        res = []
-        from_cond, to_cond = equalize(start[idx][0], end[idx][0])
-        from_pooled = start[idx][1].get("pooled_output")
-        to_pooled = end[idx][1].get("pooled_output")
-        start_at = start_at if start_at is not None else from_step
-        end_at = end_at if end_at is not None else to_step
-        total_steps = int(round((to_step - from_step) / step, 0))
-        num_steps = int(round((end_at - from_step) / step, 0))
-        start_on = int(round((start_at - from_step) / step, 0))
-        start_pct = start_at
-        log.debug(
-            f"interpolate_cond {idx=} {from_step=} {to_step=} {start_at=} {end_at=} {total_steps=} {num_steps=} {start_on=} {step=}"
-        )
-        x = 1 / (total_steps + 1)
-        for s in range(start_on, num_steps):
-            factor = round((s + 1) * x, 2)
-            new_cond = from_cond + (to_cond - from_cond) * factor
-            if from_pooled is not None and to_pooled is not None:
-                from_pooled, to_pooled = equalize(from_pooled, to_pooled)
-                new_pooled = from_pooled + (to_pooled - from_pooled) * factor
-            elif from_pooled is not None:
-                new_pooled = from_pooled
-
-            n = [new_cond, start[idx][1].copy()]
-            if new_pooled is not None:
-                n[1]["pooled_output"] = new_pooled
-            n[1]["start_percent"] = round(start_pct, 2)
-            n[1]["end_percent"] = min(round((start_pct + step), 2), 1.0)
-            start_pct += step
-            start_pct = round(start_pct, 2)
-            if prompt_start:
-                n[1]["prompt"] = f"linear:{round(1.0 - factor, 2)} / {factor}"
-            log.debug(
-                "Interpolating at step %s with factor %s (%s, %s)...",
-                s,
-                factor,
-                n[1]["start_percent"],
-                n[1]["end_percent"],
-            )
-            res.append(n)
-        if res:
-            res[-1][1]["end_percent"] = round(end_at, 2)
-            all_res.extend(res)
-    return all_res
-
-
-def get_control_points(schedule, steps, encoder):
-    assert len(steps) > 1
-    new_steps = set(steps)
-
-    for step in (s[0] for s in schedule if s[0] >= steps[0] and s[0] <= steps[-1]):
-        new_steps.add(step)
-    control_points = [(s, encoder(schedule.at_step(s)[1])) for s in new_steps]
-    log.debug("Actual control points for interpolation: %s (from %s)", new_steps, steps)
-    return sorted(control_points, key=lambda x: x[0])
-
-
-def linear_interpolator(control_points, step, start_pct, end_pct):
-    o_start, start = control_points[0]
-    o_end, _ = control_points[-1]
-    t_start = o_start
-    conds = []
-    for t_end, end in control_points[1:]:
-        if t_start < start_pct:
-            t_start, start = t_end, end
-            continue
-        if t_start >= end_pct:
-            break
-        cs = linear_interpolate_cond(start, end, o_start, o_end, step, start_at=t_start, end_at=end_pct)
-        if cs:
-            conds.extend(cs)
-        else:
-            break
-        t_start = t_end
-        start = end
-    return conds
-
-
-class ScheduleToCond:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {"clip": ("CLIP",), "prompt_schedule": ("PROMPT_SCHEDULE",)},
-        }
-
-    RETURN_TYPES = ("CONDITIONING",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, clip, prompt_schedule):
-        with Timer("ScheduleToCond"):
-            r = (control_to_clip_common(clip, prompt_schedule),)
-        return r
-
-
-class EditableCLIPEncode:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "clip": ("CLIP",),
-                "text": ("STRING", {"multiline": True}),
-            },
-            "optional": {"filter_tags": ("STRING", {"default": ""})},
-        }
-
-    RETURN_TYPES = ("CONDITIONING",)
-    CATEGORY = "promptcontrol/old"
-    FUNCTION = "parse"
-
-    def parse(self, clip, text, filter_tags=""):
-        parsed = parse_prompt_schedules(text).with_filters(filter_tags)
-        return (control_to_clip_common(clip, parsed),)
-
-
-def get_sdxl(text, defaults):
-    # Defaults fail to parse and get looked up from the defaults dict
-    text, sdxl = get_function(text, "SDXL", ["none", "none", "none"])
-    if not sdxl:
-        return text, {}
-    args = sdxl[0]
-    d = defaults
-    w, h = parse_floats(args[0], [d.get("sdxl_width", 1024), d.get("sdxl_height", 1024)], split_re="\\s+")
-    tw, th = parse_floats(args[1], [d.get("sdxl_twidth", 1024), d.get("sdxl_theight", 1024)], split_re="\\s+")
-    cropw, croph = parse_floats(args[2], [d.get("sdxl_cwidth", 0), d.get("sdxl_cheight", 0)], split_re="\\s+")
-
-    opts = {
-        "width": int(w),
-        "height": int(h),
-        "target_width": int(tw),
-        "target_height": int(th),
-        "crop_w": int(cropw),
-        "crop_h": int(croph),
-    }
-    return text, opts
-
-
-def get_style(text, default_style="comfy", default_normalization="none"):
-    text, styles = get_function(text, "STYLE", [default_style, default_normalization])
-    if not styles:
-        return default_style, default_normalization, text
-    style, normalization = styles[0]
-    style = style.strip()
-    normalization = normalization.strip()
-    if style not in AVAILABLE_STYLES:
-        log.warning("Unrecognized prompt style: %s. Using %s", style, default_style)
-        style = default_style
-
-    if normalization not in AVAILABLE_NORMALIZATIONS:
-        log.warning("Unrecognized prompt normalization: %s. Using %s", normalization, default_normalization)
-        normalization = default_normalization
-
-    return style, normalization, text
-
-
-def encode_regions(clip, tokens, regions, weight_interpretation="comfy", token_normalization="none"):
-    from custom_nodes.ComfyUI_Cutoff.cutoff import CLIPSetRegion, finalize_clip_regions
-
-    clip_regions = {
-        "clip": clip,
-        "base_tokens": tokens,
-        "regions": [],
-        "targets": [],
-        "weights": [],
-    }
-
-    strict_mask = 1.0
-    start_from_masked = 1.0
-    mask_token = ""
-
-    for region in regions:
-        region_text, target_text, w, sm, sfm, mt = region
-        if w is not None:
-            w = safe_float(w, 0)
-        else:
-            w = 1.0
-        if sm is not None:
-            strict_mask = safe_float(sm, 1.0)
-        if sfm is not None:
-            start_from_masked = safe_float(sfm, 1.0)
-        if mt is not None:
-            mask_token = mt
-        log.info("Region: text %s, target %s, weight %s", region_text.strip(), target_text.strip(), w)
-        (clip_regions,) = CLIPSetRegion.add_clip_region(None, clip_regions, region_text, target_text, w)
-    log.info("Regions: mask_token=%s strict_mask=%s start_from_masked=%s", mask_token, strict_mask, start_from_masked)
-
-    (r,) = finalize_clip_regions(
-        clip_regions, mask_token, strict_mask, start_from_masked, token_normalization, weight_interpretation
-    )
-    cond, pooled = r[0][0], r[0][1].get("pooled_output")
-    return cond, pooled
-
-
-SHUFFLE_GEN = torch.Generator(device="cpu")
-
-
-def shuffle_chunk(shuffle, c):
-    func, shuffle = shuffle
-    shuffle_count = int(safe_float(shuffle[0], 0))
-    _, separator, joiner = shuffle
-    if separator == "default":
-        separator = ","
-
-    if not separator:
-        separator = ","
-
-    joiner = {
-        "default": ",",
-        "separator": separator,
-    }.get(joiner, joiner)
-
-    log.info("%s arg=%s sep=%s join=%s", func, shuffle_count, separator, joiner)
-    separated = c.split(separator)
-    if func == "SHIFT":
-        shuffle_count = shuffle_count % len(separated)
-        permutation = separated[shuffle_count:] + separated[:shuffle_count]
-    elif func == "SHUFFLE":
-        SHUFFLE_GEN.manual_seed(shuffle_count)
-        permutation = [separated[i] for i in torch.randperm(len(separated), generator=SHUFFLE_GEN)]
-    else:
-        # ??? should never get here
-        permutation = separated
-
-    permutation = [p for p in permutation if p.strip()]
-    if permutation != separated:
-        c = joiner.join(permutation)
-    return c
-
-
-def fix_word_ids(tokens):
-    """Fix word indexes. Tokenizing separately (when BREAKs exist) causes the indexes to restart which causes problems with some weighting algorithms that rely on them"""
-    for key in tokens:
-        max_idx = 0
-        for group in range(len(tokens[key])):
-            for i, token in enumerate(tokens[key][group]):
-                if len(token) < 3:
-                    # No need to fix ids when they don't exist
-                    return tokens
-                # Ignore zeros, they represent the padding token
-                if token[2] != 0 and token[2] < max_idx:
-                    tokens[key][group][i] = (token[0], token[1], token[2] + max_idx)
-            max_idx = max(max_idx, max(x for _, _, x in tokens[key][group]))
-    return tokens
-
-
-def encode_prompt(clip, text, default_style="comfy", default_normalization="none"):
-    style, normalization, text = get_style(text, default_style, default_normalization)
-    sculpts = []
-    if can_sculpt:
-        text, sculpts = get_function(text, "SCULPT", ["1.0", "forward", "none"])
-    text, regions = parse_cuts(text)
-    # defaults=None means there is no argument parsing at all
-    text, l_prompts = get_function(text, "CLIP_L", defaults=None)
-    chunks = re.split(r"\bBREAK\b", text)
-    token_chunks = []
-    need_word_ids = len(regions) > 0 or (have_advanced_encode and style != "perp")
-    for c in chunks:
-        c, shuffles = get_function(c.strip(), "(SHIFT|SHUFFLE)", ["0", "default", "default"], return_func_name=True)
-        r = c
-        for s in shuffles:
-            r = shuffle_chunk(s, r)
-        if r != c:
-            log.info("Shuffled prompt chunk to %s", r)
-            c = r
-        if sculpts:
-            w, method, norm = sculpts[0]
-            log.info("Using vector sculptor with method=%s norm=%s w=%s", method, norm, w)
-            w = safe_float(w, 1.0)
-            t = vector_sculptor_tokens(clip, c, method, norm, w)
-        else:
-            # Tokenizer returns padded results
-            t = clip.tokenize(c, return_word_ids=need_word_ids)
-        token_chunks.append(t)
-    tokens = token_chunks[0]
-
-    for key in tokens:
-        for c in token_chunks[1:]:
-            tokens[key].extend(c[key])
-
-    # Non-SDXL has only "l"
-    if "g" in tokens and l_prompts:
-        text_l = " ".join(l_prompts)
-        log.info("Encoded SDXL CLIP_L prompt: %s", text_l)
-        tokens["l"] = clip.tokenize(text_l, return_word_ids=need_word_ids)["l"]
-
-    if "g" in tokens and "l" in tokens and len(tokens["l"]) != len(tokens["g"]):
-        empty = clip.tokenize("", return_word_ids=need_word_ids)
-        while len(tokens["l"]) < len(tokens["g"]):
-            tokens["l"] += empty["l"]
-        while len(tokens["l"]) > len(tokens["g"]):
-            tokens["g"] += empty["g"]
-
-    tokens = fix_word_ids(tokens)
-
-    if len(regions) > 0:
-        return encode_regions(clip, tokens, regions, style, normalization)
-
-    if style == "perp":
-        if normalization != "none":
-            log.warning("Normalization is not supported with perp style weighting. Ignored '%s'", normalization)
-        return perp_encode(clip, tokens)
-
-    if "t5xxl" not in tokens and have_advanced_encode and not sculpts:
-        if "g" in tokens:
-            embs_l = None
-            embs_g = None
-            pooled = None
-            if "l" in tokens:
