v2.0.0-rc.6

I can't figure out why from_masked works differently from down_weight
which also used that batching function but could be replaced with a simple
torch.cat

.github/workflows/tests_with_comfy.yml

@@ -23,5 +23,6 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: '3.11'
-      - run: pip install -r requirements.txt
+      - run: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+      - run: pip install -r requirements.txt -r ComfyUI/requirements.txt
       - run: PYTHONPATH=ComfyUI python -m prompt_control.test_graph

Makefile

@@ -11,4 +11,10 @@ test:
 test_graph:
 	PYTHONPATH=../../ python -m prompt_control.test_graph
 
+test_encode:
+	PYTHONPATH=../../ python -m prompt_control.test_encode
+
+manual_test:
+	PYTHONPATH=../../ python -im prompt_control.manual_test
+
 .PHONY: check format all

README.md

@@ -2,48 +2,51 @@
 
 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.
 
+Prompt Control comes with `PCTextEncode`, which provides advanced text encoding with many additional features compared to ComfyUI's base `CLIPTextEncode`.
+
 A `Basic Text to Image` template is included with the extension, and can be loaded from ComfyUI's template library.
 
+## What can it do?
+
+You can use text prompts to control the following:
+
+- A1111-style prompt scheduling and filtering without noodle soup.
+- LoRA loading and scheduling via ComfyUI's hook system
+- Masking, composition and area control (regional prompting) with an implementation of [Attention Couple](doc/attention_couple.md), also fully schedulable.
+- Per-encoder prompts for models with multiple text encoders, such as SDXL and Flux
+- Prompt combinators like `BREAK`, as well as `CAT`, `AVG()` and `AND` corresponding to ComfyUI's `ConditioningConcat`, `ConditioningAverage` and `ConditioningCombine` nodes.
+- Different weight interpretation types (ComfyUI, A1111, compel, etc.)
+- Prompt masking with an implementation of [cutoff](https://github.com/BlenderNeko/ComfyUI_Cutoff)
+- Simple prompt macros with `DEF`
+- And a bunch more
+
+All features are fully schedulable unless otherwise stated. See the [syntax documentation](doc/syntax.md) for details on how to use each feature.
+
+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](example_workflows/Workflow%20Comparison.json?raw=1) shows LoRA scheduling and prompt editing and compares it with the same prompt implemented with built-in ComfyUI nodes. You can also find it in the template library.
+
+## Compatibility
+
+Prompt Control uses graph generation,  and tries to delegate functionality to core ComfyUI wherever possible, implementing any hooks and patches in a way that is maximally compatible. This means that it should just work in most cases, even with models and nodes not explicitly supported.
+
+If you encounter issues as a user or if you're a node developer and Prompt Control somehow breaks something, feel free to file a bug report.
+
 ## 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. 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 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.
 
-
 ### 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?
-
-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
-
-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
 
 For LoRA scheduling to work, you'll need at least version 0.3.7 of ComfyUI (0.3.36 of ComfyUI desktop).
@@ -93,75 +96,6 @@ This node attaches masks to a `CLIP` model so that they can be referred to when
 
 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.
-
-## Advanced CLIP encoding
-
-If you use `PCTextEncode`, advanced encodings are available automatically. Thanks to BlenderNeko for the original code.
-
-Use the syntax `STYLE(weight_interpretation, normalization)` in a prompt to affect how prompts are interpreted.
-
-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:
-
-```
-STYLE(A1111) a (red:1.1) cat with (brown:0.9) spots and a long tail AND an (old:0.5) dog AND a (green:1.4) (balloon:1.1)
-```
-will interpret everything as A1111, but
-```
-a (red:1.1) cat with (brown:0.9) spots and a long tail AND STYLE(A1111) an (old:0.5) dog AND a (green:1.4) (balloon:1.1)
-```
-Will interpret the first one using the default ComfyUI behaviour, the second prompt with A1111 and the last prompt with the default again
-
-For things (ie. the code imports) to work, the nodes must be cloned in a directory named exactly `ComfyUI_ADV_CLIP_emb`.
-
-## Cutoff
-
-NOTE: Cutoff syntax might change at some point; it's pretty clunky.
-
-`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:_]
-```
-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.
-
-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 `\`.
-
-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=+`
-
-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:
-
-`[CUT:white cat:white:0.5] and [CUT:black parrot, flying:black:1.0:0.5] and [CUT:green apple:green]`
-
-`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.
-
-The parameters affect how the masked and unmasked prompts are combined to produce the final embedding. Just play around with them.
-
 # Known issues
 
 - ComfyUI's caching mechanism has an issue that makes it unnecessarily invalidate caches for certain inputs; you'll still get some benefit from the lazy nodes, but changing inputs that shouldn't affect downstream nodes (especially if using filtering) will still cause them to be recomputed because ComfyUI doesn't realize the inputs haven't changed.
@@ -169,3 +103,5 @@ The parameters affect how the masked and unmasked prompts are combined to produc
 If you want to enable a hack to fix this, set `PROMPTCONTROL_ENABLE_CACHE_HACK=1` in your environment. Unset it to disable.
 
 It's a purely optional performance optimization that allows Prompt Control nodes to override their cache keys in a way that should not interfere with other nodes. Note that the optimization only works if the text input to the lazy nodes is a constant (so either directly on the node or from a primitive); outputs from other nodes can't be optimized.
+
+- Cutoff does not work with models that use non-CLIP text encoders, like Flux. This might be fixable, but it's uncertain if cutoff even makes sense for those models.

__init__.py

@@ -29,22 +29,9 @@ cache_hack.init()
 NODE_CLASS_MAPPINGS = {}
 NODE_DISPLAY_NAME_MAPPINGS = {}
 
-nodes = ["base", "lazy", "tools"]
-optional_nodes = ["attnmask"]
-if importlib.util.find_spec("comfy.hooks"):
-    nodes.extend(["hooks"])
-else:
-    log.error("Your ComfyUI version is too old, can't import comfy.hooks. Update your installation.")
+nodes = ["base", "lazy", "tools", "hooks"]
 
 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)
-
-for node in optional_nodes:
-    try:
-        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)
-    except ImportError:
-        log.info(f"Could not import optional nodes: {node}; continuing anyway")

doc/attention_couple.md

@@ -0,0 +1,39 @@
+# Attention Couple
+
+NOTE: This is still considered an experimental feature, so the syntax may change.
+
+Attention Couple is an attention-based implementation of regional prompting. it is faster and often more flexible than latent-based masking.
+
+The implementation is based on the one by [pamparamm](https://github.com/pamparamm/ComfyUI-ppm.git), modified to use ComfyUI's hook system. This enables it to work with prompt scheduling.
+
+By default, the implementation produces slightly different results from Pamparamm's implementation because ComfyUI will only run the hook for conds that have it attached and can't batch negative conditionings.
+
+As a consequence of this, however, you can also use `ATTN()` in your negative prompt, and it will work correctly.
+
+To enable batching negative prompts, run your positive and negative prompt through the `PPCAttentionCoupleBatchNegative` node. This will make the outputs identical to pamparamm's implementation and will also improve performance. It will fall back to the default behaviour in cases where batching can't be done, so it should always be safe to use.
+
+
+## Syntax
+
+See also the main syntax documentation for `MASK` etc.
+
+### ATTN: Trigger Attention Couple
+
+Use `ATTN()` to mark a prompt to be used with Attention Couple. `ATTN()` needs to be combined with either `MASK()` or `IMASK()` to work correctly.
+
+If no mask is specified, an implicit `MASK()` is assumed.
+
+For attention masking to take effect, you need at least two prompt segments with the `ATTN()` marker (separated with `AND`). A single prompt with `ATTN()` will simply ignore the marker.
+
+For the first prompt (and the first prompt only) you can also use `FILL()` to automatically mask all parts not masked by other prompt segments.
+
+For example:
+```
+dog FILL() ATTN() AND cat MASK(0.5 1) ATTN()
+```
+
+If typing `ATTN() MASK()` feels bothersome, try the following macro:
+```
+DEF(AM=ATTN() MASK($1))
+```
+and then use it like `MASK`: `AM(0 1, 0.5 1)`

doc/syntax.md

@@ -1,6 +1,8 @@
-# Scheduling syntax
+# Prompt Control Syntax
 
-Syntax is like A1111 for now, but only fractions are supported for steps. LoRAs are scheduled by including them in a scheduling expression.
+If you're viewing this on GitHub, I recommend opening the outline by clicking the button in the top right corner of the text view (it is annoyingly easy to miss).
+
+Scheduling syntax is similar to A1111, 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]
@@ -12,7 +14,7 @@ a [large::0.1] [cat|dog:0.05] [<lora:somelora:0.5:0.6>::0.5]
 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.
+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. Either prompt can also be empty.
 For example:
 ```
 a [red:blue:0.5] cat
@@ -24,22 +26,25 @@ 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 equivalent to `[:cat:0.5]` meaning it switches from empty to `cat` at 0.5.
+For convenience `[cat:0.5]` is equivalent to `[:cat:0.5]` meaning it switches from empty to `cat` at 0.5.
+
+### Range expressions
+
+The most general form of a schedule is a range expression: For example, in `[before:during:after:0.3,0.7]`, The prompt be `a before` until 0.3, `a during` until 0.7, and then `a after`. This form is equivalent to `[before:[during:after:0.7]:0.3]`
+
+For convenience, `[during:0.1,0.4]` is equivalent to `[:during::0.1,0.4]` and `[during:after:0.1,0.4]` is equivalent to `[:during:after:0.1,0.4]`.
+
+`[before:during:after:0.1]` is the same as `[before:during:after:0.1,1.0]` which is same as `[before:during:0.1]`
+
 
 ### Using step numbers with the Advanced nodes
 
-If you provide a non-zero value to `num_steps` to the `Advanced` versions of the lazy LoRA loaders, you will be able to use step numbers in prompts.
+If you provide a non-zero value to `num_steps` to the `Advanced` versions of the scheduling nodes, you will be able to use step numbers in prompts.
 
 For now, a value between 0 and 1.0 will be interpreted as a percentage if it contains a ., and as an absolute step otherwise.
 
 This is just syntactic sugar. Behind the scenes, the values are converted to percentages and have normal ComfyUI scheduling behaviour.
 
-### 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:
 ```
@@ -55,6 +60,7 @@ 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`
 
+The three prompt form `[a:b:c:TAG]` is parsed, but ignores `b` and is equivalent to `[a:c:TAG]`.
 
 ## LoRA Scheduling
 When using the lazy graph building nodes, LoRAs can be scheduled by referring to them in a scheduling expression, like so:
@@ -89,48 +95,129 @@ generates a LoRA schedule based on a sinewave
 
 # Basic prompt syntax
 
-This syntax is also available in outside scheduled prompts, where applicable.
+This syntax is also available in outside scheduled with the `PCTextEncode` node, where applicable.
 
-## LoRA loading
+## Combining prompts
 
-The A111-style syntax `<lora:loraname:weight>` can be used to load LoRAs via the prompt. See LoRA scheduling above.
+### AND
 
-## Combining prompts, A1111-style
+`AND` can be used to create "prompt segments". By default, it works as if you had combined the different prompts with `ConditioningCombine`. 
 
-- 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.
+It is also used with regional prompting to separate different prompts; see `MASK` and `ATTN` below.
 
-`AND` can be used to combine prompts. You can also use a weight at the end. It does a weighted sum of each prompt,
+Prompts can have a weight at the end:
 ```
 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`
+`AND` is processed after schedule parsing, so you can change the weight mid-prompt: `cat:[1:2:0.5] AND dog`
+
+The weight defaults to 1. If a prompt's weight is set to 0, it's **skipped entirely.** This can be useful when scheduling to completely disable a prompt:
+
+```
+cat [\:0::0.5] AND dog
+```
+Note that the `:` needs to be escaped with a `\` or it will be interpreted as scheduling syntax.
+
+## Note about processing order
+
+Prompt operators are processed in the following order, meaning that all features "below" another can be affected by the feature above it. That is, `BREAK` can go inside a `TE()` call, but not `AND` or `CAT`.
+
+- DEF macros are expanded
+- Scheduling is expanded
+- Prompts are split by AND
+- Most functions (like STYLE, MASK) and cutoffs are evaluated
+- prompts are split by AVG()
+- prompts are split by CAT
+- the TE() function is evaluated to set per-encoder prompts
+- BREAK is evaluated
+- Everything else
 
 ## Functions
 
-There are some "functions" that can be included in a prompt to do various things. 
+There are some "functions" that can be included in a prompt to affect how it is interpreted.
 
-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 `\,`
+Functions have the form `FUNCNAME(param1, param2, ...)`. How parameters are interpreted is up to the function. 
 
-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.
+In general, function parameters will have default values that are used if the parameter is left empty.
 
-### SDXL
+Note: Whitespace is usually *not* stripped from string parameters by default. Commas can be escaped with `\,`
+
+Like `AND`, functions are parsed after regular scheduling syntax has been expanded, allowing things like `[AREA:MASK:0.3](...)`, in case that's somehow useful.
+
+### BREAK
+The keyword `BREAK` causes the prompt to be tokenized in separate chunks, padding each chunk to the text encoder's maximum size before encoding.
+
+For some text encoders (like t5), this operation doesn't really make sense and BREAKs are simply ignored.
+
+### CAT
+
+`CAT` encodes each prompt separately before concatenating the resulting tensors into a single conditioning. It behaves identically to ComfyUI's `ConditioningConcat`.
+
+### AVG()
+
+`prompt1 AVG(weight) prompt2` encodes prompt1 and prompt2 separately, and then combines them using `ConditioningAverage`. The default for `weight` is `0.5`.
+
+`AVG` is processed before `BREAK` but after `AND`
+
+`p1 AVG() p2 AVG() p3` combines `p1` and `p2` first, then combines the result with `p3`.
+
+## Prompt weighting (also known as "Advanced CLIP Encode")
+
+### STYLE
+
+Use the syntax `STYLE(weight_interpretation, normalization)` in a prompt to affect how prompts are interpreted.
+
+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:
+
+```
+STYLE(A1111) a (red:1.1) cat with (brown:0.9) spots and a long tail AND an (old:0.5) dog AND a (green:1.4) (balloon:1.1)
+```
+will interpret everything as A1111, but
+```
+a (red:1.1) cat with (brown:0.9) spots and a long tail AND STYLE(A1111) an (old:0.5) dog AND a (green:1.4) (balloon:1.1)
+```
+Will interpret the first one using the default ComfyUI behaviour, the second prompt with A1111 and the last prompt with the default again
+
+### SDXL: Configure SDXL prompting parameters
 
 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)`.
+### TE: Per-encoder prompts for multi-encoder models
+
+You can specify per-encoder prompts using the `TE` function. The syntax is as follows:
+`TE(encoder_name=prompt)`. Whitespace surrounding the prompt and encoder name are ignored.
+
+For example:
+```
+TE(l=cat) TE(g = (dog:1.1)) TE(t5xxl=tiger)
+```
+The keys to use depend on what key ComfyUI uses for the encoder; for example `l` for CLIP L, `g` for CLIP G, and `t5xxl` for T5 XXL (Flux text encoder).
+
+Use `TE(help)` to print a help text listing available keys.
 
 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`.
+- If you set a prompt with `TE`, it will override the prompt outside the function for the specified text encoder.
+- Multiple instances of `TE` are joined with a space. That is, `TE(l=foo)TE(l=bar)` is the same as `TE(l=foo bar)`
+- `AND` and `BREAK` are processed before `TE`, so they do not do anything sensible; `TE(l=foo AND bar)` will parse as two prompts `TE(foo` and `bar)`. `SHIFT`, `SHUFFLE` and `OLDBREAK` do work, however.
 
-### SHUFFLE and SHIFT
+### SHUFFLE and SHIFT: Create prompt permutations
 
 Default parameters: `SHUFFLE(seed=0, separator=,, joiner=,)`, `SHIFT(steps=0, separator=,, joiner=,)`
 
@@ -155,23 +242,32 @@ For example:
 Whitespace is *not* stripped and may also be used as a joiner or separator
 - `SHIFT(1,, ) cat,dog` results in `dog cat`
 
-### NOISE
+### NOISE: Add noise to a prompt
 
-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.
+The function `NOISE(weight, seed)` adds some random noise into the cond tensor. The seed is optional, and if not specified, the global RNG is used. `weight` should be between 0 and 1.
+
+The usefulness of this is questionable, but it wasn't difficult to implement, so here it is.
+
+
+## Regional prompting
+
+See also [Attention Couple](#attention-couple) below
 
 ### 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.
 
+Multiple `MASK` or `IMASK` calls will be composited together using ComfyUI's `MaskComposite` node, using `op` as the `operation` parameter (defaulting to `multiply`).
+
 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`
+### 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
+### 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,
@@ -184,7 +280,7 @@ Note that because the default values are percentages, `MASK(0 256, 64 512)` is v
 
 Masking does not affect LoRA scheduling unless you set unet weights to 0 for a LoRA.
 
-### FEATHER
+### FEATHER: Mask operations
 
 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`.
 
@@ -199,6 +295,30 @@ gives you a mask that is a combination of 1, 2 and 3, where 1 and 3 are feathere
 
 The order of the `FEATHER` and `MASK` calls doesn't matter; you can have `FEATHER` before `MASK` or even interleave them.
 
+## Cutoff
+
+NOTE: Cutoff syntax might change at some point; it's pretty clunky.
+
+`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:_]
+```
+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.
+
+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 `\`.
+
+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=+`
+
+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:
+
+`[CUT:white cat:white:0.5] and [CUT:black parrot, flying:black:1.0:0.5] and [CUT:green apple:green]`
+
+`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.
+
+The parameters affect how the masked and unmasked prompts are combined to produce the final embedding. Just play around with 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`.
 
@@ -206,9 +326,12 @@ The order of the `FEATHER` and `MASK` calls doesn't matter; you can have `FEATHE
 
 Experimental features are unstable and may disappear or change without warning.
 
-## DEF
+## DEF: Lightweight prompt macros
+
+You can define "prompt macros" by using `DEF`.  Macros are expanded before any other parsing takes place. The expansion continues until no further changes occur. Recursion will raise an error.
+
+`PCLazyTextEncode` and `PCLazyLoraLoader` expand macros, but `PCTextEncode` **does not**. If you need to expand macros for a single prompt, use `PCMacroExpand`
 
-You can define "prompt macros" by using `DEF`:
 ```
 DEF(MYMACRO=this is a prompt)
 [(MYMACRO:0.6):(MYMACRO:1.1):0.5]
@@ -217,7 +340,7 @@ is equivalent to
 ```
 [(this is a prompt:0.5):(this is a prompt:1.1):0.5]
 ```
-
+### Macro parameters
 It's also possible to give parameters to a macro:
 ```
 DEF(MYMACRO=[(prompt $1:$2):(prompt $1:$3):$4])
@@ -227,26 +350,51 @@ gives
 ```
 [(prompt test:1.1):(prompt test:0.7):0.2]
 ```
-in this form, the variables $N (where N is any number corresponding to a positional parameter) will be replaced with the given parameter. The parameters must be separated with a semicolon.
+in this form, the variables $N (where N is any number corresponding to a positional parameter) will be replaced with the given parameter. The parameters must be separated with a semicolon, and can be empty.
+
+You can also optionally specify default values:
 
-Note that unspecified parameters will not be substituted:
 ```
-DEF(mything=a $1 b $2)
-mything
-mything(A)
+DEF(MACRO(example; 0; 1)=[$1:$2,$3])
+MACRO MACRO(test; 0.2)
 ```
 gives
 ```
-a $1 b $2
-a A b $2
+[example:0,1] [test:0.2,1]
 ```
 
-Macros are expanded before any other parsing takes place. The expansion continues until no further changes occur. Recursion will raise an error.
+```
+DEF(MACRO() = [a:$1:0.5])
+```
+sets the default value of `$1` to an empty string.
 
-## Attention masking
+### Unspecified parameters in macros
 
-Use `ATTN()` in combination with `MASK()` or `IMASK()` to enable attention masking. Currently, it's pretty slow and only works with SDXL. You need to have a recent enough version of ComfyUI for this to work.
+Unspecified parameters (either via defaults or explicitly given) will not be substituted. Compare:
+
+```
+DEF(mything=a "$1" b "$2")
+mything
+mything()
+mything(A)
+```
+
+gives
+
+```
+a "$1" b "$2"
+a "" b "$2"
+a "A" b "$2"
+```
+
+## ATTN: Attention couple
+
+See [here](doc/attention_couple.md)
 
 ## TE_WEIGHT
 
-For models using multiple text encoders, you can set weights per TE using the syntax `TE_WEIGHT(clipname=weight, clipname2=weight2, ...)` where `clipname` is one of `g`, `l`, or `t5xxl`. For example with SDXL, try `TE_WEIGHT(g=0.25, l=0.75`). The weights are applied as a multiplier to the TE output.
+For models using multiple text encoders, you can set weights per TE using the syntax `TE_WEIGHT(clipname=weight, clipname2=weight2, ...)` where `clipname` is one of the encoder names printed by `TE(help)`. For example with SDXL, try `TE_WEIGHT(g=0.25, l=0.75)`.
+
+The weights are applied as a multiplier to the TE output. You can also override pooled output multipliers using eg. `l_pooled`.
+
+To set a default value for all encoders, use `TE_WEIGHT(all=weight)`

example_workflows/Basic Text to Image.json

@@ -1,4 +1,6 @@
 {
+  "id": "e820c2fb-9502-45b7-a864-684757dddcdf",
+  "revision": 0,
   "last_node_id": 18,
   "last_link_id": 20,
   "nodes": [
@@ -21,26 +23,26 @@
         {
           "name": "MODEL",
           "type": "MODEL",
+          "slot_index": 0,
           "links": [
             2
-          ],
-          "slot_index": 0
+          ]
         },
         {
           "name": "CLIP",
           "type": "CLIP",
+          "slot_index": 1,
           "links": [
             3
-          ],
-          "slot_index": 1
+          ]
         },
         {
           "name": "VAE",
           "type": "VAE",
+          "slot_index": 2,
           "links": [
-            16
-          ],
-          "slot_index": 2
+            18
+          ]
         }
       ],
       "properties": {
@@ -66,7 +68,7 @@
       "flags": {
         "collapsed": true
       },
-      "order": 8,
+      "order": 7,
       "mode": 0,
       "inputs": [
         {
@@ -87,10 +89,10 @@
         {
           "name": "CONDITIONING",
           "type": "CONDITIONING",
+          "slot_index": 0,
           "links": [
             12
-          ],
-          "slot_index": 0
+          ]
         }
       ],
       "title": "PC: Schedule Prompt (positive)",
@@ -109,24 +111,26 @@
       "id": 3,
       "type": "PCLazyLoraLoader",
       "pos": [
-        255,
-        -765
+        257.5,
+        -745