-                embs_l, _ = advanced_encode_from_tokens(
-                    tokens["l"],
-                    normalization,
-                    style,
-                    lambda x: encode_token_weights(clip, x, encode_token_weights_l),
-                    return_pooled=False,
-                )
-            if "g" in tokens:
-                embs_g, pooled = advanced_encode_from_tokens(
-                    tokens["g"],
-                    normalization,
-                    style,
-                    lambda x: encode_token_weights(clip, x, encode_token_weights_g),
-                    return_pooled=True,
-                    apply_to_pooled=False,
-                )
-            # Hardcoded clip_balance
-            return prepareXL(embs_l, embs_g, pooled, 0.5)
-        return advanced_encode_from_tokens(
-            tokens["l"],
-            normalization,
-            style,
-            lambda x: clip.encode_from_tokens({"l": x}, return_pooled=True),
-            return_pooled=True,
-            apply_to_pooled=True,
-        )
-    else:
-        return clip.encode_from_tokens(tokens, return_pooled=True)
-
-
-def get_area(text):
-    text, areas = get_function(text, "AREA", ["0 1", "0 1", "1"])
-    if not areas:
-        return text, None
-
-    args = areas[0]
-    x, w = parse_floats(args[0], [0.0, 1.0], split_re="\\s+")
-    y, h = parse_floats(args[1], [0.0, 1.0], split_re="\\s+")
-    weight = safe_float(args[2], 1.0)
-
-    def is_pct(f):
-        return f >= 0.0 and f <= 1.0
-
-    def is_pixel(f):
-        return f == 0 or f > 1
-
-    if all(is_pct(v) for v in [h, w, y, x]):
-        area = ("percentage", h, w, y, x)
-    elif all(is_pixel(v) for v in [h, w, y, x]):
-        area = (int(h) // 8, int(w) // 8, int(y) // 8, int(x) // 8)
-    else:
-        raise Exception(
-            f"AREA specified with invalid size {x} {w}, {h} {y}. They must either all be percentages between 0 and 1 or positive integer pixel values excluding 1"
-        )
-
-    return text, (area, weight)
-
-
-def get_mask_size(text, defaults):
-    text, sizes = get_function(text, "MASK_SIZE", ["512", "512"])
-    if not sizes:
-        return text, (defaults.get("mask_width", 512), defaults.get("mask_height", 512))
-    w, h = sizes[0]
-    return text, (int(w), int(h))
-
-
-def make_mask(args, size, weight):
-    x1, x2 = parse_floats(args[0], [0.0, 1.0], split_re="\\s+")
-    y1, y2 = parse_floats(args[1], [0.0, 1.0], split_re="\\s+")
-
-    def is_pct(f):
-        return f >= 0.0 and f <= 1.0
-
-    def is_pixel(f):
-        return f == 0 or f > 1
-
-    if all(is_pct(v) for v in [x1, x2, y1, y2]):
-        w, h = size
-        xs = int(w * x1), int(w * x2)
-        ys = int(h * y1), int(h * y2)
-    elif all(is_pixel(v) for v in [x1, x2, y1, y2]):
-        w, h = size
-        xs = int(x1), int(x2)
-        ys = int(y1), int(y2)
-    else:
-        raise Exception(
-            f"MASK specified with invalid size {x1} {x2}, {y1} {y2}. They must either all be percentages between 0 and 1 or positive integer pixel values excluding 1"
-        )
-
-    mask = torch.full((h, w), 0, dtype=torch.float32, device="cpu")
-    mask[ys[0] : ys[1], xs[0] : xs[1]] = weight
-    mask = mask.unsqueeze(0)
-    log.info("Mask xs=%s, ys=%s, shape=%s, weight=%s", xs, ys, mask.shape, weight)
-    return mask
-
-
-def get_mask(text, size, input_masks):
-    """Parse MASK(x1 x2, y1 y2, weight), IMASK(i, weight) and FEATHER(left top right bottom)"""
-    # TODO: combine multiple masks
-    text, masks = get_function(text, "MASK", ["0 1", "0 1", "1", "multiply"])
-    text, imasks = get_function(text, "IMASK", ["0", "1", "multiply"])
-    text, feathers = get_function(text, "FEATHER", ["0 0 0 0"])
-    text, maskw = get_function(text, "MASKW", ["1.0"])
-    if not masks and not imasks:
-        return text, None, None
-
-    def feather(f, mask):
-        l, t, r, b, *_ = [int(x) for x in parse_floats(f[0], [0, 0, 0, 0], split_re="\\s+")]
-        mask = FeatherMask().feather(mask, l, t, r, b)[0]
-        log.info("FeatherMask l=%s, t=%s, r=%s, b=%s", l, t, r, b)
-        return mask
-
-    mask = None
-    totalweight = 1.0
-    if maskw:
-        totalweight = safe_float(maskw[0][0], 1.0)
-    i = 0
-    for m in masks:
-        weight = safe_float(m[2], 1.0)
-        op = m[3]
-        nextmask = make_mask(m, size, weight)
-        if i < len(feathers):
-            nextmask = feather(feathers[i], nextmask)
-        i += 1
-        if mask is not None:
-            log.info("MaskComposite op=%s", op)
-            mask = MaskComposite().combine(mask, nextmask, 0, 0, op)[0]
-        else:
-            mask = nextmask
-
-    for idx, w, op in imasks:
-        idx = int(safe_float(idx, 0.0))
-        w = safe_float(w, 1.0)
-        if len(input_masks) < idx + 1:
-            log.warn("IMASK index %s not found, ignoring...", idx)
-            continue
-        nextmask = input_masks[idx] * w
-        if i < len(feathers):
-            nextmask = feather(feathers[i], nextmask)
-        i += 1
-        if mask is not None:
-            mask = MaskComposite().combine(mask, nextmask, 0, 0, op)[0]
-        else:
-            mask = nextmask
-
-    # apply leftover FEATHER() specs to the whole
-    for f in feathers[i:]:
-        mask = feather(f, mask)
-
-    return text, mask, totalweight
-
-
-def get_noise(text):
-    text, noises = get_function(
-        text,
-        "NOISE",
-        ["0.0", "none"],
-    )
-    if not noises:
-        return text, None, None
-    w = 0
-    # Only take seed from first noise spec, for simplicity
-    seed = safe_float(noises[0][1], "none")
-    if seed == "none":
-        gen = None
-    else:
-        gen = torch.Generator()
-        gen.manual_seed(int(seed))
-    for n in noises:
-        w += safe_float(n[0], 0.0)
-    return text, max(min(w, 1.0), 0.0), gen
-
-
-def apply_noise(cond, weight, gen):
-    if cond is None or not weight:
-        return cond
-
-    n = torch.randn(cond.size(), generator=gen).to(cond)
-
-    return cond * (1 - weight) + n * weight
-
-
-def do_encode(clip, text, defaults, masks):
-    # First style modifier applies to ANDed prompts too unless overridden
-    style, normalization, text = get_style(text)
-    text, mask_size = get_mask_size(text, defaults)
-
-    # Don't sum ANDs if this is in prompt
-    alt_method = "COMFYAND()" in text
-    text = text.replace("COMFYAND()", "")
-
-    prompts = [p.strip() for p in re.split(r"\bAND\b", text)]
-
-    p, sdxl_opts = get_sdxl(prompts[0], defaults)
-    prompts[0] = p
-
-    def weight(t):
-        opts = {}
-        m = re.search(r":(-?\d\.?\d*)(![A-Za-z]+)?$", t)
-        if not m:
-            return (1.0, opts, t)
-        w = float(m[1])
-        tag = m[2]
-        t = t[: m.span()[0]]
-        if tag == "!noscale":
-            opts["scale"] = 1
-
-        return w, opts, t
-
-    conds = []
-    res = []
-    scale = sum(abs(weight(p)[0]) for p in prompts if not ("AREA(" in p or "MASK(" in p))
-    for prompt in prompts:
-        prompt, mask, mask_weight = get_mask(prompt, mask_size, masks)
-        w, opts, prompt = weight(prompt)
-        text, noise_w, generator = get_noise(text)
-        if not w:
-            continue
-        prompt, area = get_area(prompt)
-        prompt, local_sdxl_opts = get_sdxl(prompt, defaults)
-        cond, pooled = encode_prompt(clip, prompt, style, normalization)
-        cond = apply_noise(cond, noise_w, generator)
-        pooled = apply_noise(pooled, noise_w, generator)
-
-        settings = {"prompt": prompt}
-        if alt_method:
-            settings["strength"] = w
-        settings.update(sdxl_opts)
-        settings.update(local_sdxl_opts)
-        if area:
-            settings["area"] = area[0]
-            settings["strength"] = area[1]
-            settings["set_area_to_bounds"] = False
-        if mask is not None:
-            settings["mask"] = mask
-            settings["mask_strength"] = mask_weight
-
-        if mask is not None or area or alt_method or local_sdxl_opts:
-            if pooled is not None:
-                settings["pooled_output"] = pooled
-            conds.append([cond, settings])
-        else:
-            s = opts.get("scale", scale)
-            res.append((cond, pooled, w / s))
-
-    sumconds = [r[0] * r[2] for r in res]
-    pooleds = [r[1] for r in res if r[1] is not None]
-
-    if len(res) > 0:
-        opts = sdxl_opts
-        if pooleds:
-            opts["pooled_output"] = sum(equalize(*pooleds))
-        sumcond = sum(equalize(*sumconds))
-        conds.append([sumcond, opts])
-    return conds
-
-
-def debug_conds(conds):
-    r = []
-    for i, c in enumerate(conds):
-        x = c[1].copy()
-        if "pooled_output" in x:
-            del x["pooled_output"]
-        r.append((i, x))
-    return r
-
-
-def control_to_clip_common(clip, schedules, lora_cache=None, cond_cache=None):
-    orig_clip = clip.clone()
-    current_loras = {}
-    if lora_cache is None:
-        lora_cache = {}
-    start_pct = 0.0
-    conds = []
-    cond_cache = cond_cache if cond_cache is not None else {}
-
-    def c_str(c):
-        r = [c["prompt"]]
-        loras = c["loras"]
-        for k in sorted(loras.keys()):
-            r.append(k)
-            r.append(loras[k]["weight_clip"])
-            for lbw, val in loras[k].get("lbw", {}).items():
-                r.append(lbw)
-                r.append(val)
-        return "".join(str(i) for i in r)
-
-    def encode(c):
-        nonlocal clip
-        nonlocal current_loras
-        prompt = c["prompt"]
-        loras = c["loras"]
-        cachekey = c_str(c)
-        cond = cond_cache.get(cachekey)
-        if cond is None:
-            if loras != current_loras:
-                _, clip = utils.apply_loras_from_spec(
-                    loras, clip=orig_clip, cache=lora_cache, applied_loras=current_loras
-                )
-                current_loras = loras
-            cond_cache[cachekey] = do_encode(clip, prompt, schedules.defaults, schedules.masks)
-        return cond_cache[cachekey]
-
-    for end_pct, c in schedules:
-        interpolations = [
-            i
-            for i in schedules.interpolations
-            if (start_pct >= i[0][0] and start_pct < i[0][-1]) or (end_pct > i[0][0] and start_pct < i[0][-1])
-        ]
-        new_start_pct = start_pct
-        if interpolations:
-            min_step = min(i[1] for i in interpolations)
-            for i in interpolations:
-                control_points, _ = i
-                interpolation_end_pct = min(control_points[-1], end_pct)
-                interpolation_start_pct = max(control_points[0], start_pct)
-
-                control_points = get_control_points(schedules, control_points, encode)
-                cs = linear_interpolator(control_points, min_step, interpolation_start_pct, interpolation_end_pct)
-                conds.extend(cs)
-                new_start_pct = max(new_start_pct, interpolation_end_pct)
-        start_pct = new_start_pct
-
-        if start_pct < end_pct:
-            cond = encode(c)
-            # Node functions return lists of cond
-            cond = conditioning_set_values(
-                cond, {"start_percent": round(start_pct, 2), "end_percent": round(end_pct, 2), "prompt": c["prompt"]}
-            )
-            conds.extend(cond)
-
-        start_pct = end_pct
-        log.debug("Conds at the end: %s", debug_conds(conds))
-
-    log.debug("Final cond info: %s", debug_conds(conds))
-    return conds