       ],
       "size": [
         210,
-        98
+        78
       ],
       "flags": {},
-      "order": 7,
+      "order": 6,
       "mode": 0,
       "inputs": [
         {
           "name": "model",
+          "shape": 7,
           "type": "MODEL",
           "link": 2
         },
         {
           "name": "clip",
+          "shape": 7,
           "type": "CLIP",
           "link": 3
         },
@@ -143,19 +147,19 @@
         {
           "name": "MODEL",
           "type": "MODEL",
+          "slot_index": 0,
           "links": [
             17
-          ],
-          "slot_index": 0
+          ]
         },
         {
           "name": "CLIP",
           "type": "CLIP",
+          "slot_index": 1,
           "links": [
             5,
             9
-          ],
-          "slot_index": 1
+          ]
         }
       ],
       "properties": {
@@ -181,7 +185,7 @@
         474
       ],
       "flags": {},
-      "order": 10,
+      "order": 9,
       "mode": 0,
       "inputs": [
         {
@@ -209,10 +213,10 @@
         {
           "name": "LATENT",
           "type": "LATENT",
+          "slot_index": 0,
           "links": [
             15
-          ],
-          "slot_index": 0
+          ]
         }
       ],
       "properties": {
@@ -242,13 +246,16 @@
         225
       ],
       "flags": {},
-      "order": 5,
+      "order": 1,
       "mode": 0,
       "inputs": [],
       "outputs": [
         {
           "name": "STRING",
           "type": "STRING",
+          "widget": {
+            "name": "text"
+          },
           "links": [
             6,
             7
@@ -277,17 +284,20 @@
         225
       ],
       "flags": {},
-      "order": 4,
+      "order": 2,
       "mode": 0,
       "inputs": [],
       "outputs": [
         {
           "name": "STRING",
           "type": "STRING",
+          "widget": {
+            "name": "text"
+          },
+          "slot_index": 0,
           "links": [
             8
-          ],
-          "slot_index": 0
+          ]
         }
       ],
       "title": "Negative prompt",
@@ -314,7 +324,7 @@
       "flags": {
         "collapsed": true
       },
-      "order": 9,
+      "order": 8,
       "mode": 0,
       "inputs": [
         {
@@ -335,10 +345,10 @@
         {
           "name": "CONDITIONING",
           "type": "CONDITIONING",
+          "slot_index": 0,
           "links": [
             13
-          ],
-          "slot_index": 0
+          ]
         }
       ],
       "title": "PC: Schedule Prompt (negative)",
@@ -365,7 +375,7 @@
         106
       ],
       "flags": {},
-      "order": 1,
+      "order": 3,
       "mode": 0,
       "inputs": [],
       "outputs": [
@@ -402,7 +412,7 @@
         46
       ],
       "flags": {},
-      "order": 11,
+      "order": 10,
       "mode": 0,
       "inputs": [
         {
@@ -420,53 +430,18 @@
         {
           "name": "IMAGE",
           "type": "IMAGE",
+          "slot_index": 0,
           "links": [
             20
-          ],
-          "slot_index": 0
+          ]
         }
       ],
       "properties": {
         "cnr_id": "comfy-core",
         "ver": "0.3.18",
         "Node name for S&R": "VAEDecode"
-      }
-    },
-    {
-      "id": 11,
-      "type": "Reroute",
-      "pos": [
-        1155,
-        -900
-      ],
-      "size": [
-        75,
-        26
-      ],
-      "flags": {},
-      "order": 6,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "",
-          "type": "*",
-          "link": 16
-        }
-      ],
-      "outputs": [
-        {
-          "name": "",
-          "type": "VAE",
-          "links": [
-            18
-          ],
-          "slot_index": 0
-        }
-      ],
-      "properties": {
-        "showOutputText": false,
-        "horizontal": false
-      }
+      },
+      "widgets_values": []
     },
     {
       "id": 13,
@@ -480,7 +455,7 @@
         105
       ],
       "flags": {},
-      "order": 3,
+      "order": 4,
       "mode": 0,
       "inputs": [],
       "outputs": [],
@@ -503,7 +478,7 @@
         210
       ],
       "flags": {},
-      "order": 2,
+      "order": 5,
       "mode": 0,
       "inputs": [],
       "outputs": [],
@@ -526,7 +501,7 @@
         405
       ],
       "flags": {},
-      "order": 12,
+      "order": 11,
       "mode": 0,
       "inputs": [
         {
@@ -634,14 +609,6 @@
       0,
       "LATENT"
     ],
-    [
-      16,
-      1,
-      2,
-      11,
-      0,
-      "*"
-    ],
     [
       17,
       3,
@@ -650,14 +617,6 @@
       0,
       "MODEL"
     ],
-    [
-      18,
-      11,
-      0,
-      10,
-      1,
-      "VAE"
-    ],
     [
       20,
       10,
@@ -665,6 +624,14 @@
       18,
       0,
       "IMAGE"
+    ],
+    [
+      18,
+      1,
+      2,
+      10,
+      1,
+      "VAE"
     ]
   ],
   "groups": [],
@@ -673,10 +640,28 @@
     "ds": {
       "scale": 0.8,
       "offset": [
-        588,
-        1260
+        591.75,
+        1235
       ]
-    }
+    },
+    "linkExtensions": [
+      {
+        "id": 18,
+        "parentId": 1
+      }
+    ],
+    "reroutes": [
+      {
+        "id": 1,
+        "pos": [
+          1273.75,
+          -879.5
+        ],
+        "linkIds": [
+          18
+        ]
+      }
+    ]
   },
   "version": 0.4,
   "models": [{

prompt_control/adv_encode.py

@@ -1,6 +1,17 @@
 import torch
 import numpy as np
+from math import copysign
+import logging
 import itertools
+from .adv_encode_old import old_advanced_encode_from_tokens
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+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 _grouper(n, iterable):
@@ -12,29 +23,22 @@ def _grouper(n, iterable):
         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 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 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)
@@ -48,61 +52,6 @@ def mask_word_id(tokens, word_ids, target_id, mask_token):
     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)
@@ -112,34 +61,6 @@ def mask_inds(tokens, inds, mask_token):
     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)
@@ -147,12 +68,6 @@ def scale_emb_to_mag(base_emb, 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
@@ -171,72 +86,307 @@ def perp_weight(weights, unweighted_embs, empty_embs):
     result[~over1] = (unweighted - (1 - weights) * perp)[~over1]
     result[weights == 0.0] = zero[weights == 0.0]
 
+    # Not sure if this is an implementation bug or if this just doesn't make sense with T5
+    nans = result.isnan()
+    if nans.any():
+        log.warning("perp weight returned NaNs (known to happen with T5), replacing with 0")
+        result[nans] = 0.0
+
     return result, unweighted_pooled
 
 
+def style_comfy(encoder, tokens, **kwargs):
+    tokens = encoder.without_word_ids(tokens)
+    return encoder.encode_fn(tokens)
+
+
+def style_a1111(encoder, tokens, **kwargs):
+    base_emb, pooled = encoder.base_emb(tokens)
+    weighted_emb = base_emb * weights_like(encoder.weights(tokens), base_emb)
+    weighted_emb = (base_emb.mean() / weighted_emb.mean()) * weighted_emb  # renormalize
+    return weighted_emb, pooled
+
+
+def style_compel(encoder, tokens, **kwargs):
+    pos_tokens = encoder.weighted_with(tokens, lambda w: w if w > 1.0 else 1.0)
+    weighted_emb, pooled = encoder.encode_fn(pos_tokens)
+    weighted_emb, _, pooled = encoder.down_weight(
+        pos_tokens, encoder.weights(tokens), encoder.word_ids(tokens), weighted_emb, pooled
+    )
+    return weighted_emb, pooled
+
+
+def style_comfypp(encoder, tokens, **kwargs):
+    unweighted_tokens = encoder.unweighted(tokens)
+    base_emb, pooled_base = encoder.base_emb(tokens)
+    weighted_emb, tokens_down, _ = encoder.down_weight(
+        unweighted_tokens, encoder.weights(tokens), encoder.word_ids(tokens), base_emb, pooled_base
+    )
+    weights = encoder.weights(encoder.weighted_with(tokens, lambda w: w if w > 1.0 else 1.0))
+    embs, pooled = encoder.from_masked(
+        unweighted_tokens,
+        weights,
+        encoder.word_ids(tokens),
+        base_emb,
+        pooled_base,
+    )
+    weighted_emb += embs
+
+    return weighted_emb, pooled
+
+
+def style_downweight(encoder, tokens, **kwargs):
+    weights = scale_to_norm(encoder.weights(tokens), encoder.word_ids(tokens), encoder.w_max)
+    base_emb, pooled_base = encoder.base_emb(tokens)
+    weighted_emb, _, pooled = encoder.down_weight(
+        encoder.unweighted(tokens), weights, encoder.word_ids(tokens), base_emb, pooled_base
+    )
+
+    return weighted_emb, pooled
+
+
+def style_perp(encoder, tokens, **kwargs):
+    zero_emb, zero_pooled = encoder.encode_fn(encoder.tokenizer.tokenize_with_weights(""))
+    base_emb, pooled = encoder.base_emb(tokens)
+    return perp_weight(encoder.weights(tokens), (base_emb, pooled), (zero_emb, zero_pooled))
+
+
+def apply_negpip(encoder, emb, pooled, **kwargs):
+    original_tokens = kwargs["original_tokens"]
+    emb_negpip = torch.empty_like(emb).repeat(1, 2, 1)
+    emb_negpip[:, 0::2, :] = emb
+    emb_negpip[:, 1::2, :] = emb * weights_like(encoder.signs(original_tokens), emb)
+    return emb_negpip, pooled
+
+
+def norm_length(encoder, tokens, **kwargs):
+    word_ids = encoder.word_ids(tokens)
+    sums = dict(zip(*np.unique(word_ids, return_counts=True)))
+    sums[0] = 1
+    tokens = [[(t, _norm_mag(w, sums[id]) if id != 0 else 1.0, id) for (t, w, id) in x] for x in tokens]
+    return tokens
+
+
+def norm_mean(encoder, tokens, **kwargs):
+    weights = encoder.weights(tokens)
+    word_ids = encoder.word_ids(tokens)
+    delta = 1 - np.mean([w for x, y in zip(weights, word_ids) for w, id in zip(x, y) if id != 0])
+    tokens = [[(t, w if id == 0 else w + delta, id) for (t, w, id) in x] for x in tokens]
+    return tokens
+
+
+def norm_none(encoder, tokens, **kwargs):
+    return tokens
+
+
+class AdvancedEncoder:
+    STYLES = {
+        "A1111": style_a1111,
+        "comfy": style_comfy,
+        "comfy++": style_comfypp,
+        "compel": style_compel,
+        "down_weight": style_downweight,
+        "perp": style_perp,
+    }
+    NORMALIZATION_OPS = {
+        "none": norm_none,
+        "length": norm_length,
+        "mean": norm_mean,
+    }
+
+    @classmethod
+    def add_encoder(cls, name, fn):
+        cls.STYLES[name] = fn
+
+    def add_normalization_op(cls, name, fn):
+        cls.NORMALIZATION_OPS[name] = fn
+
+    @classmethod
+    def weighted_with(cls, tokens, fn=id, word_ids=True):
+        w = ([(t, fn(w), id) for t, w, id in x] for x in tokens)
+        if not word_ids:
+            w = cls.without_word_ids(w)
+        return list(w)
+
+    @classmethod
+    def unweighted(cls, tokens, word_ids=False):
+        return cls.weighted_with(tokens, fn=lambda w: 1.0, word_ids=word_ids)
+
+    @classmethod
+    def tokens_only(cls, tokens):
+        return list([t[0] for t in x] for x in tokens)
+
+    @classmethod
+    def weights(cls, tokens):
+        return list([t[1] for t in x] for x in tokens)
+
+    @classmethod
+    def word_ids(cls, tokens):
+        return list([t[2] for t in x] for x in tokens)
+
+    @classmethod
+    def signs(cls, tokens):
+        return list([copysign(1, t[1]) for t in x] for x in tokens)
+
+    @classmethod
+    def without_word_ids(cls, tokens):
+        return list([(t, w) for t, w, _ in x] for x in tokens)
+
+    def __init__(self, encode_fn, style, normalization, tokenizer, m_token="+", w_max=1.0, **extra_args):
+        self.encode_fn = encode_fn
+        self.preprocessors = []
+        self.postprocessors = []
+        self.tokenizer = tokenizer
+        self.extra_args = extra_args
+        self.m_token = tokenizer.tokenize_with_weights(m_token)[0][tokenizer.tokens_start]
+        self.max_length = tokenizer.max_length if tokenizer.pad_to_max_length else None
+        self.w_max = w_max
+
+        if style == "comfy++" and not self.max_length:
+            log.warning("comfy++ does not work with tokenizer %s, using default weighting", tokenizer)
+            style = "comfy"
+
+        norms = normalization.split("+")
+        assert style in self.STYLES, f"Invalid weight interpretation: {style}"
+        self.weight_fn = self.STYLES[style]
+        for n in norms:
+            n = n.strip()
+            assert n in self.NORMALIZATION_OPS, f"Invalid normalization: {normalization}"
+            self.preprocessors.append(self.NORMALIZATION_OPS[n])
+
+        negpip = extra_args.get("has_negpip")
+        if negpip:
+
+            def _encode(t):
+                emb, pooled = encode_fn(t)
+                return emb[:, 0::2, :], pooled
+
+            self.encode_fn = _encode
+            self.preprocessors.insert(lambda encoder, tokens, **kwargs: encoder.weighted_with(tokens, abs))
+            self.postprocessors.insert(0, apply_negpip)
+
+    def base_emb(self, tokens):
+        unweighted = self.unweighted(tokens)
+        return self.encode_fn(unweighted)
+
+    def down_weight(self, tokens, weights, word_ids, base_emb, pooled_base):
+        w, w_inv = np.unique(weights, return_inverse=True)
+
+        if np.sum(w < 1) == 0:
+            return (
+                base_emb,
+                tokens,
+                (
+                    base_emb[0, self.max_length - 1 : self.max_length, :]
+                    if (pooled_base is not None and self.max_length)
+                    else None
+                ),
+            )
+
+        masked_current = tokens
+        emblist = [base_emb]
+        for i in range(len(w)):
+            if w[i] >= 1:
+                continue
+            masked_current = mask_inds(masked_current, np.where(w_inv == i)[0], self.m_token)
+            masked, _ = self.encode_fn(masked_current)
+            emblist.append(masked)
+
+        embs = torch.cat(emblist)
+        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)
+        pooled = pooled_base
+        if pooled is not None and self.max_length:
+            pooled = weighted_emb[0, self.max_length - 1 : self.max_length, :]
+        return weighted_emb, masked_current, pooled
+
+    def from_masked(self, tokens, weights, word_ids, base_emb, pooled_base):
+        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), torch.zeros_like(pooled_base) if pooled_base is not None else None
+
+        weight_tensor = weights_like(weights, base_emb)
+
+        ws = []
+        masked_tokens = []
+        masks = []
+
+        # create prompts
+        for id, w in weight_dict.items():
+            masked, m = mask_word_id(tokens, word_ids, id, self.m_token)
+            masks.append(weights_like(m, base_emb))
+            masked_tokens.extend(masked)
+
+            ws.append(w)
+
+        # TODO: figure out how to get rid of this
+        embs = batched_clip_encode(masked_tokens, self.max_length, self.encode_fn, len(tokens))
+        masks = torch.cat(masks)
+
+        embs = base_emb.expand(embs.shape) - embs
+        if pooled_base is not None and self.max_length:
+            pooled = embs[0, self.max_length - 1 : self.max_length, :]
+            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)
+            pooled = pooled_base + pooled
+
+        if embs.shape[0] != masks.shape[0]:
+            embs = embs.repeat(masks.shape[0], 1, 1)
+        embs *= masks
+        embs = embs.sum(axis=0, keepdim=True)
+
+        return ((weight_tensor - 1) * embs), pooled
+
+    def __call__(self, tokens, apply_to_pooled=False, return_pooled=False):
+        normalized_tokens = tokens
+        for op in self.preprocessors:
+            normalized_tokens = op(self, normalized_tokens)
+
+        emb, pooled = self.weight_fn(self, normalized_tokens, original_tokens=tokens)
+
+        for fn in self.postprocessors:
+            emb, pooled = fn(self, emb, pooled, tokens=tokens, original_tokens=tokens)
+
+        if return_pooled:
+            if not apply_to_pooled:
+                _, pooled = self.base_emb(tokens)
+            return emb, pooled
+        return emb, None
+
+
 def advanced_encode_from_tokens(
     tokenized,
     token_normalization,
     weight_interpretation,
     encode_func,
-    m_token=266,
-    length=77,
+    m_token="+",
     w_max=1.0,
     return_pooled=False,
     apply_to_pooled=False,
-    **extra_args
+    tokenizer=None,
+    **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
+    if "old+" not in weight_interpretation:
+        enc = AdvancedEncoder(
+            encode_func, weight_interpretation, token_normalization, tokenizer, m_token, w_max, **extra_args
+        )
+        return enc(tokenized, return_pooled=return_pooled, apply_to_pooled=apply_to_pooled)
     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(""))
+        weight_interpretation = weight_interpretation.replace("old+", "")
+        log.warning("Using old implementation of %s", weight_interpretation)
+        return old_advanced_encode_from_tokens(
+            tokenized,
+            token_normalization,
+            weight_interpretation,
+            encode_func,
+            266,
+            return_pooled=return_pooled,
+            apply_to_pooled=apply_to_pooled,
         )
-
-    if return_pooled:
-        if apply_to_pooled:
-            return weighted_emb, pooled
-        else:
-            return weighted_emb, pooled_base
-    return weighted_emb, None

prompt_control/adv_encode_old.py

@@ -0,0 +1,235 @@
+import torch
+import numpy as np
+import logging
+import itertools
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+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 _grouper(n, iterable):
+    it = iter(iterable)
+    while True:
+        chunk = list(itertools.islice(it, n))
+        if not chunk:
+            return
+        yield chunk
+
+
+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 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 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 = torch.tensor(weights, dtype=base_emb.dtype, device=base_emb.device)
+    weight_tensor = weight_tensor.reshape(1, -1, 1).expand(base_emb.shape)
+
+    # 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)
+
+        m = torch.tensor(m, dtype=base_emb.dtype, device=base_emb.device)
+        m = m.reshape(1, -1, 1).expand(base_emb.shape)
+        masks.append(m)
+
+        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):
+    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 = (266, 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
+
+
+# For verification
+def A1111_renorm(base_emb, weighted_emb):
+    embeddings_final = (base_emb.mean() / weighted_emb.mean()) * weighted_emb
+    return embeddings_final
+
+
+def from_zero(weights, base_emb):
+    weight_tensor = torch.tensor(weights, dtype=base_emb.dtype, device=base_emb.device)
+    weight_tensor = weight_tensor.reshape(1, -1, 1).expand(base_emb.shape)
+    return base_emb * weight_tensor
+
+
+def old_advanced_encode_from_tokens(
+    tokenized,
+    token_normalization,
+    weight_interpretation,
+    encode_func,
+    m_token=266,
+    w_max=1.0,
+    return_pooled=False,
+    apply_to_pooled=False,
+    **extra_args,
+):
+    length = 77
+    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]
+
+    # weight normalization
+    # ====================
+
+    # distribute down/up weights over word lengths
+    if token_normalization.startswith("length"):
+        weights = divide_length(word_ids, weights)
+
+    # make mean of word tokens 1
+    if token_normalization.endswith("mean"):
+        weights = shift_mean_weight(word_ids, weights)
+
+    # weight interpretation
+    # =====================
+    pooled = None
+
+    if weight_interpretation in ["comfy", "perp"]:
+        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 = from_zero(weights, base_emb)
+        weighted_emb = A1111_renorm(base_emb, weighted_emb)
+        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 return_pooled:
+        if apply_to_pooled:
+            return weighted_emb, pooled
+        else:
+            return weighted_emb, pooled_base
+    return weighted_emb, None

prompt_control/attention_couple_ppm.py

@@ -0,0 +1,241 @@
+# Lifted from https://github.com/pamparamm/ComfyUI-ppm/blob/c3e6b673ee2d424405dcb99aeed89f21943c89ac/nodes_ppm/attention_couple_ppm.py
+# Original implementation by laksjdjf, hako-mikan, Haoming02 licensed under GPL-3.0
+# https://github.com/laksjdjf/cgem156-ComfyUI/blob/1f5533f7f31345bafe4b833cbee15a3c4ad74167/scripts/attention_couple/node.py
+# https://github.com/Haoming02/sd-forge-couple/blob/e8e258e982a8d149ba59a4bc43b945467604311c/scripts/attention_couple.py
+import itertools
+import logging
+import math
+from typing import Any
+
+import torch
+import torch.nn.functional as F
+
+from comfy.hooks import EnumHookScope, HookGroup, TransformerOptionsHook, set_hooks_for_conditioning
+from comfy.model_patcher import ModelPatcher
+
+log = logging.getLogger("comfyui-prompt-control")
+
+
+def set_cond_attnmask(base_cond, extra_conds, fill=False):
+    hook = AttentionCoupleHook()
+    c = [base_cond[0][0], base_cond[0][1].copy()]
+    # hook uses these, remove them to avoid doing latent masking
+    c[1].pop("mask", None)
+    c[1].pop("strength", None)
+    c[1].pop("mask_strength", None)
+    c = [c]
+    c.extend(base_cond[1:])
+
+    hook.initialize_regions(base_cond[0], extra_conds, fill=fill)
+    group = HookGroup()
+    group.add(hook)
+
+    return set_hooks_for_conditioning(c, hooks=group)
+
+
+def get_mask(mask, batch_size, num_tokens, extra_options):
+    activations_shape = extra_options["activations_shape"]
+    size = activations_shape[-2:]
+
+    num_conds = mask.shape[0]
+    mask_downsample = F.interpolate(mask, size=size, mode="nearest")
+    mask_downsample_reshaped = mask_downsample.view(num_conds, num_tokens, 1).repeat_interleave(batch_size, dim=0)
+
+    return mask_downsample_reshaped
+
+
+class Proxy:
+    def __init__(self, function):
+        self.function = function
+
+    def to(self, *args, **kwargs):
+        self.function.__self__.to(*args, **kwargs)
+        return self
+
+    def __call__(self, *args, **kwargs):
+        return self.function(*args, *kwargs)
+
+
+class AttentionCoupleHook(TransformerOptionsHook):
+    COND_UNCOND_COUPLE_OPTION = "cond_or_uncond_hook_couple"
+    COND = 0
+    UNCOND = 1
+
+    def __init__(self):
+        super().__init__(hook_scope=EnumHookScope.HookedOnly)
+
+        self.transformers_dict = {
+            "patches": {
+                "attn2_output_patch": [Proxy(self.attn2_output_patch)],
+                "attn2_patch": [Proxy(self.attn2_patch)],
+            }
+        }
+        self.has_negpip = False
+
+        # calculate later
+        self.conds_k: list[torch.Tensor] = None
+        self.conds_v: list[torch.Tensor] = None
+
+    def initialize_regions(self, base_cond, conds, fill):
+        self._base_cond = base_cond
+        self._conds = conds
+        self._fill = fill
+
+        self.num_conds = len(conds) + 1
+        self.base_strength = base_cond[1].get("strength", 1.0)
+        self.strengths = [cond[1].get("strength", 1.0) for cond in conds]
+        self.conds: list[torch.Tensor] = [base_cond[0]] + [cond[0] for cond in conds]
+        base_mask = base_cond[1].get("mask", None)
+        masks = [cond[1].get("mask") * cond[1].get("mask_strength") for cond in conds]
+        if len(masks) < 1:
+            raise ValueError("Attention Couple hook makes no sense without masked conds")
+
+        if any(m is None for m in masks):
+            raise ValueError("All conds given to Attention Couple must have masks")
+