prompt_control/node_lora.py

@@ -1,247 +0,0 @@
-import logging
-import torch
-
-from .utils import unpatch_model, clone_model, set_callback, apply_loras_from_spec
-from .parser import parse_prompt_schedules
-from .hijack import do_hijack
-from comfy.samplers import CFGGuider
-
-log = logging.getLogger("comfyui-prompt-control")
-
-
-def apply_lora_for_step(schedules, step, total_steps, state, original_model, lora_cache, patch=True):
-    # zero-indexed steps, 0 = first step, but schedules are 1-indexed
-    sched = schedules.at_step(step + 1, total_steps)
-    lora_spec = sched[1]["loras"]
-
-    if state["applied_loras"] != lora_spec:
-        log.debug("At step %s, applying lora_spec %s", step, lora_spec)
-        m, _ = apply_loras_from_spec(
-            lora_spec,
-            model=state["model"],
-            orig_model=original_model,
-            cache=lora_cache,
-            patch=patch,
-            applied_loras=state["applied_loras"],
-        )
-        state["model"] = m
-        state["applied_loras"] = lora_spec
-
-
-def schedule_lora_common(model, schedules, lora_cache=None):
-    do_hijack()
-    orig_model = clone_model(model)
-    orig_model.model_options["pc_schedules"] = schedules
-
-    if lora_cache is None:
-        lora_cache = {}
-
-    def sampler_cb(orig_sampler, is_custom, *args, **kwargs):
-        split_sampling = args[0].model_options.get("pc_split_sampling")
-        state = {}
-        if is_custom:
-            steps = len(args[4])
-            log.info(
-                "SamplerCustom detected, number of steps not available. LoRA schedules will be calculated based on the number of sigmas (%s)",
-                steps,
-            )
-        else:
-            log.debug("Normal sampler detected, using steps from parameter")
-            steps = args[2]
-        start_step = kwargs.get("start_step") or 0
-        # The model patcher may change if LoRAs are applied
-        state["model"] = args[0]
-        state["applied_loras"] = {}
-
-        orig_cb = kwargs["callback"]
-
-        def step_callback(*args, **kwargs):
-            current_step = args[0] + start_step
-            apply_lora_for_step(schedules, current_step, steps, state, orig_model, lora_cache, patch=True)
-            if orig_cb:
-                return orig_cb(*args, **kwargs)
-
-        kwargs["callback"] = step_callback
-
-        apply_lora_for_step(schedules, start_step, steps, state, orig_model, lora_cache, patch=True)
-
-        def filter_conds(conds, t, start_t, end_t):
-            r = []
-            for c in conds:
-                x = c[1].copy()
-                start_at = round(x["start_percent"], 2)
-                end_at = round(x["end_percent"], 2)
-                # Take any cond that has any effect before end_t, since the percentages may not perfectly match
-                if end_t > start_at and end_t <= end_at:
-                    del x["start_percent"]
-                    del x["end_percent"]
-                    r.append([c[0].clone(), x])
-                else:
-                    log.debug("Rejecting cond (%s, %s) between (%s, %s)", start_at, end_at, start_t, end_t)
-            if len(r) == 0:
-                log.error("No %s conds between (%s, %s); Try adjusting your steps", t, start_t, end_t)
-            return r
-
-        def get_steps(conds):
-            for c in conds:
-                yield round(c[1].get("end_percent", 0), 2)
-
-        if split_sampling:
-            actual_end_step = kwargs["last_step"] or steps
-            first_step = True
-            s = args[8]
-            all_steps = sorted(set(int(steps * i) for i in [1.0] + list(get_steps(args[6])) + list(get_steps(args[7]))))
-            for end_step in all_steps:
-                if end_step <= start_step:
-                    continue
-                start_t = round(start_step / steps, 2)
-                end_t = round(end_step / steps, 2)
-                new_kwargs = kwargs.copy()
-                new_args = list(args)
-                new_args[0] = state["model"]
-                new_args[6] = filter_conds(new_args[6], "positive", start_t, end_t)
-                new_args[7] = filter_conds(new_args[7], "negative", start_t, end_t)
-                new_args[8] = s
-                log.info("Sampling from %s to %s (total: %s)", start_step, end_step, actual_end_step)
-                new_kwargs["start_step"] = start_step
-                new_kwargs["last_step"] = end_step
-                if end_step >= min(steps, actual_end_step):
-                    new_kwargs["force_full_denoise"] = kwargs["force_full_denoise"]
-                else:
-                    new_kwargs["force_full_denoise"] = False
-
-                if not first_step:
-                    # disable_noise apparently does nothing currently, we need to override noise in args
-                    new_kwargs["disable_noise"] = True
-                    new_args[1] = torch.zeros_like(s)
-
-                s = orig_sampler(*new_args, **new_kwargs)
-                start_step = end_step
-                first_step = False
-        else:
-            args = list(args)
-            args[0] = state["model"]
-            s = orig_sampler(*args, **kwargs)
-
-        unpatch_model(state["model"])
-
-        return s
-
-    set_callback(orig_model, sampler_cb)
-
-    return orig_model
-
-
-class PCWrapGuider:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "guider": ("GUIDER",),
-            },
-        }
-
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-    RETURN_TYPES = ("GUIDER",)
-
-    def apply(self, guider):
-        return (PCGuider(guider),)
-
-
-class PCGuider(CFGGuider):
-    def __init__(self, original_guider):
-        if "pc_schedules" not in original_guider.model_patcher.model_options:
-            raise ValueError(
-                "The guider passed to PCWrapGuider must contain a Model that has schedules applied. Use ScheduleToModel"
-            )
-        self.schedules = original_guider.model_patcher.model_options["pc_schedules"]
-        self.guider = original_guider
-        self.lora_cache = {}
-        # sets self.model_patcher
-        super().__init__(original_guider.model_patcher)
-
-    def sample(self, *args, **kwargs):
-        orig_cb = kwargs["callback"]
-        sigmas = args[3]
-        state = {"model": self.guider.model_patcher, "applied_loras": {}}
-
-        def step_callback(*args, **kwargs):
-            apply_lora_for_step(
-                self.schedules,
-                args[0],
-                len(sigmas),
-                state,
-                self.guider.model_patcher,
-                self.lora_cache,
-                patch=True,
-            )
-            if orig_cb:
-                return orig_cb(*args, **kwargs)
-
-        kwargs["callback"] = step_callback
-        apply_lora_for_step(
-            self.schedules, 0, len(sigmas), state, self.guider.model_patcher, self.lora_cache, patch=True
-        )
-        try:
-            r = self.guider.sample(*args, **kwargs)
-        finally:
-            unpatch_model(state["model"])
-        return r
-
-
-class ScheduleToModel:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "prompt_schedule": ("PROMPT_SCHEDULE",),
-            },
-        }
-
-    RETURN_TYPES = ("MODEL",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, model, prompt_schedule):
-        return (schedule_lora_common(model, prompt_schedule),)
-
-
-class PCSplitSampling:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "split_sampling": (["enable", "disable"],),
-            },
-        }
-
-    RETURN_TYPES = ("MODEL",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, model, split_sampling):
-        model = clone_model(model)
-        model.model_options["pc_split_sampling"] = split_sampling == "enable"
-        return (model,)
-
-
-class LoRAScheduler:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "text": ("STRING", {"multiline": True}),
-            },
-        }
-
-    RETURN_TYPES = ("MODEL",)
-    CATEGORY = "promptcontrol/old"
-    FUNCTION = "apply"
-
-    def apply(self, model, text):
-        schedules = parse_prompt_schedules(text)
-        return (schedule_lora_common(model, schedules),)