+        if any(m.shape != masks[0].shape for m in masks) or (
+            base_mask is not None and base_mask.shape != masks[0].shape
+        ):
+            largest_shape = max(m.shape for m in masks)
+            if base_mask is not None:
+                largest_shape = max(largest_shape, base_mask.shape)
+            print("largest shape x", largest_shape, [m.shape for m in masks], base_mask.shape)
+            log.warning("Attention Couple: Masks are irregularly shaped, resizing them all to match the largest")
+            for i in range(len(masks)):
+                masks[i] = F.interpolate(masks[i].unsqueeze(1), size=largest_shape[1:], mode="nearest-exact").squeeze(1)
+
+            if base_mask is not None:
+                base_mask = F.interpolate(base_mask.unsqueeze(1), size=largest_shape[1:], mode="nearest-exact").squeeze(
+                    1
+                )
+
+        if base_mask is None:
+            if not fill:
+                raise ValueError("You must specify a base mask when fill=False")
+            sum = torch.stack(masks, dim=0).sum(dim=0)
+            base_mask = torch.zeros_like(sum)
+            base_mask[sum <= 0] = 1.0
+
+        mask = [base_mask] + masks
+        mask = torch.stack(mask, dim=0)
+        if mask.sum(dim=0).min() <= 0 and not fill:
+            raise ValueError("Masks contain non-filled areas")
+
+        self.mask = mask / mask.sum(dim=0, keepdim=True)
+
+    def on_apply_hooks(self, model: ModelPatcher, transformer_options: dict[str, Any]):
+        if self.conds_k is None:
+            self.has_negpip = model.model_options.get("ppm_negpip", False)
+            log.debug("AttentionCouple has_negpip=%s", self.has_negpip)
+
+            # Skip the base cond here, which is always first
+            if self.has_negpip:
+                self.conds_k = [cond[:, 0::2] for cond in self.conds[1:]]
+                self.conds_v = [cond[:, 1::2] for cond in self.conds[1:]]
+            else:
+                self.conds_k = self.conds_v = self.conds[1:]
+
+        return super().on_apply_hooks(model, transformer_options)
+
+    def clone(self):
+        c: AttentionCoupleHook = super().clone()
+        c.initialize_regions(self._base_cond, self._conds, self._fill)
+        return c
+
+    def to(self, *args, **kwargs):
+        self.conds = [c.to(*args, **kwargs) for c in self.conds]
+        self.mask = self.mask.to(*args, **kwargs)
+        return self
+
+    def attn2_patch(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, extra_options):
+        cond_or_uncond = extra_options["cond_or_uncond"]
+        cond_or_uncond_couple = extra_options[self.COND_UNCOND_COUPLE_OPTION] = list(cond_or_uncond)
+        num_chunks = len(cond_or_uncond)
+
+        lcm_tokens_k = math.lcm(k.shape[1], *(cond.shape[1] for cond in self.conds_k))
+        lcm_tokens_v = math.lcm(v.shape[1], *(cond.shape[1] for cond in self.conds_v))
+        q_chunks = q.chunk(num_chunks, dim=0)
+        k_chunks = k.chunk(num_chunks, dim=0)
+        v_chunks = v.chunk(num_chunks, dim=0)
+
+        bs = q.shape[0] // num_chunks
+
+        conds_k_tensor = conds_v_tensor = torch.cat(
+            [
+                cond.repeat(bs, lcm_tokens_k // cond.shape[1], 1) * self.strengths[i]
+                for i, cond in enumerate(self.conds_k)
+            ],
+            dim=0,
+        )
+        if self.has_negpip:
+            conds_v_tensor = torch.cat(
+                [
+                    cond.repeat(bs, lcm_tokens_v // cond.shape[1], 1) * self.strengths[i]
+                    for i, cond in enumerate(self.conds_v)
+                ],
+                dim=0,
+            )
+
+        qs, ks, vs = [], [], []
+        cond_or_uncond_couple.clear()
+
+        for i, cond_type in enumerate(cond_or_uncond):
+            q_target = q_chunks[i]
+            k_target = k_chunks[i].repeat(1, lcm_tokens_k // k.shape[1], 1)
+            v_target = v_chunks[i].repeat(1, lcm_tokens_v // v.shape[1], 1)
+            if cond_type == self.UNCOND:
+                qs.append(q_target)
+                ks.append(k_target)
+                vs.append(v_target)
+                cond_or_uncond_couple.append(self.UNCOND)
+            else:
+                qs.append(q_target.repeat(self.num_conds, 1, 1))
+                ks.append(
+                    torch.cat(
+                        [
+                            k_target * self.base_strength,
+                            conds_k_tensor,
+                        ],
+                        dim=0,
+                    )
+                )
+                vs.append(
+                    torch.cat(
+                        [
+                            v_target * self.base_strength,
+                            conds_v_tensor,
+                        ],
+                        dim=0,
+                    )
+                )
+                cond_or_uncond_couple.extend(itertools.repeat(self.COND, self.num_conds))
+
+        q = torch.cat(qs, dim=0)
+        k = torch.cat(ks, dim=0)
+        v = torch.cat(vs, dim=0)
+
+        return q, k, v
+
+    def attn2_output_patch(self, out, extra_options):
+        cond_or_uncond = extra_options[self.COND_UNCOND_COUPLE_OPTION]
+        bs = out.shape[0] // len(cond_or_uncond)
+        mask_downsample = get_mask(self.mask, bs, out.shape[1], extra_options)
+        outputs = []
+        cond_outputs = []
+        i_cond = 0
+        for i, cond_type in enumerate(cond_or_uncond):
+            pos, next_pos = i * bs, (i + 1) * bs
+
+            if cond_type == self.UNCOND:
+                outputs.append(out[pos:next_pos])
+            else:
+                pos_cond, next_pos_cond = i_cond * bs, (i_cond + 1) * bs
+                masked_output = out[pos:next_pos] * mask_downsample[pos_cond:next_pos_cond]
+                cond_outputs.append(masked_output)
+                i_cond += 1
+
+        if len(cond_outputs) > 0:
+            cond_output = torch.stack(cond_outputs).sum(0)
+            outputs.append(cond_output)
+
+        return torch.cat(outputs, dim=0)

prompt_control/cutoff.py

@@ -209,6 +209,9 @@ def encode_regions(clip_regions, encode, tokenizer):
         debug_tokens("region", region_prompt, tokenizer)
         region_emb, _ = encode(region_prompt)
         region_emb -= base_embedding_start
+        # NegPiP support:
+        if region_emb.shape[1] == 2 * region_masking.shape[1]:
+            region_masking = torch.repeat_interleave(region_masking, 2, dim=1)
         region_emb *= region_masking
 
         region_embeddings.append(region_emb)
@@ -217,6 +220,10 @@ def encode_regions(clip_regions, encode, tokenizer):
     embeddings_final_mask = torch.tensor(
         global_region_mask, dtype=base_embedding_full.dtype, device=base_embedding_full.device
     ).unsqueeze(-1)
+    # NegPiP support:
+    if region_embeddings.shape[1] == 2 * embeddings_final_mask.shape[1]:
+        embeddings_final_mask = torch.repeat_interleave(embeddings_final_mask, 2, dim=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/manual_test.py

@@ -0,0 +1,46 @@
+import main
+import nodes
+import prompt_control.adv_encode
+
+(l,) = nodes.CLIPLoader.load_clip(None, "clip_l.safetensors")
+(t5,) = nodes.CLIPLoader.load_clip(None, "t5base.safetensors")
+
+id(main)  # get rid of warning
+
+
+def adv(t, text, style="A1111", norm="none", new=True, **kwargs):
+    c = t.tokenize(text, return_word_ids=True)
+    if new:
+        style = "new+" + style
+    if t is t5:
+        te = t.patcher.model.t5base.encode_token_weights
+        token = t.tokenizer.clip_t5base
+        tok = c["t5base"]
+    else:
+        te = t.patcher.model.clip_l.encode_token_weights
+        token = t.tokenizer.clip_l
+        tok = c["l"]
+    return prompt_control.adv_encode.advanced_encode_from_tokens(tok, norm, style, te, tokenizer=token)
+
+
+def adv_all(t, text, styles=[], **kwargs):
+    r = []
+    for s in styles or prompt_control.adv_encode.AdvancedEncoder.STYLES:
+        print("Testing", s, kwargs)
+        r.append([s, adv(t, text, style=s, **kwargs)])
+    return r
+
+
+def replacenan(t):
+    t[t.isnan()] = 42.123321
+    return t
+
+
+def adv_equal(t, text, **kwargs):
+    old = adv_all(t, text, new=False, **kwargs)
+    new = adv_all(t, text, new=True, **kwargs)
+    r = {}
+    for i, o in enumerate(old):
+        n = new[i]
+        r[n[0]] = (replacenan(n[1][0]) == replacenan(o[1][0])).all()
+    return r

prompt_control/nodes_attnmask.py

@@ -1,79 +0,0 @@
-import logging
-
-log = logging.getLogger("comfyui-prompt-control")
-
-from comfy.hooks import TransformerOptionsHook, HookGroup, EnumHookScope
-from comfy.ldm.modules.attention import optimized_attention
-import torch.nn.functional as F
-import torch
-from math import sqrt
-
-
-class MaskedAttn2:
-    def __init__(self, mask):
-        self.mask = mask
-
-    def __call__(self, q, k, v, extra_options):
-        mask = self.mask
-        orig_shape = extra_options["original_shape"]
-        _, _, oh, ow = orig_shape
-        seq_len = q.shape[1]
-        mask_h = oh / sqrt(oh * ow / seq_len)
-        mask_h = int(mask_h) + int((seq_len % int(mask_h)) != 0)
-        mask_w = seq_len // mask_h
-        r = optimized_attention(q, k, v, extra_options["n_heads"])
-        mask = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="nearest").squeeze(1)
-        mask = mask.view(mask.shape[0], -1, 1).repeat(1, 1, r.shape[2])
-
-        return mask * r
-
-
-def create_attention_hook(mask):
-    attn_replacements = {}
-    mask = mask.detach().to(device="cuda", dtype=torch.float16)
-
-    masked_attention = MaskedAttn2(mask)
-
-    for id in [4, 5, 7, 8]:  # id of input_blocks that have cross attention
-        block_indices = range(2) if id in [4, 5] else range(10)  # transformer_depth
-        for index in block_indices:
-            k = ("input", id, index)
-            attn_replacements[k] = masked_attention
-    for id in range(6):  # id of output_blocks that have cross attention
-        block_indices = range(2) if id in [3, 4, 5] else range(10)  # transformer_depth
-        for index in block_indices:
-            k = ("output", id, index)
-            attn_replacements[k] = masked_attention
-    for index in range(10):
-        k = ("middle", 1, index)
-        attn_replacements[k] = masked_attention
-
-    hook = TransformerOptionsHook(
-        transformers_dict={"patches_replace": {"attn2": attn_replacements}}, hook_scope=EnumHookScope.HookedOnly
-    )
-    group = HookGroup()
-    group.add(hook)
-
-    return group
-
-
-class AttentionMaskHookExperimental:
-    @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {"mask": ("MASK",)},
-        }
-
-    RETURN_TYPES = ("HOOKS",)
-    CATEGORY = "promptcontrol/_testing"
-    FUNCTION = "apply"
-    EXPERIMENTAL = True
-    DESCRIPTION = "Experimental attention masking hook. For testing only"
-
-    def apply(self, mask):
-        return (create_attention_hook(mask),)
-
-
-NODE_CLASS_MAPPINGS = {"AttentionMaskHookExperimental": AttentionMaskHookExperimental}
-
-NODE_DISPLAY_NAME_MAPPINGS = {}

prompt_control/nodes_hooks.py

@@ -1,9 +1,13 @@
 import logging
-import comfy.utils
+
 import comfy.hooks
+import comfy.utils
 import folder_paths
-from .utils import consolidate_schedule
+from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict
+
+from .attention_couple_ppm import AttentionCoupleHook
 from .parser import parse_prompt_schedules
+from .utils import consolidate_schedule
 
 log = logging.getLogger("comfyui-prompt-control")
 
@@ -79,10 +83,51 @@ def lora_hooks_from_schedule(schedules, non_scheduled):
         return hooks
 
 
+class PCAttentionCoupleBatchNegative(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "positive": (IO.CONDITIONING, {}),
+                "negative": (IO.CONDITIONING, {}),