prompt_control/node_other.py

@@ -1,153 +0,0 @@
-import logging
-from .parser import parse_prompt_schedules
-
-log = logging.getLogger("comfyui-prompt-control")
-
-
-class FilterSchedule:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {"prompt_schedule": ("PROMPT_SCHEDULE",)},
-            "optional": {
-                "tags": ("STRING", {"default": ""}),
-                "start": ("FLOAT", {"min": 0.00, "max": 1.00, "default": 0.0, "step": 0.01}),
-                "end": ("FLOAT", {"min": 0.00, "max": 1.00, "default": 1.0, "step": 0.01}),
-            },
-        }
-
-    RETURN_TYPES = ("PROMPT_SCHEDULE",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, prompt_schedule, tags="", start=0.0, end=1.0):
-        p = prompt_schedule.with_filters(tags, start=start, end=end)
-        log.debug(
-            f"Filtered {prompt_schedule.parsed_prompt} with: ({tags}, {start}, {end}); the result is %s",
-            p.parsed_prompt,
-        )
-        return (p,)
-
-
-class PCApplySettings:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"prompt_schedule": ("PROMPT_SCHEDULE",), "settings": ("SCHEDULE_SETTINGS",)}}
-
-    RETURN_TYPES = ("PROMPT_SCHEDULE",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, prompt_schedule, settings):
-        return (prompt_schedule.with_filters(defaults=settings),)
-
-
-class PCScheduleAddMasks:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {"prompt_schedule": ("PROMPT_SCHEDULE",)},
-            "optional": {
-                "mask1": ("MASK",),
-                "mask2": ("MASK",),
-                "mask3": ("MASK",),
-                "mask4": ("MASK",),
-            },
-        }
-
-    RETURN_TYPES = ("PROMPT_SCHEDULE",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, prompt_schedule, mask1=None, mask2=None, mask3=None, mask4=None):
-        p = prompt_schedule.clone()
-        p.add_masks(mask1, mask2, mask3, mask4)
-        return (p,)
-
-
-class PCScheduleSettings:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {},
-            "optional": {
-                "steps": ("INT", {"default": 0, "min": 0, "max": 10000}),
-                "mask_width": ("INT", {"default": 512, "min": 64, "max": 4096 * 4}),
-                "mask_height": ("INT", {"default": 512, "min": 64, "max": 4096 * 4}),
-                "sdxl_width": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
-                "sdxl_height": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
-                "sdxl_target_w": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
-                "sdxl_target_h": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
-                "sdxl_crop_w": ("INT", {"default": 0, "min": 0, "max": 4096 * 4}),
-                "sdxl_crop_h": ("INT", {"default": 0, "min": 0, "max": 4096 * 4}),
-            },
-        }
-
-    RETURN_TYPES = ("SCHEDULE_SETTINGS",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(
-        self,
-        steps=0,
-        mask_width=512,
-        mask_height=512,
-        sdxl_width=1024,
-        sdxl_height=1024,
-        sdxl_target_w=1024,
-        sdxl_target_h=1024,
-        sdxl_crop_w=0,
-        sdxl_crop_h=0,
-    ):
-        settings = {
-            "steps": steps,
-            "mask_width": mask_width,
-            "mask_height": mask_height,
-            "sdxl_width": sdxl_width,
-            "sdxl_height": sdxl_height,
-            "sdxl_twidth": sdxl_target_w,
-            "sdxl_theight": sdxl_target_h,
-            "sdxl_cwidth": sdxl_crop_w,
-            "sdxl_cheight": sdxl_crop_h,
-        }
-        return (settings,)
-
-
-class PCPromptFromSchedule:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "prompt_schedule": ("PROMPT_SCHEDULE",),
-                "at": ("FLOAT", {"min": 0.0, "max": 1.0, "step": 0.01}),
-            },
-            "optional": {"tags": ("STRING", {"default": ""})},
-        }
-
-    RETURN_TYPES = ("STRING",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "apply"
-
-    def apply(self, prompt_schedule, at, tags=""):
-        p = prompt_schedule.with_filters(tags, start=at, end=at).parsed_prompt[-1][1]
-        log.info("Prompt at %s:\n%s", at, p["prompt"])
-        log.info("LoRAs: %s", p["loras"])
-        return (p["prompt"],)
-
-
-class PromptToSchedule:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "text": ("STRING", {"multiline": True}),
-            },
-        }
-
-    RETURN_TYPES = ("PROMPT_SCHEDULE",)
-    CATEGORY = "promptcontrol"
-    FUNCTION = "parse"
-
-    def parse(self, text, settings=None):
-        schedules = parse_prompt_schedules(text)
-        return (schedules,)

prompt_control/nodes_base.py

@@ -0,0 +1,30 @@
+import logging
+from .prompts import encode_prompt
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+class PCTextEncode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"clip": ("CLIP",), "text": ("STRING", {"multiline": True})},
+        }
+
+    RETURN_TYPES = ("CONDITIONING",)
+    CATEGORY = "promptcontrol"
+    FUNCTION = "apply"
+
+    def apply(self, clip, text):
+        defaults = clip.patcher.model_options.get("x-promptcontrol.defaults", {})
+        masks = clip.patcher.model_options.get("x-promptcontrol.masks", None)
+        return (encode_prompt(clip, text, 0, 1.0, defaults, masks),)
+
+
+NODE_CLASS_MAPPINGS = {
+    "PCTextEncode": PCTextEncode,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PCTextEncode": "PC Text Encode (no scheduling)",
+}

prompt_control/nodes_hooks.py

@@ -0,0 +1,88 @@
+import logging
+import comfy.utils
+import comfy.hooks
+import folder_paths
+from .utils import consolidate_schedule
+from .parser import parse_prompt_schedules
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+class PCLoraHooksFromText:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"text": ("STRING",)},
+        }
+
+    RETURN_TYPES = ("HOOKS",)
+    OUTPUT_TOOLTIPS = ("set of hooks created from the prompt schedule",)
+    CATEGORY = "promptcontrol/v2"
+    FUNCTION = "apply"
+    EXPERIMENTAL = True
+
+    def apply(self, text):
+        prompt_schedule = parse_prompt_schedules(text)
+        consolidated = consolidate_schedule(prompt_schedule)
+        hooks = lora_hooks_from_schedule(consolidated, {})
+        return (hooks,)
+
+
+def lora_hooks_from_schedule(schedules, non_scheduled):
+    start_pct = 0.0
+    lora_cache = {}
+    all_hooks = []
+
+    def create_hook(loraspec, start_pct, end_pct, non_scheduled):
+        nonlocal lora_cache
+        hooks = []
+        hook_kf = comfy.hooks.HookKeyframeGroup()
+        for path, info in loras.items():
+            if non_scheduled.get(path) == info:
+                log.info("Skipping %s from hook, it's loaded directly on model", path)
+                continue
+            if path not in lora_cache:
+                lora_cache[path] = comfy.utils.load_torch_file(
+                    folder_paths.get_full_path("loras", path), safe_load=True
+                )
+            new_hook = comfy.hooks.create_hook_lora(
+                lora_cache[path], strength_model=info["weight"], strength_clip=info["weight_clip"]
+            )
+            # Set hook_ref so that identical hooks compare equal
+            new_hook.hooks[0].hook_ref = f"pc-{path}-{info['weight']}-{info['weight_clip']}"
+            hooks.append(new_hook)
+        if start_pct > 0.0:
+            kf = comfy.hooks.HookKeyframe(strength=0.0, start_percent=0.0)
+            hook_kf.add(kf)
+        kf = comfy.hooks.HookKeyframe(strength=1.0, start_percent=start_pct)
+        hook_kf.add(kf)
+        if end_pct < 1.0:
+            kf = comfy.hooks.HookKeyframe(strength=0.0, start_percent=end_pct)
+            hook_kf.add(kf)
+        hooks = comfy.hooks.HookGroup.combine_all_hooks(hooks)
+        if hooks:
+            hooks.set_keyframes_on_hooks(hook_kf=hook_kf)
+        return hooks
+
+    for end_pct, loras in schedules:
+        log.info("Creating LoRA hook from %s to %s: %s", start_pct, end_pct, loras)
+        hook = create_hook(loras, start_pct, end_pct, non_scheduled)
+        all_hooks.append(hook)
+        start_pct = end_pct
+
+    del lora_cache
+
+    all_hooks = [x for x in all_hooks if x]
+
+    if all_hooks:
+        hooks = comfy.hooks.HookGroup.combine_all_hooks(all_hooks)
+        return hooks
+
+
+NODE_CLASS_MAPPINGS = {
+    "PCLoraHooksFromText": PCLoraHooksFromText,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PCLoraHooksFromText": "PC LoRA Hooks From Text (non-lazy)",
+}