+            },
+        }
+
+    RETURN_TYPES = (IO.CONDITIONING, IO.CONDITIONING)
+    RETURN_NAMES = ("positive", "negative")
+    CATEGORY = "promptcontrol/v2"
+    FUNCTION = "batch"
+    EXPERIMENTAL = True
+
+    # May cause side-effects?
+    # TODO: Support scheduling in negative prompt
+    def batch(self, positive, negative):
+        if len(negative) != 1:
+            log.warning("Batching scheduled negatives is not supported yet")
+            return (positive, negative)
+
+        negative_batch = []
+        for p in positive:
+            n = [negative[0][0], negative[0][1].copy()]
+            n_hook_group: comfy.hooks.HookGroup = n[1].get("hooks", comfy.hooks.HookGroup()).clone()
+            p_hook_group: comfy.hooks.HookGroup = p[1].get("hooks", comfy.hooks.HookGroup())
+            attn_couple = [hook for hook in p_hook_group.hooks if isinstance(hook, AttentionCoupleHook)]
+            for hook in attn_couple:
+                n_hook_group.add(hook)
+            n[1]["hooks"] = p_hook_group if n_hook_group.hooks == p_hook_group.hooks else n_hook_group
+            n[1]["start_percent"] = p[1].get("start_percent", 0.0)
+            n[1]["end_percent"] = p[1].get("end_percent", 1.0)
+            negative_batch.append(n)
+
+        return (positive, negative_batch)
+
+
 NODE_CLASS_MAPPINGS = {
     "PCLoraHooksFromText": PCLoraHooksFromText,
+    "PCAttentionCoupleBatchNegative": PCAttentionCoupleBatchNegative,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
     "PCLoraHooksFromText": "PC: LoRA Hooks From Text (non-lazy)",
+    "PCAttentionCoupleBatchNegative": "PC: Attention Couple (batch negative)",
 }

prompt_control/nodes_tools.py

@@ -1,5 +1,5 @@
 import logging
-from .parser import parse_prompt_schedules
+from .parser import parse_prompt_schedules, expand_macros
 from .nodes_lazy import NODE_CLASS_MAPPINGS as LAZY_NODES
 import json
 import folder_paths
@@ -209,6 +209,24 @@ class PCExtractScheduledPrompt:
         return (prompt_text,)
 
 
+class PCMacroExpand:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "text": ("STRING", {"multiline": True}),
+            },
+        }
+
+    RETURN_TYPES = ("STRING",)
+    CATEGORY = "promptcontrol/tools"
+    FUNCTION = "apply"
+    DESCRIPTION = "Expands DEF macros in a string and returns the result"
+
+    def apply(self, text):
+        return (expand_macros(text),)
+
+
 NODE_CLASS_MAPPINGS = {
     "PCSetPCTextEncodeSettings": PCSetPCTextEncodeSettings,
     "PCAddMaskToCLIP": PCAddMaskToCLIP,
@@ -216,6 +234,7 @@ NODE_CLASS_MAPPINGS = {
     "PCSetLogLevel": PCSetLogLevel,
     "PCExtractScheduledPrompt": PCExtractScheduledPrompt,
     "PCSaveExpandedWorkflow": PCSaveExpandedWorkflow,
+    "PCMacroExpand": PCMacroExpand,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
@@ -225,4 +244,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "PCSetLogLevel": "PC: Configure Logging (for debug)",
     "PCExtractScheduledPrompt": "PC: Extract Scheduled Prompt",
     "PCSaveExpandedWorkflow": "PC: Save Expanded Workflow (for debug)",
+    "PCMacroExpand": "PC: Expand Macros",
 }

prompt_control/parser.py

@@ -8,7 +8,7 @@ log = logging.getLogger("comfyui-prompt-control")
 import re
 
 from functools import lru_cache
-from .utils import get_function
+from .utils import get_function, find_closing_paren
 
 if lark.__version__ == "0.12.0":
     from sys import executable
@@ -31,8 +31,9 @@ prompt: (emphasized | embedding | scheduled | alternate | sequence | loraspec |
 !emphasized: "(" prompt? ")"
         | "(" prompt ":" prompt ")"
         | "[" prompt "]"
-scheduled: "[" [[prompt] ":"] [prompt] ":" _WS? NUMBER ["," NUMBER] "]"
-        | "[" [[prompt] ":"] [prompt] ":" _WS? TAG "]"
+promptlist: ([prompt] ":")~1..3
+scheduled: "[" promptlist _WS? NUMBER ["," NUMBER] "]"
+        | "[" promptlist _WS? TAG "]"
 sequence.5:  "[SEQ" ":" [prompt] ":" NUMBER (":" [prompt] ":" NUMBER)* "]"
 alternate: "[" [prompt] ("|" [prompt])+ [":" NUMBER] "]"
 loraspec.99: "<lora:" FILENAME lora_weights [lora_block_weights] ">"
@@ -91,7 +92,7 @@ def parse_cuts(text):
 
 
 def flatten(x):
-    if type(x) in [str, tuple] or isinstance(x, dict) and "type" in x:
+    if type(x) in [str, tuple, int, type(None)] or isinstance(x, dict) and "type" in x:
         yield x
     else:
         for g in x:
@@ -160,31 +161,38 @@ def get_steps(tree, num_steps):
 def at_step(step, filters, tree):
     class AtStep(lark.Transformer):
         def scheduled(self, args):
+            before = None
+            during = None
+            after = None
             when_end = None
-            before, after, when, *rest = args
-            if isinstance(when, str):
-                return before or "" if when not in filters else after or ""
-
+            pl, when, *rest = args
             if rest:
                 when_end = rest[0]
 
-            if when_end is not None and step <= when and before is not None:
-                return ""
+            pl = list(pl)
+            if len(pl) == 1:
+                (during,) = pl  # [after:0.5] == [::after:0.5,0.5]
+                if when_end is None:
+                    when_end = when
+                    after = during
+            elif len(pl) == 2:
+                during, after = pl  # [during:after:0.5] = [before::after:0.5,0.5]
+                if when_end is None:
+                    when_end = when
+                    before = during
+            else:
+                before, during, after = pl  # [before:during:after:0.5,0.8]
 
-            if when_end is not None and (step > when and step <= when_end):
-                # handle [a:0,1]
-                if before is None:
-                    return after or ""
-                return before or ""
+            if isinstance(when, str):
+                return before or "" if when not in filters else after or ""
 
-            if when_end is not None and step >= when_end:
-                # handle [a:0,1]
-                if before is None:
-                    return ""
-                return after or ""
+            if when_end is None:
+                when_end = 1000_000
 
             if step <= when:
                 return before or ""
+            if when < step <= when_end:
+                return during or ""
             else:
                 return after or ""
 
@@ -303,6 +311,7 @@ class PromptSchedule(object):
         except lark.exceptions.LarkError as e:
             log.error("Prompt editing parse error: %s", e)
             parsed = [[1.0, {"prompt": self.prompt, "loras": {}}]]
+            raise
 
         # Tag filtering may return redundant prompts, so filter them out here
         res = []
@@ -359,17 +368,38 @@ class PromptSchedule(object):
         return len(self.parsed_prompt) - 1, self.parsed_prompt[-1]
 
 
-def replace_def(text):
+def parse_search(search):
+    arg_start = search.find("(")
+    args = ""
+    name = search.strip()
+    if arg_start > 0:
+        arg_end = find_closing_paren(search, arg_start)
+        name = search[:arg_start].strip()
+        args = search[arg_start + 1 : arg_end - 1]
+
+    if not name:
+        return None
+    args = args.strip()
+    # If using the form DEF(F()=$1) then the default value of $1 is the empty string
+    if arg_start > 0:
+        args = [a.strip() for a in args.split(";")]
+    else:
+        args = []
+    return name, args
+
+
+def expand_macros(text):
     text, defs = get_function(text, "DEF", defaults=None)
     res = text
     prevres = text
     replacements = []
     for d in defs:
         r = d.split("=", 1)
-        if len(r) != 2 or not r[0].strip():
+        search = parse_search(r[0].strip())
+        if not search or len(r) != 2:
             log.warning("Ignoring invalid DEF(%s)", d)
             continue
-        replacements.append((r[0].strip(), r[1].strip()))
+        replacements.append((search, r[1].strip()))
     iterations = 0
     while True:
         iterations += 1
@@ -389,22 +419,30 @@ def replace_def(text):
 
 
 def substitute_def(text, search, replace):
+    search, default_args = search
+    for i, v in enumerate(default_args):
+        replace = re.sub(rf"\${i+1}\b", v, replace)
     return re.sub(rf"\b{re.escape(search)}\b", replace, text)
 
 
 def substitute_defcall(text, search, replace):
-    text, defns = get_function(text, search, defaults=None, placeholder=f"DEFNCALL{search}")
+    name, default_args = search
+    text, defns = get_function(text, name, defaults=None, placeholder=f"DEFNCALL{search}")
     for i, defn in enumerate(defns):
         ph = f"\0DEFNCALL{search}{i}\0"
         paramvals = [x.strip() for x in defn.split(";")]
         r = replace
         for i, v in enumerate(paramvals):
             r = re.sub(rf"\${i+1}\b", v, r)
+
+        for i, v in enumerate(default_args):
+            r = re.sub(rf"\${i+1}\b", v, r)
+
         text = text.replace(ph, r)
     return text
 
 
 @lru_cache
 def parse_prompt_schedules(prompt, **kwargs):
-    prompt = replace_def(prompt)
+    prompt = expand_macros(prompt)
     return PromptSchedule(prompt, **kwargs)

prompt_control/prompts.py

@@ -3,26 +3,14 @@ import re
 import torch
 from functools import partial
 from comfy_extras.nodes_mask import FeatherMask, MaskComposite
+from nodes import ConditioningAverage
 
 from .utils import safe_float, get_function, parse_floats, smarter_split
 from .adv_encode import advanced_encode_from_tokens
 from .cutoff import process_cuts
 from .parser import parse_cuts
 
-try:
-    from .nodes_attnmask import create_attention_hook
-    from comfy.hooks import set_hooks_for_conditioning
-
-    def set_cond_attnmask(cond, mask):
-        hook = create_attention_hook(mask)
-        return set_hooks_for_conditioning(cond, hooks=hook)
-
-except ImportError:
-
-    def set_cond_attnmask(cond, mask):
-        log.info("Attention masking is not available")
-        return cond
-
+from .attention_couple_ppm import set_cond_attnmask
 
 log = logging.getLogger("comfyui-prompt-control")
 
@@ -77,13 +65,14 @@ def get_style(text, default_style="comfy", default_normalization="none"):
     style, normalization = styles[0]
     style = style.strip()
     normalization = normalization.strip()
-    if style not in AVAILABLE_STYLES:
+    if style.replace("old+", "") 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
+    for part in normalization.split("+"):
+        if part not in AVAILABLE_NORMALIZATIONS:
+            log.warning("Unrecognized prompt normalization: %s. Using %s", normalization, default_normalization)
+            normalization = default_normalization
+            break
 
     return style, normalization, text
 
@@ -138,6 +127,83 @@ def fix_word_ids(tokens):
     return tokens
 
 
+def tokenize_chunks(clip, text, need_word_ids, can_break):
+    chunks = re.split(r"\bBREAK\b", text)
+    token_chunks = []
+    shuffled_chunks = []
+    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)
+        shuffled_chunks.append(r)
+        t = clip.tokenize(c, return_word_ids=need_word_ids)
+        token_chunks.append(t)
+
+    tokens = token_chunks[0]
+    full_prompt = "".join(shuffled_chunks)
+    full_tokenized = tokens
+    if len(chunks) > 1:
+        full_tokenized = clip.tokenize(full_prompt, return_word_ids=need_word_ids)
+        for key in tokens:
+            if not can_break.get(key):
+                log.warning("BREAK does not make sense for %s, tokenizing as one chunk. Use CAT instead.", key)
+                tokens[key] = full_tokenized[key]
+                continue
+            for c in token_chunks[1:]:
+                tokens[key].extend(c[key])
+
+    return tokens
+
+
+def tokenize(clip, text, can_break, empty_tokens):
+    # defaults=None means there is no argument parsing at all
+    text, l_prompts = get_function(text, "CLIP_L", defaults=None)