prompt_control/nodes_lazy.py

@@ -0,0 +1,271 @@
+import logging
+from .parser import parse_prompt_schedules
+from comfy_execution.graph_utils import GraphBuilder
+
+from .prompts import get_function
+
+log = logging.getLogger("comfyui-prompt-control")
+
+from .utils import consolidate_schedule, find_nonscheduled_loras
+
+
+def create_lora_loader_nodes(graph, model, clip, loras):
+    for path, info in loras.items():
+        log.info("Creating LoraLoader for %s", path)
+        loader = graph.node("LoraLoader")
+        loader.set_input("model", model)
+        loader.set_input("clip", clip)
+        loader.set_input("strength_model", info["weight"])
+        loader.set_input("strength_clip", info["weight_clip"])
+        loader.set_input("lora_name", path)
+        model = loader.out(0)
+        clip = loader.out(1)
+    return model, clip
+
+
+def create_hook_nodes_for_lora(graph, path, info, existing_node, start_pct, end_pct):
+    next_keyframe = None
+    if not existing_node:
+        log.debug("Creating hook for %s", path)
+        hook_node = graph.node("CreateHookLora")
+        hook_node.set_input("lora_name", path)
+        hook_node.set_input("strength_model", info["weight"])
+        hook_node.set_input("strength_clip", info["weight_clip"])
+        prev_hook_kf = None
+        if start_pct > 0:
+            log.debug("Creating KF (0, %s) for %s", start_pct, path)
+            prev_keyframe = graph.node("CreateHookKeyframe")
+            prev_keyframe.set_input("strength_mult", 0.0)
+            prev_keyframe.set_input("start_percent", 0.0)
+            prev_hook_kf = prev_keyframe.out(0)
+    else:
+        log.debug("Hook already created for %s", path)
+        hook_node, prev_keyframe = existing_node
+        prev_hook_kf = prev_keyframe.out(0)
+
+    if (
+        prev_keyframe
+        and prev_keyframe.get_input("start_pct") == start_pct
+        and prev_keyframe.get_input("strength_mult") == 0.0
+    ):
+        next_keyframe = prev_keyframe
+        log.debug("Previous keyframe for %s starts at %s and has 0 strength, overriding", path, start_pct)
+    else:
+        log.debug("Creating keyframe for %s, start=%s ", path, start_pct)
+        next_keyframe = graph.node("CreateHookKeyframe")
+        next_keyframe.set_input("start_percent", start_pct)
+        next_keyframe.set_input("prev_hook_kf", prev_hook_kf)
+
+    next_keyframe.set_input("strength_mult", 1.0)
+    prev_hook_kf = next_keyframe.out(0)
+    if end_pct < 1.0:
+        log.debug("Creating end keyframe for %s, start=%s", path, start_pct)
+        next_keyframe = graph.node("CreateHookKeyframe")
+        next_keyframe.set_input("strength_mult", 0.0)
+        next_keyframe.set_input("start_percent", end_pct)
+        next_keyframe.set_input("prev_hook_kf", prev_hook_kf)
+    return hook_node, next_keyframe
+
+
+def build_lora_schedule(graph, schedule, model, clip, apply_hooks=True, return_hooks=True):
+    # This gets rid of non-existent LoRAs
+    consolidated = consolidate_schedule(schedule)
+    non_scheduled = find_nonscheduled_loras(consolidated)
+    model, clip = create_lora_loader_nodes(graph, model, clip, non_scheduled)
+
+    hook_nodes = {}
+    start_pct = 0.0
+
+    def key(lora, info):
+        return f"{lora}-{info['weight']}-{info['weight_clip']}"
+
+    for end_pct, loras in consolidated:
+        for lora, info in loras.items():
+            if non_scheduled.get(lora) == info:
+                continue
+            k = key(lora, info)
+            existing_node = hook_nodes.get(k)
+            hook_nodes[k] = create_hook_nodes_for_lora(graph, lora, info, existing_node, start_pct, end_pct)
+        start_pct = end_pct
+
+    hooks = []
+    # Attach the keyframe chain to the hook node
+    for hook, kfs in hook_nodes.values():
+        n = graph.node("SetHookKeyframes")
+        n.set_input("hooks", hook.out(0))
+        n.set_input("hook_kf", kfs.out(0))
+        hooks.append(n)
+
+    res = None
+    # Finally, combine all hooks and optionally apply
+    if len(hooks) > 0:
+        res = hooks[0]
+        for h in hooks[:1]:
+            n = graph.node("CombineHooks2")
+            n.set_input("hooks_A", res.out(0))
+            n.set_input("hooks_B", h.out(0))
+            res = n
+        res = res.out(0)
+    if apply_hooks:
+        n = graph.node("SetClipHooks")
+        n.set_input("clip", clip)
+        n.set_input("hooks", res)
+        n.set_input("apply_to_conds", True)
+        n.set_input("schedule_clip", True)
+        clip = n.out(0)
+
+    r = graph.finalize()
+
+    if return_hooks:
+        ret = (model, clip, res)
+    else:
+        ret = (model, clip)
+
+    return {"result": ret, "expand": r}
+
+
+class PCLazyLoraLoaderAdvanced:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "text": ("STRING", {"multiline": True}),
+                "model": ("MODEL", {"rawLink": True}),
+                "clip": ("CLIP", {"rawLink": True}),
+                "apply_hooks": ("BOOLEAN", {"default": True}),
+            },
+            "optional": {
+                "tags": ("STRING", {"default": ""}),
+                "start": ("FLOAT", {"min": 0.0, "max": 1.0, "default": 0.0, "step": 0.01}),
+                "end": ("FLOAT", {"min": 0.0, "max": 1.0, "default": 1.0, "step": 0.01}),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("MODEL", "CLIP", "HOOKS")
+    OUTPUT_TOOLTIPS = ("Returns a model and clip with LoRAs scheduled",)
+    CATEGORY = "promptcontrol"
+    FUNCTION = "apply"
+
+    def apply(self, model, clip, text, apply_hooks, unique_id, tags="", start=0.0, end=1.0):
+        schedule = parse_prompt_schedules(text).with_filters(filters=tags, start=start, end=end)
+        graph = GraphBuilder(f"PCLazyLoraLoaderAdvanced-{unique_id}")
+        return build_lora_schedule(graph, schedule, model, clip, apply_hooks=apply_hooks, return_hooks=True)
+
+
+class PCLazyLoraLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "model": ("MODEL", {"rawLink": True}),
+                "clip": ("CLIP", {"rawLink": True}),
+                "text": ("STRING", {"multiline": True}),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = (
+        "MODEL",
+        "CLIP",
+    )
+    OUTPUT_TOOLTIPS = ("Returns a model and clip with LoRAs scheduled",)
+    CATEGORY = "promptcontrol"
+    FUNCTION = "apply"
+
+    def apply(self, model, clip, text, unique_id):
+        graph = GraphBuilder(f"PCLazyLoraLoader-{unique_id}")
+        schedule = parse_prompt_schedules(text)
+        return build_lora_schedule(graph, schedule, model, clip, apply_hooks=True, return_hooks=False)
+
+
+def build_scheduled_prompts(graph, schedules, clip):
+    nodes = []
+    start_pct = 0.0
+    prompt_cache = {}
+    for end_pct, c in schedules:
+        p = c["prompt"]
+        p, classnames = get_function(p, "NODE", ["PCTextEncode", "text"])
+        classname = "PCTextEncode"
+        paramname = "text"
+        if classnames:
+            classname = classnames[0][0]
+            paramname = classnames[0][1]
+        node = prompt_cache.get((p, classname, paramname))
+        if not node:
+            node = graph.node(classname)
+            node.set_input("clip", clip)
+            node.set_input(paramname, p)
+            prompt_cache[(p, classname, paramname)] = node
+        timestep = graph.node("ConditioningSetTimestepRange")
+        timestep.set_input("conditioning", node.out(0))
+        timestep.set_input("start", start_pct)
+        timestep.set_input("end", end_pct)
+        nodes.append(timestep)
+        start_pct = end_pct
+    node = nodes[0]
+    for othernode in nodes[1:]:
+        combiner = graph.node("ConditioningCombine")
+        combiner.set_input("conditioning_1", node.out(0))
+        combiner.set_input("conditioning_2", othernode.out(0))
+        node = combiner
+
+    return {"result": (node.out(0),), "expand": graph.finalize()}
+
+
+class PCLazyTextEncode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"clip": ("CLIP", {"rawLink": True}), "text": ("STRING", {"multiline": True})},
+            # "optional": {"defaults": ("SCHEDULE_DEFAULTS",)},
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("CONDITIONING",)
+    OUTPUT_TOOLTIPS = ("A fully encoded and scheduled conditioning",)
+    CATEGORY = "promptcontrol"
+    FUNCTION = "apply"
+
+    def apply(self, clip, text, unique_id):
+        schedules = parse_prompt_schedules(text)
+        graph = GraphBuilder(f"PCEncodeLazy-{unique_id}")
+        return build_scheduled_prompts(graph, schedules, clip)
+
+
+class PCLazyTextEncodeAdvanced:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"clip": ("CLIP", {"rawLink": True}), "text": ("STRING", {"multiline": True})},
+            "optional": {
+                "tags": ("STRING", {"default": ""}),
+                "start": ("FLOAT", {"min": 0.0, "max": 1.0, "default": 0.0, "step": 0.01}),
+                "end": ("FLOAT", {"min": 0.0, "max": 1.0, "default": 1.0, "step": 0.01}),
+            },
+            "hidden": {"unique_id": "UNIQUE_ID"},
+        }
+
+    RETURN_TYPES = ("CONDITIONING",)
+    CATEGORY = "promptcontrol"
+    FUNCTION = "apply"
+
+    def apply(self, clip, text, unique_id, tags="", start=0.1, end=1.0):
+        schedules = parse_prompt_schedules(text).with_filters(start=start, end=end, filters=tags)
+        graph = GraphBuilder(f"PCLazyTextEncodeAdvanced-{unique_id}")
+        return build_scheduled_prompts(graph, schedules, clip)
+
+
+NODE_CLASS_MAPPINGS = {
+    "PCLazyTextEncode": PCLazyTextEncode,
+    "PCLazyTextEncodeAdvanced": PCLazyTextEncodeAdvanced,
+    "PCLazyLoraLoader": PCLazyLoraLoader,
+    "PCLazyLoraLoaderAdvanced": PCLazyLoraLoaderAdvanced,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PCLazyTextEncode": "PC: Schedule Prompt",
+    "PCLazyTextEncodeAdvanced": "PC: Schedule prompt (Advanced)",
+    "PCLazyLoraLoader": "PC: Schedule LoRAs",
+    "PCLazyLoraLoaderAdvanced": "PC: Schedule LoRAs (Advanced)",
+}

prompt_control/nodes_tools.py

@@ -0,0 +1,110 @@
+import logging
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+class PCAddMaskToCLIP:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"clip": ("CLIP",)},
+            "optional": {
+                "mask": ("MASK",),
+            },
+        }
+
+    RETURN_TYPES = ("CLIP",)
+    CATEGORY = "promptcontrol/v2"
+    FUNCTION = "apply"
+
+    def apply(self, clip, mask=None):
+        return PCAddMaskToCLIPMany().apply(clip, mask1=mask)
+
+
+class PCAddMaskToCLIPMany:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"clip": ("CLIP",)},
+            "optional": {
+                "mask1": ("MASK",),
+                "mask2": ("MASK",),
+                "mask3": ("MASK",),
+                "mask4": ("MASK",),
+            },
+        }
+
+    RETURN_TYPES = ("CLIP",)
+    CATEGORY = "promptcontrol/v2"
+    FUNCTION = "apply"
+
+    def apply(self, clip, mask1=None, mask2=None, mask3=None, mask4=None):
+        clip = clip.clone()
+        current_masks = clip.patcher.model_options.get("x-promptcontrol.masks", [])
+        current_masks.extend(m for m in (mask1, mask2, mask3, mask4) if m is not None)
+        clip.patcher.model_options["x-promptcontrol.masks"] = current_masks
+        return (clip,)
+
+
+class PCSetPCTextEncodeSettings:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {"clip": ("CLIP",)},
+            "optional": {
+                "steps": ("INT", {"default": 0, "min": 0, "max": 10000}),
+                "mask_width": ("INT", {"default": 512, "min": 64, "max": 4096 * 4}),
+                "mask_height": ("INT", {"default": 512, "min": 64, "max": 4096 * 4}),
+                "sdxl_width": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
+                "sdxl_height": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
+                "sdxl_target_w": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
+                "sdxl_target_h": ("INT", {"default": 1024, "min": 0, "max": 4096 * 4}),
+                "sdxl_crop_w": ("INT", {"default": 0, "min": 0, "max": 4096 * 4}),
+                "sdxl_crop_h": ("INT", {"default": 0, "min": 0, "max": 4096 * 4}),
+            },
+        }
+
+    RETURN_TYPES = ("CLIP",)
+    CATEGORY = "promptcontrol/v2"
+    FUNCTION = "apply"
+
+    def apply(
+        self,
+        clip,
+        steps=0,
+        mask_width=512,
+        mask_height=512,
+        sdxl_width=1024,
+        sdxl_height=1024,
+        sdxl_target_w=1024,
+        sdxl_target_h=1024,
+        sdxl_crop_w=0,
+        sdxl_crop_h=0,
+    ):
+        settings = {
+            "steps": steps,
+            "mask_width": mask_width,
+            "mask_height": mask_height,
+            "sdxl_width": sdxl_width,
+            "sdxl_height": sdxl_height,
+            "sdxl_twidth": sdxl_target_w,
+            "sdxl_theight": sdxl_target_h,
+            "sdxl_cwidth": sdxl_crop_w,
+            "sdxl_cheight": sdxl_crop_h,
+        }
+        clip = clip.clone()
+        clip.patcher.model_options["x-promptcontrol.settings"] = settings
+        return (clip,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "PCSetPCTextEncodeSettings": PCSetPCTextEncodeSettings,
+    "PCAddMaskToCLIP": PCAddMaskToCLIP,
+    "PCAddMaskToCLIPMany": PCAddMaskToCLIPMany,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PCSetTextEncodeSettings": "PC: Configure PCTextEncode",
+    "PCAddMaskToCLIP": "PC: Attach Mask",
+    "PCAddMaskToCLIPMany": "PC: Attach Mask (multi)",
+}

prompt_control/parser.py

@@ -5,19 +5,22 @@ from math import ceil
 logging.basicConfig()
 log = logging.getLogger("comfyui-prompt-control")
 