+    text, te_prompts = get_function(text, "TE", defaults=None)
+    need_word_ids = True
+    tokens = tokenize_chunks(clip, text, need_word_ids, can_break)
+
+    per_te_prompts = {}
+    if l_prompts:
+        log.warning("Note: CLIP_L is deprecated. Use TE(l=prompt) instead")
+        per_te_prompts["l"] = l_prompts
+
+    for prompt in te_prompts:
+        if prompt.strip() == "help":
+            log.info("Encoders available for TE: %s", ", ".join(tokens.keys()))
+            continue
+        params = prompt.split("=", 1)
+        if len(params) != 2:
+            log.warning("Invalid TE call, ignoring: %s", prompt)
+            continue
+        te = params[0].strip()
+        prompt = params[1].strip()
+        if te not in tokens:
+            log.warning("Invalid TE call, no TE with key '%s', ignoring: %s", te)
+            log.info("Encoders available for TE: %s", ", ".join(tokens.keys()))
+            continue
+        l = per_te_prompts.get(te, [])
+        l.append(prompt)
+        per_te_prompts[te] = l
+
+    if per_te_prompts:
+        for key in per_te_prompts:
+            prompt = " ".join(per_te_prompts[key])
+            tokens[key] = tokenize_chunks(clip, prompt, need_word_ids, can_break)[key]
+            log.info("Encoded prompt with TE '%s': %s", key, prompt)
+
+    maxlen = max([0] + [len(tokens[k]) for k in tokens if can_break[k]])
+    for k in tokens:
+        if not can_break[k]:
+            continue
+        while len(tokens[k]) < maxlen:
+            tokens[k] += empty_tokens[k]
+
+    return fix_word_ids(tokens)
+
+
 def encode_prompt_segment(
     clip,
     text,
@@ -155,52 +221,56 @@ def encode_prompt_segment(
     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 = True
-    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
-        t = clip.tokenize(c, return_word_ids=need_word_ids)
-        token_chunks.append(t)
-    tokens = token_chunks[0]
+    empty = clip.tokenize("", return_word_ids=True)
+    can_break = {}
+    for k in empty:
+        tokenizer = getattr(clip.tokenizer, f"clip_{k}", getattr(clip.tokenizer, k, None))
+        can_break[k] = tokenizer and tokenizer.pad_to_max_length
 
-    for key in tokens:
-        for c in token_chunks[1:]:
-            tokens[key].extend(c[key])
+    clip = hook_te(clip, empty.keys(), style, normalization, extra)
 
-    # 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"]
+    # Chunks to ConditioningAverage:
 
-    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"]
+    text, averages = get_function(text, "AVG", ["0.5"], return_dict=True)
+    prev = 0
+    prompts_to_avg = []
+    for avg in averages:
+        w = safe_float(avg["args"][0], 0.5)
+        p = text[prev : avg["position"]], w
+        prompts_to_avg.append(p)
+        prev = avg["position"]
+    prompts_to_avg.append((text[prev:], 1.0))
 
-    tokens = fix_word_ids(tokens)
+    conds_to_avg = []
+    for prompt, weight in prompts_to_avg:
+        conds_to_cat = []
+        chunks = re.split(r"\bCAT\b", prompt)
+        for c in chunks:
+            tokens = tokenize(clip, c, can_break, empty)
+            conds_to_cat.append(clip.encode_from_tokens_scheduled(tokens, add_dict=settings))
 
-    tes = []
-    for k in tokens:
-        if k in ["g", "l"]:
-            tes.append(f"clip_{k}")
-        else:
-            tes.append(k)
+        base = conds_to_cat[0]
+        for cond in conds_to_cat[1:]:
+            assert len(cond) == len(base), "Conditioning length mismatch"
+            # Pooled gets ignored
+            for i in range(len(base)):
+                c1 = base[i][0]
+                c2 = cond[i][0]
+                base[i][0] = torch.cat((c1, c2), 1)
+        conds_to_avg.append((base, weight))
 
-    clip = hook_te(clip, tes, style, normalization, extra)
+    base, w = conds_to_avg[0]
+    for cond, next_w in conds_to_avg[1:]:
+        assert len(base) == len(cond), "Conditioning length mismatch"
+        if w == 1.0:
+            w = next_w
+            continue
+        for i in range(len(base)):
+            (cond,) = ConditioningAverage.addWeighted(None, [base[i]], [cond[i]], w)
+            base[i] = cond[0]
+        w = next_w
 
-    return clip.encode_from_tokens_scheduled(tokens, add_dict=settings)
+    return base
 
 
 def apply_weights(output, te_name, spec):
@@ -211,21 +281,29 @@ def apply_weights(output, te_name, spec):
     if te_name.startswith("clip_"):
         te_name = te_name[5:]
 
+    default = spec.get("all", None)
+
     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]
+        if te_name in spec or pkey in spec or default is not None:
+            w = spec.get(te_name, default)
+            pooled_w = spec.get(pkey, w)
+            if w is None:
+                w = 1.0
+            if pooled_w is None:
+                pooled_w = 1.0
+            log.info("Weighting %s output by %s, pooled by %s", te_name, w, pooled_w)
+            out = out * w
+            if pooled is not None:
+                pooled = pooled * pooled_w
 
         return out, pooled
     else:
-        if te_name in spec:
-            log.info("Weighting %s output by %s", te_name, spec[te_name])
-            output = output * spec[te_name]
+        if te_name in spec or default is not None:
+            w = spec.get(te_name, default)
+            log.info("Weighting %s output by %s", te_name, w)
+            output = output * w
         return output
 
 
@@ -246,15 +324,24 @@ def hook_te(clip, te_names, style, normalization, extra):
         return clip
     newclip = clip.clone()
     for te_name in te_names:
-        if hasattr(clip.patcher.model, te_name):
+        tokenizer = getattr(clip.tokenizer, f"clip_{te_name}", getattr(clip.tokenizer, te_name, None))
+        if tokenizer:
             x = extra.copy()
-            x["tokenizer"] = getattr(clip.tokenizer, te_name)
-            log.debug("Hooked into %s with style=%s, normalization=%s", te_name, style, normalization)
+            x["tokenizer"] = tokenizer
+            if not hasattr(clip.patcher.model, te_name):
+                te_name = "clip_" + te_name
+            if not hasattr(clip.patcher.model, te_name):
+                log.warning("TE model %s not found on model patcher. Skipping...", te_name)
+                continue
+
+            log.debug("Hooked into te=%s with style=%s, normalization=%s", te_name, style, normalization)
+            encode = clip.patcher.get_model_object(f"{te_name}.encode_token_weights")
+            x["has_negpip"] = clip.patcher.model_options.get("ppm_negpip", False)
             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"),
+                    encode,
                     normalization,
                     style,
                     x,
@@ -262,7 +349,7 @@ def hook_te(clip, te_names, style, normalization, extra):
             )
         # '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)
+            log.warning("Tokens contain items with key %s but no tokenizer found on object with that name.", te_name)
     return newclip
 
 
@@ -328,7 +415,7 @@ def make_mask(args, size, weight):
     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)
+    log.debug("Mask xs=%s, ys=%s, shape=%s, weight=%s", xs, ys, mask.shape, weight)
     return mask
 
 
@@ -444,19 +531,25 @@ def encode_prompt(clip, text, start_pct, end_pct, defaults, masks):
     conds = []
     # TODO: is this still needed?
     # scale = sum(abs(weight(p)[0]) for p in prompts if not ("AREA(" in p or "MASK(" in p))
+    attnmasked_prompts = []
+    fill = False
     for prompt in prompts:
-        attn = False
+        attn_couple = False
+        prompt_has_fill = False
         if "ATTN()" in prompt:
             prompt = prompt.replace("ATTN()", "")
-            attn = True
-            log.info("Using attention masking for prompt segment")
+            attn_couple = True
+        if "FILL()" in prompt:
+            prompt = prompt.replace("FILL()", "")
+            prompt_has_fill = True
         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)
+        # Get weight last so other syntax doesn't interfere with it
+        w, opts, prompt = weight(prompt)
+        if not w:
+            continue
         settings = {"prompt": prompt}
         settings["strength"] = w
         settings.update(sdxl_opts)
@@ -471,12 +564,41 @@ def encode_prompt(clip, text, start_pct, end_pct, defaults, masks):
 
         settings["start_percent"] = start_pct
         settings["end_percent"] = end_pct
-        x = encode_prompt_segment(clip, prompt, settings, style, normalization)
-        if attn and mask is not None:
-            mask = settings.pop("mask")
-            strength = settings.pop("mask_strength")
-            x = set_cond_attnmask(x, mask * strength)
 
-        conds.extend(x)
+        x = encode_prompt_segment(clip, prompt, settings, style, normalization)
+        if attn_couple:
+            if prompt_has_fill:
+                if attnmasked_prompts:
+                    log.warning("FILL() can only be used for the first prompt, ignoring")
+                elif mask is not None:
+                    log.warning("MASK() and FILL() can't be used together, ignoring FILL()")
+                else:
+                    fill = True
+            attnmasked_prompts.extend(x)
+        else:
+            conds.extend(x)
+
+    def ensure_mask(c):
+        if "mask" not in c[1]:
+            _, mask, _ = get_mask("MASK()", mask_size, masks)
+            c[1]["mask"] = mask
+            c[1]["mask_strength"] = 1.0
+        return c
+
+    if attnmasked_prompts:
+        base_cond = attnmasked_prompts[0]
+        if not fill:
+            ensure_mask(base_cond)
+        # else, set_cond_attnmask will have the base mask fill any unspecified areas
+        base_cond = [base_cond]
+        if len(attnmasked_prompts) > 1:
+            base_cond = set_cond_attnmask(
+                base_cond,
+                [ensure_mask(c) for c in attnmasked_prompts[1:]],
+                fill=fill,
+            )
+        else:
+            log.warning("You must specify at least two prompt segments with ATTN() for attention couple to work")
+        conds.extend(base_cond)
 
     return conds

prompt_control/test_encode.py

@@ -0,0 +1,106 @@
+import unittest
+import numpy.testing as npt
+
+clip_l = None
+dual = None
+
+
+def run(f, *args):
+    return getattr(f, f.FUNCTION)(*args)
+
+
+class TestEncode(unittest.TestCase):
+    def tensorsEqual(self, t1, t2):
+        npt.assert_equal(t1.detach().numpy(), t2.detach().numpy())
+
+    def condEqual(self, c1, c2, key=None, key_assert=None):
+        self.assertEqual(len(c1), len(c2))
+        for i in range(len(c1)):
+            a, b = c1[i], c2[i]
+            if key:
+                (key_assert or self.assertEqual)(a[1][key], b[1][key])
+            else:
+                self.tensorsEqual(a[0], b[0])
+
+    def test_basic_encode(self):
+        pc = PCTextEncode()
+        comfy = nodes.CLIPTextEncode()
+        combine = nodes.ConditioningCombine()
+        concat = nodes.ConditioningConcat()
+        zeroout = nodes.ConditioningZeroOut()
+        for k, clip in [("l", clip_l), ("dual", dual)]:
+            with self.subTest(k):
+                with self.subTest("No exceptions"):
+                    run(
+                        pc,
+                        clip,
+                        "test AND test (test:1.2) BREAK test AND TE_WEIGHT(all=0) SDXL() AND AREA(,,) test CAT test",
+                    )
+                with self.subTest("Basic"):
+                    (c1,) = run(pc, clip, "test")
+                    (c2,) = run(comfy, clip, "test")
+                    c = c2  # Used in later tests
+                    self.condEqual(c1, c2)
+
+                    (c1,) = run(pc, clip, "(test:1.2)")