+if lark.__version__ == "0.12.0":
+    x = "Your lark package reports an ancient version (0.12.0) and will not work. If you have the 'lark-parser' package in your Python environment, remove that and *reinstall* lark!"
+    log.error(x)
+    raise ImportError(x)
+
+
 prompt_parser = lark.Lark(
     r"""
 !start: (prompt | /[][():|]/+)*
-prompt: (emphasized | embedding | scheduled | alternate | sequence | interpolate | loraspec | PLAIN | /</ | />/ | WHITESPACE)+
+prompt: (emphasized | embedding | scheduled | alternate | sequence | loraspec | PLAIN | /</ | />/ | WHITESPACE)+
 !emphasized: "(" prompt? ")"
         | "(" prompt ":" prompt ")"
         | "[" prompt "]"
 scheduled: "[" [prompt ":"] [prompt] ":" _WS? NUMBER ["," NUMBER] "]"
         | "[" [prompt ":"] [prompt] ":" _WS? TAG "]"
 sequence:  "[SEQ" ":" [prompt] ":" NUMBER (":" [prompt] ":" NUMBER)+ "]"
-interpolate.100: "[INT" ":" interp_prompts ":" interp_steps "]"
-interp_prompts: prompt (":" [prompt])+
-interp_steps: NUMBER ("," NUMBER)+ [":" NUMBER]
 alternate: "[" [prompt] ("|" [prompt])+ [":" NUMBER] "]"
 loraspec.99: "<lora:" FILENAME lora_weights [lora_block_weights] ">"
 lora_weights.1: (":" _WS? NUMBER)~1..2
@@ -88,7 +91,6 @@ def clamp(a, b, c):
 
 def get_steps(tree):
     res = [100]
-    interpolation_steps = []
 
     def tostep(s):
         w = float(s) * 100
@@ -110,7 +112,6 @@ def get_steps(tree):
             for i, _ in enumerate(tree.children[:-1]):
                 tree.children[i] = tostep(tree.children[i])
 
-            interpolation_steps.append((tuple(tree.children[:-1]), tree.children[-1]))
             res.extend(tree.children[:-1])
 
         def sequence(self, tree):
@@ -128,7 +129,7 @@ def get_steps(tree):
 
     CollectSteps().visit(tree)
 
-    return sorted(set(interpolation_steps)), sorted(set(res))
+    return sorted(set(res))
 
 
 def at_step(step, filters, tree):
@@ -174,24 +175,6 @@ def at_step(step, filters, tree):
                     previous_step = s
             return ""
 
-        def interpolate(self, args):
-            prompts, starts = args
-            starts = starts[:-1]
-            prev_prompt = None
-            if step < starts[0]:
-                return prompts[0]
-            for i, x in enumerate(starts):
-                prev_prompt = prompts[i]
-                if x >= step:
-                    break
-            return prev_prompt
-
-        def interp_steps(self, args):
-            return list(args)
-
-        def interp_prompts(self, args):
-            return ["".join(flatten(a or [])) for a in args]
-
         def alternate(self, args):
             step_size = args[-1]
             idx = ceil(step / step_size)
@@ -262,22 +245,15 @@ def at_step(step, filters, tree):
 
 
 class PromptSchedule(object):
-    def __init__(self, prompt, filters="", start=0.0, end=1.0, defaults=None, masks=None):
+    def __init__(self, prompt, filters="", start=0.0, end=1.0):
         self.filters = filters
         self.start = start
         self.end = end
         self.prompt = prompt.strip()
         self.defaults = {}
-        if defaults:
-            self.defaults = defaults
         self.loaded_loras = {}
 
-        self.interpolations = None
-        self.parsed_prompt = None
-        self.interpolations, self.parsed_prompt = self._parse()
-        self.masks = masks
-        if masks is None:
-            self.masks = []
+        self.parsed_prompt = self._parse()
 
     def __iter__(self):
         # Filter out zero, it's only useful for interpolation
@@ -287,27 +263,14 @@ class PromptSchedule(object):
         filters = [x.strip() for x in self.filters.upper().split(",")]
         try:
             parsed = []
-            interpolations = set()
             tree = prompt_parser.parse(self.prompt)
-            interpolation_steps, steps = get_steps(tree)
-            log.debug("Interpolation steps: %s", interpolation_steps)
+            steps = get_steps(tree)
 
             def f(x):
                 return round(x / 100, 2)
 
             for t in steps:
                 p = at_step(t, filters, tree)
-                for control_points, step in interpolation_steps:
-                    interp_start = None
-                    interp_end = None
-                    if t == control_points[-1]:
-                        interp_start = max(control_points[0], int(self.start * 100))
-                        interp_end = min(control_points[-1], int(self.end * 100))
-                        control_points = tuple(
-                            sorted(set(f(c) for c in control_points if c >= interp_start or c <= interp_end))
-                        )
-                    if interp_start is not None and interp_end is not None and interp_end > interp_start:
-                        interpolations.add((control_points, f(step)))
                 parsed.append([f(t), p])
 
         except lark.exceptions.LarkError as e:
@@ -316,14 +279,9 @@ class PromptSchedule(object):
 
         # Tag filtering may return redundant prompts, so filter them out here
         res = []
-        prev_p = None
         prev_end = -1
 
         for end_at, p in parsed:
-            # Preserve prompt if it ends at the start of an interpolation, otherwise bump its end time
-            if p == prev_p and res[-1][0] not in [x[0][0] for x in interpolations]:
-                res[-1][0] = end_at
-                continue
             if end_at < self.start:
                 continue
             elif end_at <= self.end:
@@ -332,18 +290,12 @@ class PromptSchedule(object):
             elif end_at > self.end and prev_end < self.end:
                 res.append([end_at, p])
                 break
-            prev_p = p
 
         # Always use the last prompt if everything was filtered
         if len(res) == 0:
             res = [[1.0, parsed[-1][1]]]
 
-        return interpolations, res
-
-    def add_masks(self, *masks):
-        for mask in masks:
-            if mask is not None:
-                self.masks.append(mask)
+        return res
 
     def clone(self):
         return self.with_filters()
@@ -357,8 +309,6 @@ class PromptSchedule(object):
             filters=ifspecified(filters, self.filters),
             start=ifspecified(start, self.start),
             end=ifspecified(end, self.end),
-            defaults=ifspecified(defaults, self.defaults),
-            masks=self.masks[:],
         )
         return p
 
@@ -372,23 +322,6 @@ class PromptSchedule(object):
                 return i, x
         return len(self.parsed_prompt) - 1, self.parsed_prompt[-1]
 
-    def interpolation_at(self, step, total_steps=1):
-        i, x = self.at_step_idx(step, total_steps)
-        for y in self.parsed_prompt[i:]:
-            step = min(y[0], 1.0)
-            if x[1]["prompt"] != y[1]["prompt"]:
-                return step, y
-        return 1.0, self.parsed_prompt[-1]
-
-    def load_loras(self, lora_cache=None):
-        from .utils import Timer, load_loras_from_schedule
-
-        if lora_cache is not None:
-            self.loaded_loras = lora_cache
-        with Timer("PromptSchedule.load_loras()"):
-            self.loaded_loras = load_loras_from_schedule(self.parsed_prompt, self.loaded_loras)
-        return self.loaded_loras
-
 
 def parse_prompt_schedules(prompt):
     return PromptSchedule(prompt)

prompt_control/perp_weight.py

@@ -1,70 +0,0 @@
-import torch
-
-
-# Copied and adapted from https://github.com/bvhari/ComfyUI_PerpWeight/blob/main/clipperpweight.py
-def perp_encode(clip, tokens):
-    empty_tokens = clip.tokenize("")
-    sdxl_flag = "g" in tokens
-    empty_cond, empty_cond_pooled = clip.encode_from_tokens(empty_tokens, return_pooled=True)
-    unweighted_tokens = {}
-    for k in ["l", "g"]:
-        if k not in tokens:
-            continue
-        unweighted_tokens[k] = [[(t, 1.0) for t, _ in x] for x in tokens[k]]
-    unweighted_cond, unweighted_pooled = clip.encode_from_tokens(unweighted_tokens, return_pooled=True)
-    cond = torch.clone(unweighted_cond)
-
-    if sdxl_flag:
-        for i in range(unweighted_cond.shape[0]):
-            for j in range(unweighted_cond.shape[1]):
-                weight_l = tokens["l"][(j // 77)][(j % 77)][1]
-                if weight_l != 1.0:
-                    token_vector_l = unweighted_cond[i][j][:768]
-                    zero_vector_l = empty_cond[0][(j % 77)][:768]
-                    perp_l = (
-                        (torch.mul(zero_vector_l, token_vector_l).sum()) / (torch.norm(token_vector_l) ** 2)
-                    ) * token_vector_l
-                    if weight_l > 1.0:
-                        cond[i][j][:768] = token_vector_l + (weight_l * perp_l)
-                    elif (weight_l > 0.0) and (weight_l < 1.0):
-                        cond[i][j][:768] = token_vector_l - ((1 - weight_l) * perp_l)
-                    elif weight_l < 0.0:
-                        cond[i][j][:768] = token_vector_l + (weight_l * perp_l)
-                    elif weight_l == 0.0:
-                        cond[i][j][:768] = empty_cond[0][(j % 77)][:768]
-
-                weight_g = tokens["g"][(j // 77)][(j % 77)][1]
-                if weight_g != 1.0:
-                    token_vector_g = unweighted_cond[i][j][768:]
-                    zero_vector_g = empty_cond[0][(j % 77)][768:]
-                    perp_g = (
-                        (torch.mul(zero_vector_g, token_vector_g).sum()) / (torch.norm(token_vector_g) ** 2)
-                    ) * token_vector_g
-                    if weight_g > 1.0:
-                        cond[i][j][768:] = token_vector_g + (weight_g * perp_g)
-                    elif (weight_g > 0.0) and (weight_g < 1.0):
-                        cond[i][j][768:] = token_vector_g - ((1 - weight_g) * perp_g)
-                    elif weight_g < 0.0:
-                        cond[i][j][768:] = token_vector_g + (weight_g * perp_g)
-                    elif weight_g == 0.0:
-                        cond[i][j][768:] = empty_cond[0][(j % 77)][768:]
-    else:
-        tokens = tokens["l"]
-        for i in range(unweighted_cond.shape[0]):
-            for j in range(unweighted_cond.shape[1]):
-                weight = tokens[(j // 77)][(j % 77)][1]
-                if weight != 1.0:
-                    token_vector = unweighted_cond[i][j]
-                    zero_vector = empty_cond[0][(j % 77)]
-                    perp = (
-                        (torch.mul(zero_vector, token_vector).sum()) / (torch.norm(token_vector) ** 2)
-                    ) * token_vector
-                    if weight > 1.0:
-                        cond[i][j] = token_vector + (weight * perp)
-                    elif (weight > 0.0) and (weight < 1.0):
-                        cond[i][j] = token_vector - ((1 - weight) * perp)
-                    elif weight < 0.0:
-                        cond[i][j] = token_vector + (weight * perp)
-                    elif weight == 0.0:
-                        cond[i][j] = empty_cond[0][(j % 77)]
-    return cond, unweighted_pooled

prompt_control/prompts.py

@@ -1,58 +1,18 @@
 import logging
 import re
 import torch
-from .utils import safe_float, get_function, parse_floats, lora_name_to_file
+from functools import partial
 from comfy_extras.nodes_mask import FeatherMask, MaskComposite
-import comfy.utils
-import comfy.hooks
-import folder_paths
+
+from .utils import safe_float, get_function, parse_floats
+from .adv_encode import advanced_encode_from_tokens
+from .cutoff import process_cuts
+from .parser import parse_cuts
 
 log = logging.getLogger("comfyui-prompt-control")
 
-AVAILABLE_STYLES = ["comfy"]
-AVAILABLE_NORMALIZATIONS = ["none"]
-
-have_advanced_encode = False
-try:
-    import custom_nodes.ComfyUI_ADV_CLIP_emb.adv_encode as adv_encode
-
-    have_advanced_encode = True
-    AVAILABLE_STYLES.extend(["A1111", "compel", "comfy++", "down_weight"])
-    AVAILABLE_NORMALIZATIONS.extend(["mean", "length", "length+mean"])
-except ImportError:
-    pass
-
-
-class PCLoraHooksFromSchedule:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {"prompt_schedule": ("PROMPT_SCHEDULE",)},
-        }
-
-    RETURN_TYPES = ("HOOKS",)
-    OUTPUT_TOOLTIPS = ("set of hooks created from the prompt schedule",)
-    CATEGORY = "promptcontrol/_unstable"
-    FUNCTION = "apply"
-
-    def apply(self, prompt_schedule):
-        return (lora_hooks_from_schedule(prompt_schedule),)
-
-
-class PCEncodeSchedule:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {"clip": ("CLIP",), "prompt_schedule": ("PROMPT_SCHEDULE",)},
-        }
-
-    RETURN_TYPES = ("CONDITIONING",)
-    CATEGORY = "promptcontrol/_unstable"
-    FUNCTION = "apply"
-
-    def apply(self, clip, prompt_schedule):
-        return (encode_schedule(clip, prompt_schedule),)
-
+AVAILABLE_STYLES = ["comfy", "perp", "A1111", "compel", "comfy++", "down_weight"]
+AVAILABLE_NORMALIZATIONS = ["none", "mean", "length", "length+mean"]
 
 SHUFFLE_GEN = torch.Generator(device="cpu")
 
@@ -79,6 +39,22 @@ def get_sdxl(text, defaults):
     return text, opts
 
 
+def get_clipweights(text, existing_spec=None):
+    text, spec = get_function(text, "TE_WEIGHT", defaults=None)
+    if not spec:
+        return existing_spec or {}, text
+    args = spec[0].strip()
+    res = {}
+    for arg in args.split(","):
+        try:
+            te, val = arg.strip().split("=")
+            te, val = te.strip(), float(val.strip())
+            res[te] = val
+        except ValueError:
+            log.warning("Invalid TE weight spec '%s', ignoring...", arg.strip())
+    return res, text
+
+
 def get_style(text, default_style="comfy", default_normalization="none"):
     text, styles = get_function(text, "STYLE", [default_style, default_normalization])
     if not styles:
@@ -146,15 +122,28 @@ def fix_word_ids(tokens):
     return tokens
 
 
-def encode_prompt(
-    clip, text, settings, default_style="comfy", default_normalization="none"
+def encode_prompt_segment(
+    clip,
+    text,
+    settings,
+    default_style="comfy",
+    default_normalization="none",
+    clip_weights=None,
 ) -> list[tuple[torch.Tensor, dict[str]]]:
     style, normalization, text = get_style(text, default_style, default_normalization)
+    clip_weights, text = get_clipweights(text, clip_weights)
+    text, cuts = parse_cuts(text)
+    extra = {}
+    if clip_weights:
+        extra["clip_weights"] = clip_weights
+    if cuts:
+        extra["cuts"] = cuts
+
     # defaults=None means there is no argument parsing at all
     text, l_prompts = get_function(text, "CLIP_L", defaults=None)
     chunks = re.split(r"\bBREAK\b", text)
     token_chunks = []
-    need_word_ids = have_advanced_encode or style == "comfy" and normalization == "none"
+    need_word_ids = True
     for c in chunks:
         c, shuffles = get_function(c.strip(), "(SHIFT|SHUFFLE)", ["0", "default", "default"], return_func_name=True)
         r = c
@@ -186,31 +175,79 @@ def encode_prompt(
 
     tokens = fix_word_ids(tokens)
 
-    newclip = clip
-    if have_advanced_encode:
-        newclip = clip.clone()
-        if hasattr(clip.patcher.model, "clip_g"):
-            newclip.patcher.add_object_patch(
-                "clip_g.encode_token_weights",
-                encoder_patch(style, normalization, clip.patcher.get_model_object("clip_g.encode_token_weights")),
-            )
-        if hasattr(clip.patcher.model, "clip_l"):
-            newclip.patcher.add_object_patch(
-                "clip_l.encode_token_weights",
-                encoder_patch(style, normalization, clip.patcher.get_model_object("clip_l.encode_token_weights")),
-            )
+    tes = []
+    for k in tokens:
+        if k in ["g", "l"]:
+            tes.append(f"clip_{k}")
+        else:
+            tes.append(k)
 
-    return newclip.encode_from_tokens_scheduled(tokens, add_dict=settings)
+    clip = hook_te(clip, tes, style, normalization, extra)
+
+    return clip.encode_from_tokens_scheduled(tokens, add_dict=settings)
 
 
-def encoder_patch(style, normalization, orig_fn):
-    if not have_advanced_encode or style == "comfy" and normalization == "none":
-        return orig_fn
+def apply_weights(output, te_name, spec):
+    """Applies weights to TE outputs"""
+    if not spec:
+        return output
+
+    if te_name.startswith("clip_"):
+        te_name = te_name[5:]
+
+    if isinstance(output, tuple):
+        out, pooled = output
+        if te_name in spec:
+            log.info("Weighting %s output by %s", te_name, spec[te_name])
+            out = out * spec[te_name]
+        pkey = te_name + "_pooled"
+        if pkey in spec:
+            log.info("Weighting %s pooled output by %s", te_name, spec[pkey])
+            pooled = pooled * spec[pkey]
+
+        return out, pooled
     else:
-        log.debug("Encoding with style=%s, normalization=%s", style, normalization)
-        return lambda t: adv_encode.advanced_encode_from_tokens(
-            t, normalization, style, orig_fn, return_pooled=True, apply_to_pooled=False
+        if te_name in spec:
+            log.info("Weighting %s output by %s", te_name, spec[te_name])
+            output = output * spec[te_name]
+        return output
+
+
+def make_patch(te_name, orig_fn, normalization, style, extra):
+    def encode(t):
+        r = advanced_encode_from_tokens(
+            t, normalization, style, orig_fn, return_pooled=True, apply_to_pooled=False, **extra
         )
+        return apply_weights(r, te_name, extra.get("clip_weights"))
+
+    if "cuts" in extra:
+        return partial(process_cuts, encode, extra)
+    return encode
+
+
+def hook_te(clip, te_names, style, normalization, extra):
+    if style == "comfy" and normalization == "none" and not extra:
+        return clip
+    newclip = clip.clone()
+    for te_name in te_names:
+        if hasattr(clip.patcher.model, te_name):
+            x = extra.copy()
+            x["tokenizer"] = getattr(clip.tokenizer, te_name)
+            log.debug("Hooked into %s with style=%s, normalization=%s", te_name, style, normalization)
+            newclip.patcher.add_object_patch(
+                f"{te_name}.encode_token_weights",
+                make_patch(
+                    te_name,
+                    clip.patcher.get_model_object(f"{te_name}.encode_token_weights"),
+                    normalization,
+                    style,
+                    x,
+                ),
+            )
+        # 'g' and 'l' exist in these are clip_g and clip_l
+        else:
+            log.debug("Tokens contain items with key %s but no TE found on object with that name.", te_name)
+    return newclip
 
 
 def get_area(text):
@@ -365,7 +402,7 @@ def apply_noise(cond, weight, gen):
     return cond * (1 - weight) + n * weight
 
 
-def do_encode(clip, text, start_pct, end_pct, defaults, masks):
+def encode_prompt(clip, text, start_pct, end_pct, defaults, masks):
     # First style modifier applies to ANDed prompts too unless overridden
     style, normalization, text = get_style(text)
     text, mask_size = get_mask_size(text, defaults)
@@ -389,7 +426,8 @@ def do_encode(clip, text, start_pct, end_pct, defaults, masks):
         return w, opts, t
 
     conds = []
-    scale = sum(abs(weight(p)[0]) for p in prompts if not ("AREA(" in p or "MASK(" in p))
+    # TODO: is this still needed?
+    # scale = sum(abs(weight(p)[0]) for p in prompts if not ("AREA(" in p or "MASK(" in p))
     for prompt in prompts:
         prompt, mask, mask_weight = get_mask(prompt, mask_size, masks)
         w, opts, prompt = weight(prompt)
@@ -412,97 +450,7 @@ def do_encode(clip, text, start_pct, end_pct, defaults, masks):
 
         settings["start_percent"] = start_pct
         settings["end_percent"] = end_pct
-        x = encode_prompt(clip, prompt, settings, style, normalization)
+        x = encode_prompt_segment(clip, prompt, settings, style, normalization)
         conds.extend(x)
 
     return conds
-
-
-def debug_conds(conds):
-    r = []
-    for i, c in enumerate(conds):
-        x = c[1].copy()
-        if "pooled_output" in x:
-            del x["pooled_output"]
-        r.append((i, x))
-    return r
-
-
-def lora_hooks_from_schedule(schedules):
-    start_pct = 0.0
-
-    lora_cache = {}
-
-    all_hooks = []
-
-    prev_loras = {}
-
-    def create_hook(loraspec, start_pct, end_pct):
-        nonlocal lora_cache
-        hooks = []
-        hook_kf = comfy.hooks.HookKeyframeGroup()
-        for lora, info in loras.items():
-            path = lora_name_to_file(lora)
-            if not path:
-                continue
-            if path not in lora_cache:
-                lora_cache[path] = comfy.utils.load_torch_file(
-                    folder_paths.get_full_path("loras", path), safe_load=True
-                )
-            new_hook = comfy.hooks.create_hook_lora(
-                lora_cache[path], strength_model=info["weight"], strength_clip=info["weight_clip"]
-            )
-            # Set hook_ref so that identical hooks compare equal
-            new_hook.hooks[0].hook_ref = f"pc-{path}-{info['weight']}-{info['weight_clip']}"
-            hooks.append(new_hook)
-        if start_pct > 0.0:
-            kf = comfy.hooks.HookKeyframe(strength=0.0, start_percent=0.0)
-            hook_kf.add(kf)
-        kf = comfy.hooks.HookKeyframe(strength=1.0, start_percent=start_pct)
-        hook_kf.add(kf)
-        if end_pct < 1.0:
-            kf = comfy.hooks.HookKeyframe(strength=0.0, start_percent=end_pct)
-            hook_kf.add(kf)
-        hooks = comfy.hooks.HookGroup.combine_all_hooks(hooks)
-        if hooks:
-            hooks.set_keyframes_on_hooks(hook_kf=hook_kf)
-        return hooks
-
-    consolidated = []
-
-    prev_loras = {}
-    for end_pct, c in reversed(list(schedules)):
-        loras = c["loras"]
-        if loras != prev_loras:
-            consolidated.append((end_pct, loras))
-        prev_loras = loras
-    consolidated = reversed(consolidated)
-
-    for end_pct, loras in consolidated:
-        log.info("Creating LoRA hook from %s to %s: %s", start_pct, end_pct, loras)
-        hook = create_hook(loras, start_pct, end_pct)
-        all_hooks.append(hook)
-        start_pct = end_pct
-
-    del lora_cache
-
-    all_hooks = [x for x in all_hooks if x is not None]
-
-    if all_hooks:
-        hooks = comfy.hooks.HookGroup.combine_all_hooks(all_hooks)
-        return hooks
-
-
-def encode_schedule(clip, schedules):
-    start_pct = 0.0
-    conds = []
-    for end_pct, c in schedules:
-        if start_pct < end_pct:
-            prompt = c["prompt"]
-            cond = do_encode(clip, prompt, start_pct, end_pct, schedules.defaults, schedules.masks)
-            conds.extend(cond)
-        start_pct = end_pct
-        log.debug("Conds at the end: %s", debug_conds(conds))
-
-    log.debug("Final cond info: %s", debug_conds(conds))
-    return conds

prompt_control/utils.py

@@ -1,60 +1,51 @@
-from collections import namedtuple
-from os import environ
 from pathlib import Path
 import re
-from math import lcm
-import time
 import logging
-import torch
 