+                    (c2,) = run(comfy, clip, "(test:1.2)")
+
+                with self.subTest("Concat"):
+                    (c1,) = run(pc, clip, "test CAT test")
+                    (c2,) = run(concat, c, c)
+                    self.condEqual(c1, c2)
+
+                with self.subTest("Combine"):
+                    (c1,) = run(pc, clip, "test AND test")
+                    (c2,) = run(combine, c, c)
+                    self.condEqual(c1, c2)
+
+                with self.subTest("Zero out"):
+                    (c1,) = run(pc, clip, "test TE_WEIGHT(all=0)")
+                    (c2,) = run(zeroout, c)
+                    self.condEqual(c1, c2)
+
+    def test_styles(self):
+        pc = PCTextEncode()
+        comfy = nodes.CLIPTextEncode()
+        for k, clip in [("l", clip_l), ("dual", dual)]:
+            (no_weights,) = run(comfy, clip, "this prompt has no weights")
+            for style in ["comfy", "A1111", "comfy++", "compel", "down_weight", "perp"]:
+                with self.subTest(f"TE {k} style {style} no weights equal comfy"):
+                    (c,) = run(pc, clip, "this prompt has no weights")
+                    self.condEqual(no_weights, c)
+                with self.subTest(f"TE {k} style {style} does not fail when encoding weights"):
+                    for normalization in ["none", "mean", "length", "mean+length", "length+mean"]:
+                        with self.subTest(f"TE {k} style {style} normalization {normalization}"):
+                            (c,) = run(
+                                pc,
+                                clip,
+                                f"STYLE({style}, {normalization}) (this prompt) (has weights:0.9), (a:1.2) (b:1.2)",
+                            )
+
+    def test_masks(self):
+        pc = PCTextEncode()
+        comfy = nodes.CLIPTextEncode()
+        solidmask = comfy_extras.nodes_mask.SolidMask()
+        setMask = nodes.ConditioningSetMask()
+        for k, clip in [("l", clip_l), ("dual", dual)]:
+            (c1,) = run(pc, clip, "test MASK()")
+            (c2,) = run(comfy, clip, "test")
+            (c2,) = run(setMask, c2, run(solidmask, 1.0, 512, 512)[0], "default", 1.0)
+            self.condEqual(c1, c2)
+            self.condEqual(c1, c2, "mask", self.tensorsEqual)
+
+
+if __name__ == "__main__":
+    print("Loading ComfyUI")
+    import main
+
+    id(main)  # get rid of flake warning
+    import nodes
+    import comfy_extras.nodes_mask
+    from .nodes_base import PCTextEncode
+
+    (clip_l,) = nodes.CLIPLoader().load_clip("clip_l.safetensors")
+    (dual,) = nodes.DualCLIPLoader().load_clip("clip_l.safetensors", "t5xxl_fp16.safetensors", "flux")
+    print("Starting tests")
+    unittest.main()

prompt_control/test_encode_refactor.py

@@ -0,0 +1,56 @@
+import unittest
+import numpy.testing as npt
+
+clip_l = None
+dual = None
+
+
+def run(f, *args):
+    return getattr(f, f.FUNCTION)(*args)
+
+
+class TestEncode(unittest.TestCase):
+    def tensorsEqual(self, t1, t2):
+        npt.assert_equal(t1.detach().numpy(), t2.detach().numpy())
+
+    def condEqual(self, c1, c2, key=None, key_assert=None):
+        self.assertEqual(len(c1), len(c2))
+        for i in range(len(c1)):
+            a, b = c1[i], c2[i]
+            if key:
+                (key_assert or self.assertEqual)(a[1][key], b[1][key])
+            else:
+                self.tensorsEqual(a[0], b[0])
+
+    def test_styles(self):
+        pc = PCTextEncode()
+        for k, clip in [("l", clip_l), ("dual", dual)]:
+            for style in ["comfy++", "A1111", "comfy++", "compel", "down_weight"]:
+                with self.subTest(f"TE {k} style {style} does not fail when encoding weights"):
+                    for normalization in ["none", "mean", "length", "length+mean"]:
+                        with self.subTest(f"TE {k} style {style} normalization {normalization}"):
+                            (c,) = run(
+                                pc,
+                                clip,
+                                f"STYLE(old+{style}, {normalization}) this prompt has weights, (a:1.2) (b:1.2)",
+                            )
+                            (c2,) = run(
+                                pc,
+                                clip,
+                                f"STYLE({style}, {normalization}) this prompt has weights, (a:1.2) (b:1.2)",
+                            )
+                            self.condEqual(c, c2)
+
+
+if __name__ == "__main__":
+    print("Loading ComfyUI")
+    import main
+
+    id(main)  # get rid of flake warning
+    import nodes
+    from .nodes_base import PCTextEncode
+
+    (clip_l,) = nodes.CLIPLoader().load_clip("clip_l.safetensors")
+    (dual,) = nodes.DualCLIPLoader().load_clip("clip_l.safetensors", "clip_g.safetensors", "sdxl")
+    print("Starting tests")
+    unittest.main()

prompt_control/test_graph.py

@@ -17,7 +17,7 @@ def apply(cls, text, **kwargs):
 
 
 @mock.patch("prompt_control.utils.lora_name_to_file", find_file)
-@mock.patch.dict("sys.modules", nodes=mock.MagicMock())
+@mock.patch("torch.cuda.current_device", lambda: "cpu")
 class GraphTests(unittest.TestCase):
     maxDiff = 4096
 

prompt_control/test_parser.py

@@ -1,5 +1,5 @@
 import unittest
-from .parser import parse_prompt_schedules as parse
+from .parser import parse_prompt_schedules as parse, expand_macros
 
 
 def prompt(until, text, *loras):
@@ -18,6 +18,19 @@ class TestParser(unittest.TestCase):
         self.assertEqual(p.at_step(0.5), expected)
         self.assertEqual(p.at_step(1), expected)
 
+    def test_equivalences(self):
+        eqs = [
+            [parse(p) for p in ["[a:0.1]", "[:a:0.1]", "[:a:0,0.1]", "[:a::0.1,1.0]", "[:a::0.1]"]],
+            [parse(p) for p in ["[before:during:after:0.1]", "[before:during:after:0.1,1.0]", "[before:during:0.1]"]],
+            [parse(p) for p in ["[a:0.1,0.5]", "[[a:0.1]::0.5]", "[:a::0.1,0.5]", "[a::0.1,0.5]"]],
+            [parse(p) for p in ["[a:b:0.5]", "[a::b:0.5,0.5]"]],
+            [parse(p) for p in ["[a::0.5]", "[a:::0.5,0.5]"]],
+        ]
+        for group in eqs:
+            for p in group[1:]:
+                with self.subTest(p):
+                    self.assertEqual(group[0].parsed_prompt, p.parsed_prompt)
+
     def test_basic(self):
         p = parse(
             "This is a (basic:0.6) (prompt) with (very [[simple]:(basic:0.6):0.5]:1.1) [features::0.8][ and this is ignored:1]"
@@ -71,6 +84,19 @@ class TestParser(unittest.TestCase):
         self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
         self.assertEqual(p3.parsed_prompt, p4.parsed_prompt)
 
+    def test_range(self):
+        p = parse("test [excluded::excluded2:0.1,0.4] test")
+        self.assertPrompt(p, 0, 0.1, "test excluded test")
+        self.assertPrompt(p, 0.2, 0.4, "test  test")
+        self.assertPrompt(p, 0.45, 1.0, "test excluded2 test")
+        p = parse("test [[:included::0.2,0.8]|[excluded::excluded2:0.4,0.9]:0.1] test")
+        self.assertPrompt(p, 0, 0.1, "test  test")
+        self.assertPrompt(p, 0.25, 0.3, "test included test")
+        self.assertPrompt(p, 0.15, 0.2, "test excluded test")
+        self.assertPrompt(p, 0.25, 0.3, "test included test")
+        self.assertPrompt(p, 0.55, 0.6, "test  test")
+        self.assertPrompt(p, 0.95, 1.0, "test excluded2 test")
+
     def test_nested(self):
         p = parse(
             "This [prompt is [SEQ:[crazy:weird:0.2] stuff:0.5:<lora:cool:1>:0.7:nesting:1.0]:completely ignored with tags:HR]"
@@ -101,14 +127,33 @@ class TestParser(unittest.TestCase):
 
         p = parse("DEF(X=[($1):($1:$2):$2])X(test;0.7)")
         p2 = parse("[(test):(test:0.7):0.7]")
-        self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
+        with self.subTest("parameters"):
+            self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
 
         p = parse("DEF(X=[($1):($1:$2):$2])DEF(Y=X(test;$1))Y(0.7) Y(0.5)")
         p2 = parse("[(test):(test:0.7):0.7] [(test):(test:0.5):0.5]")
-        self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
+        with self.subTest("two functions"):
+            self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
+
+        p = expand_macros("DEF(X(a;b)=$1 $2 $3 d)X(A) X(A;B;C)")
+        with self.subTest("defaults"):
+            self.assertEqual(p, "A b $3 d A B C d")
+
+        p = expand_macros("DEF(MACRO()=[empty:$1:$2])MACRO MACRO(;) MACRO(;0.5) MACRO(a;0.5)")
+        with self.subTest("Empty default for $1"):
+            self.assertEqual(p, "[empty::$2] [empty::] [empty::0.5] [empty:a:0.5]")
+
+        p = expand_macros("DEF(X=$1)DEF(Y()=$1)[X Y][X() Y()][X(1) Y(1)]")
+        with self.subTest("defaults, DEF=X vs DEF=X()"):
+            self.assertEqual(p, "[$1 ][ ][1 1]")
+
+        p = parse("DEF(test(1)=prompt $1)DEF(test2((a); (test))=[$1:$2:0.5])test test2")
+        p2 = parse("prompt 1 [(a):(prompt 1):0.5]")
+        with self.subTest("defaults, nested parens"):
+            self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
 
         with self.assertRaises(ValueError) as c:
-            parse("DEF(X=recurse Y) DEF(Y=recurse X) X")
+            expand_macros("DEF(X=recurse Y) DEF(Y=recurse X) X")
         self.assertTrue("Unable to resolve DEFs" in str(c.exception))
 
     def test_misc(self):
@@ -148,11 +193,13 @@ class TestParser(unittest.TestCase):
         self.assertEqual(p.parsed_prompt, p2.parsed_prompt)
         for i, x in enumerate(["cat", "wolf", "tiger", "cat", "dog", "tiger", "cat", "wolf", "tiger", "cat"]):
             step = round((i * 0.1) + 0.1, 2)
-            self.assertPrompt(p3, step, step, x)
+            with self.subTest(step):
+                self.assertPrompt(p3, step, step, x)
 
         for i, x in enumerate([["cat"], ["dog"], ["cat"], ["wolf", ("canine", 1.0, 1.0)], ["cat"]]):
             step = round((i * 0.2) + 0.2, 2)
-            self.assertPrompt(p4, step, step, *x)
+            with self.subTest(step):
+                self.assertPrompt(p4, step, step, *x)
         self.assertPrompt(p4, 0.7, 0.8, "wolf", ("canine", 1.0, 1.0))
 
 

prompt_control/utils.py

@@ -87,7 +87,7 @@ def find_closing_paren(text, start):
     return len(text)
 
 
-def get_function(text, func, defaults, return_func_name=False, placeholder=""):
+def get_function(text, func, defaults, return_func_name=False, placeholder="", return_dict=False):
     rex = re.compile(rf"\b{func}\(", re.MULTILINE)
     instances = []
     match = rex.search(text)
@@ -98,7 +98,19 @@ def get_function(text, func, defaults, return_func_name=False, placeholder=""):
         funcname = text[start : after_first_paren - 1]
         end = find_closing_paren(text, after_first_paren)
         args = parse_strings(text[after_first_paren:end], defaults)
-        if return_func_name:
+        ph = None
+        if placeholder:
+            ph = f"\0{placeholder}{count}\0"
+        if return_dict:
+            instances.append(
+                {
+                    "name": funcname,
+                    "args": args,
+                    "position": start,
+                    "placeholder": ph,
+                }
+            )
+        elif return_func_name:
             instances.append((funcname, args))
         else:
             instances.append(args)

pyproject.toml

@@ -1,14 +1,13 @@
 [project]
 name = "comfyui-prompt-control"
 description = "Nodes for convenient prompt editing, making many common operations prompt-controllable"
-version = "2.0.0-beta.11"
+version = "2.0.0-rc.6"
 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"]
 
 [project.urls]
 Repository = "https://github.com/asagi4/comfyui-prompt-control"
-#  Used by Comfy Registry https://comfyregistry.org
 
 [tool.comfy]
 PublisherId = "asagi4"