-
-import nodes
 import folder_paths
 
-import comfy.model_management
-
 log = logging.getLogger("comfyui-prompt-control")
 
-FORCE_CPU_OFFLOAD = bool(environ.get("COMFYUI_PC_CPU_OFFLOAD"))
+
+def consolidate_schedule(prompt_schedule):
+    prev_loras = {}
+    not_found = []
+    consolidated = []
+    for end_pct, c in reversed(list(prompt_schedule)):
+        loras = {}
+        for k, v in c["loras"].items():
+            if k in not_found:
+                continue
+            path = lora_name_to_file(k)
+            if path is None:
+                not_found.append(k)
+                continue
+            loras[path] = v
+
+        if loras != prev_loras:
+            consolidated.append((end_pct, loras))
+        prev_loras = loras
+    for k in not_found:
+        log.warning("LoRA '%s' not found, ignoring...", k)
+    return list(reversed(consolidated))
 
 
-# Minimal Modelpatcher that doesn't do anything, for LoRA loading when not
-# interested in either CLIP or unet
-class DummyModelPatcher:
-    class DummyTorchModel:
-        def __init__(self):
-            dummyconf = {
-                "num_res_blocks": [],
-                "channel_mult": [],
-                "transformer_depth": [],
-                "transformer_depth_output": [],
-                "transformer_depth_middle": 0,
-            }
-            self.model_config = namedtuple("DummyConfig", ["unet_config"])(dummyconf)
-
-        def state_dict(self):
-            return {}
-
-    def __init__(self):
-        self.model = self.DummyTorchModel()
-        self.cond_stage_model = self.DummyTorchModel()
-        self.weight_inplace_update = True
-        self.model_options = {}
-
-    def add_patches(self, patches, *args, **kwargs):
-        return []
-
-    def patch_model(self):
-        pass
-
-    def unpatch_model(self):
-        pass
-
-    def clone(self):
-        return self
-
-
-DUMMY_MODEL = DummyModelPatcher()
+def find_nonscheduled_loras(consolidated_schedule):
+    consolidated_schedule = list(consolidated_schedule)
+    if not consolidated_schedule:
+        return {}
+    last_end, candidate_loras = consolidated_schedule[0]
+    print(candidate_loras)
+    to_remove = set()
+    for candidate, weights in candidate_loras.items():
+        for end, loras in consolidated_schedule[1:]:
+            last_end = end
+            if loras.get(candidate) != weights:
+                to_remove.add(candidate)
+    # No candidates if the schedule does not span full time
+    if last_end < 1.0:
+        return {}
+    return {k: v for (k, v) in candidate_loras.items() if k not in to_remove}
 
 
 def find_closing_paren(text, start):
@@ -118,15 +109,6 @@ def parse_strings(string, defaults, split_re=r"(?<!\\),", replace=(r"\,", ",")):
     return parse_args(splits, spec, strip=False)
 
 
-def equalize(*tensors):
-    if all(t.shape[1] == tensors[0].shape[1] for t in tensors):
-        return tensors
-
-    x = lcm(*(t.shape[1] for t in tensors))
-
-    return (t.repeat(1, x // t.shape[1], 1) for t in tensors)
-
-
 def safe_float(f, default):
     if f is None:
         return default
@@ -136,89 +118,6 @@ def safe_float(f, default):
         return default
 
 
-def unpatch_model(model):
-    if model:
-        log.info("Unpatching model")
-        model.unpatch_model()
-
-
-def clone_model(model):
-    if not model:
-        return None
-    model = model.clone()
-    if not environ.get("PC_NO_INPLACE_UPDATE"):
-        model.weight_inplace_update = True
-    return model
-
-
-def add_patches(model, patches, weight):
-    model.add_patches(patches, weight)
-
-
-def patch_model(model, forget=False, orig=None):
-    global FORCE_CPU_OFFLOAD
-    try:
-        return _patch_model(model, forget, orig, FORCE_CPU_OFFLOAD)
-    except comfy.model_management.OOM_EXCEPTION:
-        FORCE_CPU_OFFLOAD = True
-        log.error("Ran out of memory while applying LoRAs, Forcing CPU offload from now on")
-        # Unpatch to restore partially applied weights
-        unpatch_model(model)
-        raise
-
-
-def _patch_model(model, forget=False, orig=None, offload_to_cpu=False):
-    if not model:
-        return None
-    if offload_to_cpu:
-        saved_offload = model.offload_device
-        model.offload_device = torch.device("cpu")
-    log.info(
-        "Patching model, model.load_device=%s model.model.device=%s cpu_offload=%s",
-        model.load_device,
-        model.model.device,
-        model.offload_device == torch.device("cpu"),
-    )
-    if orig:
-        model.backup = orig.backup
-    model.patch_model(device_to=model.load_device)
-    if offload_to_cpu:
-        model.offload_device = saved_offload
-    if forget:
-        model.patches = {}
-        model.object_patches = {}
-    return model
-
-
-def get_callback(model):
-    return model.model_options.get("prompt_control_callback")
-
-
-def set_callback(model, cb):
-    model.model_options["prompt_control_callback"] = cb
-
-
-# Hack to temporarily override printing to stdout to stop log spam
-def suppress_print(f):
-    def noop(*args):
-        pass
-
-    p = print
-    __builtins__["print"] = noop
-    rootlogger = logging.getLogger()
-    oldlevel = rootlogger.level
-    try:
-        rootlogger.setLevel(logging.ERROR)
-        x = f()
-    except BaseException:
-        __builtins__["print"] = p
-        rootlogger.setLevel(oldlevel)
-        raise
-    __builtins__["print"] = p
-    rootlogger.setLevel(oldlevel)
-    return x
-
-
 def lora_name_to_file(name):
     filenames = folder_paths.get_filename_list("loras")
     # Return exact matches as is
@@ -231,121 +130,3 @@ def lora_name_to_file(name):
             if p.name == n or str(p) == n:
                 return f
     return None
-
-
-def load_lbw():
-    return nodes.NODE_CLASS_MAPPINGS.get("LoraLoaderBlockWeight //Inspire")
-
-
-def make_loader(filename, lbw):
-    if not lbw:
-        l = nodes.LoraLoader()
-
-        def loader(model, clip, model_weight, clip_weight, lbw):
-            return suppress_print(lambda: l.load_lora(model, clip, filename, model_weight, clip_weight))
-
-    else:
-        # This is already checked before calling make_loader
-        l = load_lbw()()
-
-        def loader(model, clip, model_weight, clip_weight, lbw):
-            spec = lbw["LBW"]
-            lbw_a = safe_float(lbw.get("A"), 4.0)
-            lbw_b = safe_float(lbw.get("B"), 1.0)
-            m = model or DUMMY_MODEL
-            c = clip or DUMMY_MODEL
-            m, c, _ = suppress_print(
-                lambda: l.doit(m, c, filename, model_weight, clip_weight, False, 0, lbw_a, lbw_b, "", spec)
-            )
-            if m is DUMMY_MODEL:
-                m = None
-            if c is DUMMY_MODEL:
-                c = None
-            return m, c
-
-    return loader
-
-
-def apply_loras_from_spec(
-    loraspec, model=None, clip=None, orig_model=None, orig_clip=None, patch=False, cache=None, applied_loras=None
-):
-    if applied_loras is None:
-        applied_loras = {}
-    actual_loraspec = {}
-    additive = True
-    for key in loraspec:
-        if key in applied_loras and applied_loras[key] == loraspec[key]:
-            continue
-        if key in applied_loras and applied_loras[key] != loraspec[key]:
-            additive = False
-        actual_loraspec[key] = loraspec[key]
-
-    for key in applied_loras:
-        if key not in loraspec:
-            actual_loraspec = loraspec
-            additive = False
-
-    backup_model = model
-    if not additive:
-        unpatch_model(model)
-        # Reset clip to unpatched
-        if clip:
-            clip = orig_clip or clip
-
-    if cache is None:
-        cache = {}
-    if not loraspec:
-        return model, clip
-
-    for name, params in actual_loraspec.items():
-        m, c = model, clip
-        w, w_clip = params["weight"], params["weight_clip"]
-        if w == 0:
-            m = None
-        if w_clip == 0:
-            c = None
-        if not w and not c:
-            continue
-
-        lbw = params.get("lbw")
-        if lbw and not load_lbw():
-            log.warning("LoraBlockWeight not available, ignoring LBW parameters")
-            lbw = None
-
-        # Cache the loader instance so that it doesn't reload the LoRA from disk all the time
-        cache_key = name, bool(lbw)
-        loader = cache.get(cache_key)
-        if not loader:
-            f = lora_name_to_file(name)
-            if not f:
-                log.warning("Lora %s not found", name)
-                continue
-            log.info("Loading LoRA: %s", f)
-            loader = make_loader(f, bool(lbw))
-            cache[cache_key] = loader
-
-        m, c = loader(m, c, w, w_clip, lbw)
-        model = m or model
-        clip = c or clip
-        if model:
-            log.info("Applying LoRA: %s:%s, LBW=%s, additive=%s", name, params["weight"], bool(lbw), additive)
-        if clip:
-            log.info("Applying CLIP LoRA: %s:%s, LBW=%s, additive=%s", name, params["weight_clip"], bool(lbw), additive)
-
-    # forget patches so we don't double-patch
-    model = patch_model(model, forget=True, orig=backup_model)
-    return model, clip
-
-
-class Timer:
-    def __init__(self, name):
-        self.name = name
-        self.start = None
-
-    def __enter__(self):
-        self.start = time.time()
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        elapsed = time.time() - self.start
-        if environ.get("PC_SHOW_TIMINGS"):
-            log.info("Executed %s in %s seconds", self.name, elapsed)

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "comfyui-prompt-control"
 description = "Nodes for convenient prompt editing, making many common operations prompt-controllable"
-version = "1.2.0"
+version = "2.0.0-beta.2"
 license = { file = "LICENSE" }
 # some lark versions older than 1.1.9 apparently have a bug that breaks things, see https://github.com/asagi4/comfyui-prompt-control/issues/35
 dependencies = ["lark >= 1.1.9"]